In the 2000’s, Adam and Eve were a pair of prototype lab robots, built to fulfill a dream. They were a step on the way to freeing scientists from the painstaking tedium of traditional labwork: gathering samples, pipetting fluid from one place to another, carefully calibrating the placement of sensors and lasers, and noting down measurements in lab notebooks. For decades scientists hoped there was a way to smooth away that work, using robots to fill in the tedious and elementary tasks so they could focus on bigger, deeper questions.

When the first useful robot arms became available in the 1980’s, scientists adopted them almost immediately. The first prototype automated labs were enormously expensive, but over time the technology improved. Specialized autosamplers and automated probe stations filled in the most routine parts of lab work in a reliable way, and it became increasingly feasible to jury-rig together different low-cost systems into a functioning whole.

At the same time, others pursued the idea of more generally capable robot scientists, able to do a variety of experiments and function for days without input. Some, like the creators of Adam and Eve, achieved flashy results even before the rise of modern AI. Now, more advanced lab automation and LLM integration has led to the rise of true “zero-click” labs that can seamlessly perform routine procedures and record their results, and even closed-loop experiments that test basic hypotheses by themselves. As a result, science is progressing faster than ever.

From Hospitals and Chip Fabs to Research Scientists

Some of the first big successes for lab robots involved clinical work. One of the most famous was the work of Masahide Sasaki at Kochi Medical School in Nankoku, Japan. With a team of lab technologists, he designed a lab where robots and conveyor belts carried test tube racks and patient samples, while robot arms pipetted serum for assays. The lab was a real-world success, performing all clinical lab work for a 600-bed hospital, and inspired imitators around the world.

By the 1990’s, clinicians were touting the possibility of “Total Laboratory Automation”, or TLA systems, large-scale systems that could organize and test patient samples. These systems were an enormous boon to hospitals and some universities, but their multi-million-dollar cost kept them out of the labs of most individual scientists.

At the same time, though, more specialized machines were getting more affordable. Different fields had different needs. In biology and some kinds of materials research, where one needs to test a variety of substances in precisely measured doses, all while keeping them pure, autosamplers were a particularly important innovation: systems that could regularly take samples, pipetting liquids in a biology lab or moving metal samples to and from measuring stations. In other fields, like optics, physical precision was the most important. Automated probe stations could position sensors with the utmost consistency, leading to their adoption first in chip manufacturing, and soon after in research labs.

As these systems became more available, a major challenge has been getting machines from different manufacturers to work with each other. Rather than waiting for manufacturers to link up their systems, scientists found low-cost ways to fill in the gaps. With open-source software, researchers in smaller labs could jury-rig these expensive systems together, so that even if each system is controlled with its own software a common lab system would manage the whole. Thus, research labs can increasingly live the TLA dream.

The Genesis of Robot Scientists

Researchers have even bigger dreams, though. Autosamplers and automated probe stations could perform well-defined tasks, collecting and testing samples. But in a sense, the simplest science experiments are also well-defined tasks. The scientific method we learn in school is a cycle: use a hypothesis to design an experiment, perform the experiment, analyze the data, and use your observations to propose the next hypothesis. And while the most daring hypotheses require scientific expertise, the scientific method happens in smaller ways as well. A scientist might look for a fast-growing strain of cell or a molecular tweak to a new drug, testing smaller hypotheses about one effect or another on the way to a more impactful conclusion.

The Robot Scientist project took that small-scale scientific method, and gave it a body. Its designers billed it as the first machine to discover scientific knowledge autonomously. In the 2000’s their first creation, Adam, was built to study the interplay of genes and enzymes in the metabolism of baker’s yeast. Using precise cameras, Adam tracked the growth of the yeast under different conditions, autonomously forming hypotheses and designing experiments. It carried out over one hundred experiments each day, working around the clock, its only interaction with the outside world when technicians dropped off reagents and took away waste. Long before today’s more powerful AI systems, Adam was nonetheless able to perform routine scientific work to an impressive level.

The team went on to develop another Robot Scientist, which they named Eve. Eve was designed to be more practical, a machine that could actually be employed more broadly. It was a drug development platform, developed to make the creation of new pharmaceutical drugs faster and cheaper. With more advanced AI and synthetic biology capabilities, Eve was able to search through large libraries of candidate chemicals, test to confirm successes, and hypothesize relationships that let it predict new promising candidates, a system which proved itself dramatically more efficient than a brute-force search for drug candidates.

Science Without a Click

Now, AI systems are dramatically more powerful. While older systems were best with structured data, now systems equipped with language models can incorporate information from the scientific literature. At the same time, advances in machine learning fundamentals have made techniques like reinforcement learning more powerful, training AI systems to discover new strategies. And with more efficient software and hardware, all of this becomes more widely available.

Instead of jury-rigging systems together, scientists can use commercial platforms designed for “zero-click science”. The automated systems can move samples, carry out assays, reposition sensors, and even record results in a digital lab notebook, keeping everything organized so scientists don’t have to step in and manually transfer materials or data back and forth.

For those working beyond these platforms, at the cutting edge, the possibilities are even greater. Increasingly, scientists are experimenting with true closed-loop experiments. In these setups, a system doesn’t just perform experiments one by one at scientists’ direction. It makes choices, using the data found at each step to decide the next experiment.

One recent open-source standout, BacterAI, is a reinforcement learning-powered AI scientist that can optimize strains of bacteria, finding which bacteria will thrive in which environments. A key element of BacterAI is its automated lab setup, the result of years of effort custom-building a system that can keep up with AI-powered insights. When the team began in 2020 with a more or less normal lab robotics system, they could carry out 300 experiments per day. After a year refining the systems, they could do ten thousand.

In our last post, we mentioned OpenAI’s success with closed-loop experimentation earlier this year, where they used an AI scientist to lower the cost of cell-free protein synthesis. An essential contributor to that success was an automated cloud laboratory run by their collaborator Gingko Bioworks. The robotized lab was able to handle liquids, incubate samples, and perform fluorescence measurements, all using a modular array of reconfigurable automation carts, or RACs, devices which can house a variety of instruments from different manufacturers in a standardized, robotizable setup.

Another impressive system was demonstrated at MIT last year. Billed as the first AI-driven robotics platform for free-space optics, their system didn’t just work with a fixed setup. Using robot arms, cameras, and LiDAR, their robotic platform was able to take natural language instructions (like “build an interferometer”) and assemble an experiment to meet them, carefully testing to make sure every piece was aligned exactly where it needed to be. The versatile setup could pick up a wide range of tools to make many different types of measurements, covering a wide range of functionality without the painful calibration by hand that used to slow such experiments down. The work is part of a broader revolution in self-driving labs that is exciting the optics community.

Today’s most advanced robots are being used for some of the most ambitious projects imaginable. The successors of robots like Adam and Eve are curing cancer, editing genes, and finding new nanomaterials.

We see more on the horizon. At e184, we’re working on the toughest challenges yet. The technology we’re building will break past limits that have held us back through our evolutionary history, from restrictions on how we reproduce, to how we think and interact with the increasingly technological world around us, to the limits of our genomes that can hold us back from a healthy, long life.

We’re actively hiring specific roles for scientists and engineers, but we also have other opportunities. We want to hear from people with a plan to push science to the next level, experts in systems like AI and robotics who are ready to take the possibilities of the present and bring the transformations of the future. If that sounds like you, you can submit a general application here.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

The remarkable thing is that increasingly, one can fuel the other. AI progress is powering scientific progress in a multitude of ways, bringing us closer to the world of our dreams. And it’s poised to do a whole lot more.

A Dog, and a Preview

Last March, people were captivated by a series of X posts by Australian tech entrepreneur Paul Conyngham, describing his AI-powered quest to treat his dog Rosie’s cancer. After Rosie developed leg tumors, Conyngham turned to AI tools to make a plan to save her. The AI talked him through different emerging technologies for customized medicines, showing different techniques scientists had refined to quickly build new treatments, like finding a protein to shut off the cancer-causing mutation or mRNA vaccines. It helped him find labs to handle sequencing, use tools like AlphaFold to find protein structures, and get in contact with professors at several universities. Consulting with those scientists, Conyngham and his AI “assistants” settled on an mRNA vaccine as the best path forward, and as of March the treatment seemed to be working, with the dog’s tumors no longer growing.

While Conyngham had the resources to be on the cutting edge of these techniques, it will likely only get easier to personalize medicine in this way in future, with more people benefiting from the groundwork laid by the scientists who worked on these ideas. Some companies are already advertising AI-powered personalized medicine, while others have made their AI models widely available, to encourage researchers and members of the public to use them.

AI has been enormously helpful to Conyngham, and to Rosie. But Conyngham was clear in his post as to what was accomplished via AI, and what was not:

“What the chat bots did NOT do:

They did not collect samples. They did not isolate or sequence the DNA. They did not physically manufacture the vaccine. They did not administer it. Many brilliant scientists were required - including Professor Pall Thordarson at the UNSW mRNA Institute who manufactured the vaccine, Professor Rachel Allavena & Dr. José Granados at the University of Queensland who administered it, and Professor Martin Smith who provided expert guidance on the bioinformatics throughout.”

The chat bots were a powerful tool for navigating scientific publications, letting Conyngham benefit from the work of many others. They didn’t replace the inspiration and inventiveness of scientists, of course. But also, they didn’t do the more routine tasks: those had to be done by scientists as well. They didn’t sequence DNA or manufacture vaccines. But what if they could? What if scientists didn’t need to spend their time on procedures like those, and could focus on the science itself?

AI-Powered Labs

Since scientists were first exploring the potential of AI, some speculated that it could fuel the labs of the future. The first ideas researchers explored involved using AI to optimize lab setups. Often the AI would propose strange new techniques that led to surprising improvements. Some of these scientists figured out ways to have the AI optimize the experimental setup on the fly, like this result from 2016 that used AI to speed up the creation of an extremely quantum state of matter called a Bose-Einstein Condensate.

Now, more powerful AI systems can not only optimize processes, but behave in many ways like junior scientists, discussing and refining ideas. And just as self-driving cars are on more and more streets in more and more cities, more and more researchers are building self-driving labs. These labs use robots to build and carry out experiments, streamlining key tasks so an AI can iterate with minimal human involvement, and sometimes, do so on an enormous scale.

Researchers have built platforms to construct and optimize optical experiments, using robots to achieve the precision required for such delicate tasks. Others are using AI-powered self-driving labs to discover new materials, improving everything from power transmission to solar cells to insulation. Enough different labs are exploring these techniques that some researchers have proposed benchmarks to compare the different approaches, finding that many self-driving labs improve materials four or even six times faster than conventional techniques.

The potential of these techniques has led many institutions and companies to invest in self-driving lab technology, or other forms of AI-powered science. Oak Ridge National Labs in Tennessee now has a number of prototype self-driving “Labs of the Future”, while the University of Illinois at Urbana-Champaign is building a cloud-based autonomous laboratory for enzyme engineering, so researchers around the world can request custom-built proteins. Naturally, OpenAI itself is pursuing this technology, and recently announced that they used a self-driving lab to find a cheaper way to manufacture proteins, an important cost for pharmaceutical companies. The AI-powered lab was able to not only experiment, but search the scientific literature, finding papers that suggested new techniques to try and then proposing methods to improve on them. The result sparked a debate in the scientific community, speculating about just how far these systems will be able to go, and discussing how the role of scientists will change with the ability to operate on such an extended scale.

What Comes Next?

At e184, we’re eager to overcome the limits of human biology, enabling people to think faster, live healthier, longer lives, and build families on their own terms. The rapid progress of AI-powered science inspires and excites us. It shows that more is possible now than was ever possible before. Ideas that used to be overambitious, that could never be completed within a single career, now can be sped up and streamlined. It’s the right time to dream big: to look beyond incremental progress and set out to do something world-changing.

If the potential of AI-powered science excites you too, or if you’re already an expert, consider getting involved. Our sister company 42 is hiring researchers and engineers for their AI-powered science platform. You could work to improve the AI itself, or develop the infrastructure for autonomous labs to speed up scientific progress. Or if it’s our mission at e184 specifically that inspires you, check out our career page, or submit a general application. The labs of the future are coming: will you build them?

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

So many people need IVF that it has defenders across the political spectrum, and many countries have bipartisan support for making it as accessible as possible. For the one in six couples who struggle with infertility, IVF is often their last hope when other techniques like intrauterine insemination fall short.

One of our readers shared what IVF meant to her:

“IVF was not an easy decision, and it’s not an option for everybody, for many reasons. But when I see my child grow, when I hear her laugh, when I hug her, I know that every step was worth it.”

It’s a feeling that today many parents recognize.

But IVF wasn’t always seen this way. In the 1970’s, when pioneering biologists like Robert Edwards first developed IVF, the media told sensationalist, science-fictional stories of where it might lead, while scientific authorities downplayed IVF’s value and exaggerated its risks.

In the end, families ignored the pressure. IVF didn’t lead to chaos or a science fiction dystopia. It let millions of couples build lives without the spectre of infertility. And as new reproductive technologies develop, we expect them to do the same.

A time when medical researchers believed infertility should not be treated

It’s hard to imagine today, but in the 1960’s and 1970’s, very few researchers worked on infertility. A small number of visionaries understood its importance and tried to find solutions, including Landrum Shettles in the US and the team of Robert Edwards, Jean Purdy, and Patrick Steptoe in the UK. These pioneers were often looked down on by their colleagues, who didn’t appreciate just how many couples struggled with infertility, or just how much those struggles can impact a family.

In 1971, when Edwards, Purdy, and Steptoe applied for funding to develop IVF, they were told that the work was of low priority. Edwards recalled in particular that the committee had a “belief that infertility should not be treated because the world was overpopulated”. They viewed any new procedure to treat infertility as experimental, and imposed additional restrictions beyond those they applied to research they considered more important, like contraception.

Later in 1971, Edwards published an essay in the journal Nature, based on a talk he had given a few years earlier. In it, he argued that overpopulation should not be a reason to deny infertile parents the opportunity to have children. He emphasized the suffering that an inability to have children can cause,

“The desire to have children must be among the most basic of human instincts, and denying it can lead to considerable psychological and social difficulties,”

and argued that the medical profession had a duty to help,

“Infertility seems to be a clinical defect to be remedied if possible by medical attention. It cannot be a rational basis for denying on ethical or other grounds the right of some couples to have children.”

Addressing that era’s worries about overpopulation, he wrote,

“Social concern for overpopulation is now widespread. Our view is that a campaign against overpopulation should be directed to all parents generally and not enforced on a selected few.”

Despite the passionate arguments of people like Edwards, the number of researchers working to cure infertility remained small, and bias against them from the rest of the medical research community remained widespread. Edwards’ colleagues at Cambridge even tried to dissuade other researchers from working with him due to his research focus. Martin Johnson, who chose Edwards as his PhD advisor in 1966, recalls “the other staff members took me to one side and asked why was I going to waste my time on research with that charlatan, who worked on stuff ‘down there’ and spoke to the press.”

A Fascinated Public

Those conversations with the press were unusual in other areas of medical research at the time. But for the IVF pioneers, it was important to keep the public informed. And the public, in turn, was fascinated.

In 1971, Look Magazine published an article about the work of US IVF researcher Landrum Shettles, describing him as developing a “test-tube baby”. The name stuck, despite the fact that in vitro fertilization used a petri dish, not a test tube. A 1974 article in the New York Times went further, alluding to the science fiction novel “Brave New World” and quoting James Watson, one of the discoverers of the structure of DNA, who said that as a consequence of IVF, “All hell will break loose, politically and morally, all over the world.”

Edwards complained that these speculations were “dramatically doom-lit and gaudily coloured by science-fiction fantasies and visions – fantasies of horror and disaster, and visions of white-coated, heartless men, breeding and rearing embryos in the laboratory to bring forth Frankenstein genetic monsters.” He saw it as his duty to talk to the press and correct the misconceptions, arguing that the technology he was developing was going to be used to benefit ordinary families.

It is striking, looking at the 1974 New York Times article, to see how much of what they treated dramatically then is ordinary today. The article talks not only about IVF, but about the possibility of surrogate mothers, egg donors, and parents choosing sperm donors based on their health, rather than choosing at random as people were obliged to do back then. At the time, all of these were viewed as scary possibilities that would test the limits of ethics and bring us closer to disaster. Instead, they now support millions of families all over the world.

IVF Becomes a Reality

Edwards, Purdy, and Steptoe had to walk a long road until they were ready to help their first family. Without funding from the UK’s Medical Research Council, they were supported by private foundations, in particular a Californian heiress named Lillian Lincoln Howell. Their first patients were sworn to secrecy for their own protection, and the hospital even faced a bomb threat.

But in 1978, shortly before Steptoe was due to retire, the team succeeded, and helped Lesley and John Brown have their first child, Louise Joy Brown.

Soon, the team was receiving hundreds of letters, as couples wrote hoping for the chance for treatment. Lucy Daniel Raby was one of those early patients.

“It was all new and a bit sci-fi,” said Raby. “We were the early pioneers, and part of this exciting experimental process. I didn’t have a second thought about it once I knew it was the only way I could get pregnant. We were lucky that it was available.”

Over time, IVF has become more and more reliable, and more prospective parents gained access. Doctors and researchers came to understand just how much good it could do, and the practice expanded.

Saving more families

Edwards died in 2013, the last of his team. He lived long enough to see in vitro fertilization go from scary media stories to a vital part of fertility treatment. In 2010, just three years before his death, he received the Nobel Prize in Physiology or Medicine. Finally, the world recognized just how much good IVF had done.

Edwards’ ambition was to cure infertility, but infertility is still with us today. While medical researchers in the 1970’s were worried about overpopulation, now fertility rates are falling around the world. The movie Joy, a dramatization of the development of IVF, suggests that Edwards’ collaborator Jean Purdy had endometriosis. While we can’t know that with certainty, endometriosis can interfere with fertility in ways that IVF can’t address, disrupting the ovarian environment or making it difficult to carry a child. There are better treatments now than in Purdy’s day, but many women with endometriosis still can’t have children.

At e184, we’re working to change that. We are developing in vitro gametogenesis, the next step after in vitro fertilization: not merely uniting an egg and sperm in a clinic, but creating eggs or sperm using cells from a patient’s body, so that even women who have difficulty ovulating can have children. And we are working towards ectogenesis, technology that can support a fetus outside of the womb, so that women who have difficulty carrying a pregnancy, or cannot carry a pregnancy at all, can still have a child on their own terms.

Right now, some are afraid of these technologies, just as some were afraid of IVF in the 1970’s. But the history of IVF is reassuring. Far from a science-fiction dystopia, IVF did exactly what its developers proposed, and exactly what families around the world asked for: it helped parents have children. The technologies we build will do the same.

In an upcoming post, we’ll talk about the pressures parents face now, and how our new technologies will address them. Like Edwards, we want to stand up for your right to have children. We’d love to hear from you about the barriers you’ve faced, and what it could take to overcome them. You can contact us at stories@e184.com to tell your story.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

The Foresight Institute has been bringing together futurists and scientists for forty years. This year, they reached out to partners like us to craft an event honoring those forty years of progress, and looking ahead to the next forty. We had worked with them before on a Neurotechnology workshop, but this would be something different. It was a chance to touch on more than just one aspect or research program, but instead to highlight progress across the full arc of human life. More than that, it was an opportunity to showcase what we are building at e184: a set of technologies focused on overcoming biology’s most long-standing limits.

The community met that showcase with aplomb. We were blown away by their passion, curiosity, and open-mindedness. Presenting to Foresight’s community of experts and dreamers, we saw our core values reflected back at us. This was a community that won’t let barriers stand in its way – one that cares deeply about getting things right and making a difference on the largest scales.

Hearing from experts across two days of talks, we found those values emphasized again and again. Foresight managed to draw on a worldwide community for the event: not just UK locals, but experts from the US and elsewhere. It was a fast-paced cavalcade of insights, both from established voices and eager newcomers.

The evening before the talks, a VIP gathering brought together sponsors, speakers, and the Fellows and Grantees supported by the Foresight Institute. A highlight of the evening was a rapid-fire series of introductions from the fellows, where each had one minute to describe what they’re building. Standouts of the evening included Ninon Masclef, an artist at MIT Media Lab who uses brain data from sleeping participants to make evocative images linked to dreams, Sven Truckenbrodt, who works on mapping the brain using expansion microscopy, Sobia Hamid, who talked about her work with Infinicell Bio building the groundwork for regenerative medicine, Alberto Privitera, who develops technologies linking quantum spin and light on molecular scales, and Léo Pio-Lopez, whose work at Tufts University’s Allen Center ties together cognition, the logic behind biological processes, and AI.

On Saturday, researchers in Emerging AI Paradigms talked about building new capabilities such as letting AI agents learn over time and embedding them into the world, and new approaches to ensure AI is used with safety and privacy in mind. Seeing the dialogue around safety in the AI community made us more optimistic about the dialogues to come in brain-computer interfaces and reproductive technologies, where key conversations are just beginning.

Our Life Unlimited track followed that morning, starting with a keynote by our Chief Operating Officer Marianna Krell on our core mission and the technologies we’re building. The following talks showcased the potential benefits of greater control of the fundamental building blocks of biology, especially for the challenge of increasing longevity, a subject where João Pedro de Magalhães’ talk was particularly exciting.

The session on Neurotechnology, Whole Brain Emulation, and BCI later in the day was a chance to catch up with old friends in the neurotech space and meet new faces. Sergey Stavisky’s talk stood out for showing the impressive level that speech decoding can now reach with implantable BCI systems, while Christopher Rozell’s talk showcased impressive progress for neurotechnology that treats depression. The last session of scheduled talks for the day was a mix of emerging technologies, from fusion and high-temperature superconductors to building with DNA strands and new approaches to thinking about what makes life special in the universe.

Sunday’s talks were, for the most part, more practical. The speakers in the Funding X and Pathways to Implementation tracks focused on the tricky challenges of how to build value and fund breakthrough research in a VC ecosystem focused on fast returns and distracted by hype. There were a number of encouraging funding models, showing that more people in this space are thinking hard about how to enable real progress.

Finally, the last track concerned a new theme of Foresight’s, Existential Hope in the Age of AI. This theme, of hope for the future, was especially well-suited to Foresight’s 40th anniversary. Christine Peterson, one of Foresight’s founders, talked about how the last forty years inspire her to look forward to the next forty, and described the wide range of emerging technologies whose development Foresight fosters. Anders Sandberg laid out a particularly inspiring vision, arguing that a Dyson swarm capable of capturing the energy of a star could be constructed in a surprisingly short amount of time, reasoning based on physical limits for self-replicating space probes. Others talked about efforts to support science going forward.

The passion at Vision Weekend UK was infectious. We’re enormously glad to have contributed to bringing so many talented, insightful people together, and profoundly grateful for the warm welcome and keen interest we found. It’s a community we’ll keep our eye on in future, looking forward to great things.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

Over its forty-year history, the Foresight Institute has had a central role in articulating those dreams. Founded in 1986 by Christine Peterson and Eric Drexler, Foresight has inspired researchers and innovators to aim for the stars, advocating for work towards technologies from nanotech to artificial general intelligence. Over that time, they have become the core of a substantial community dedicated to the question of what comes next.

Now, many of the technologies Foresight’s community has dreamed about are closer than they have ever been. Brain-computer interfaces are increasingly capable, augmenting the capabilities of a small group of pioneers, while technologies like artificial wombs have gone from science fiction to viable research goals. Scientific progress has yielded increasingly fine-grained control over the building-blocks of life itself. As a result, we may well be the last generation that has to accept biology as it is.

Last year, we worked with Foresight to highlight just one aspect of that story, that of neurotechnology. We’ve been in contact with them since then, discussing ways we can support the community of dreamers they foster, and showcase how close some of those dreams are to being achieved. Inspired by those discussions, this year, we wanted to tell a broader story: one that doesn’t concern just one of our focus areas, but the whole picture.

That’s why this year, we’re partnering with Foresight to offer a preview of what’s in store. For their Vision Weekend UK event on June 5-7, we worked with them to select speakers for a special track. Called Life Unlimited, this track brings together experts in areas of the life sciences where we see the greatest potential to overcome long-standing human limitations, giving humanity unprecedented control of both how life begins and how it ends. We plumbed both Foresight’s network of experts and our own contacts, finding speakers who could give an exciting peek at the world of tomorrow.

The session begins with a keynote address by our founder, Kirill Eves. In his talk, he will talk about our core mission at e184, focused on extending the limits of human potential in reproduction, the genome, and cognition. He’ll talk about the technologies, like in vitro gametogenesis, artificial wombs, and non-implantable brain-computer interfaces, that we are pursuing to achieve that mission. And he’ll talk about how we plan to grow in future, with directions like genome editing letting us tackle biological limits at their roots.

The next talk, titled “Hacking Aging Biology”, is by João Pedro de Magalhães, the Chair of Molecular Biogerontology at the University of Birmingham. A world leader in using big data to study aging, de Magalhães’s works to decipher how the human genome impacts how we age, with the potential to radically increase longevity, extend health, or even reverse aging. As a futurist and thinker, he exemplifies the community that Foresight fosters.

From our own expert contacts, we invited Google DeepMind’s Žiga Avsec to give the third talk, titled “Advancing Regulatory Variant Effect Prediction with AlphaGenome”. In a world of increasingly impressive AI achievements, AlphaGenome stands out for its potential. Able to predict the effects of individual mutations on a wide range of biological processes, Avsec’s work with AlphaGenome paves the way towards an era where we can treat DNA like computer code.

Finally, University of Oxford professor Lynne Cox is an expert on cell senescence, one of the most important processes behind aging on the cellular level. She has identified the chemical rapamycin as a “genoprotector” that can protect aging cells from accumulating DNA damage. Her work connects to the Foresight Institute’s longstanding interest in longevity, and the scientific question of whether we can one day live longer without diseases like cancer holding us back.

We’re proud to support Foresight in choosing and showing this set of speakers, and celebrating forty years of dreaming big. If these talks excite you like they excite us, then we look forward to seeing you on June 6 at 11am for the Life Unlimited track at Foresight Vision Weekend UK!

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

Among our advisors, no-one exemplifies this aspect of our work like Daniele Faccio. A Professor of Quantum Technologies at the University of Glasgow, Faccio started his career working on optical telecom devices before doing a PhD in high-intensity laser physics at the University of Nice Sophia Antipolis in France. Before starting the Extreme Light Group in Glasgow, he held positions at the University of Insubria in Italy and Heriot-Watt University, Edinburgh. In that time, he expanded his work from laser and optical physics to modeling black holes and testing fundamental aspects of quantum mechanics. Now, a major focus of his lab is on applying the technologies developed there to advance brain imaging, using functional near-infrared spectroscopy, or fNIRS. We asked him about the lines tying this work together, and the developments he sees most encouraging for the future.

What follows is an edited and curated version of our discussion.

Your work covers a very wide range of topics these days, from fundamentals of quantum mechanics to new imaging techniques for the heart and brain. Thinking back to when you started your PhD in 2004, how much of that do you think you could have imagined then?

My training was as a laser physicist doing strong-field physics, working with very big lasers and looking at light-matter interaction at very extreme intensities. As a physicist it’s natural to gravitate toward big questions, and in 2008 a Science paper by Ulf Leonhardt came out describing how one could use the exact same kind of physics I was interested in with these big lasers to simulate black hole physics in the lab. And for me that was a wakeup call. I was super excited about this, and that led to a total pivot in the direction of my research.

When did you start working on brain imaging, and what motivated you there?

I had moved to Edinburgh and started collaborating with Ulf Leonhardt, who was at St. Andrews University, just an hour away. I got some successful grants, some good results, fairly interesting physics, but started at some point to want to move on. This was all about simulating black hole physics. You’re not creating a real black hole in the lab. And so I had the desire to do something a bit more grounded in the real thing.

We’d started to play around with what was back then a new generation of cameras that could detect single photons. Single-photon sensors as in a single pixel had been around for a long time, but the ability to have large arrays of these cameras that could actually take images was relatively new. And we started using those to look at the analogue black holes we were generating. Then, coming back to the point of view “What can we do that’s real?”, we started to think about what else can we do with these cameras.

And they have three features.

One is, they are sensitive to single photons, of course.

The second is, because they are operating what is called Geiger mode, so they’re detecting light in its particle form, it gives a click every time you detect a photon. That can be used to give you very very high timing-precision in terms of when the photon actually hits the sensor. You can start a clock and then you can see when the photons are hitting the detector, and that allows you to then build these movies at a trillion frames per second.

That’s one aspect, and the other aspect is that not only can you have this trillion-frame-per-second resolution, but you can also keep on reacquiring these movies at very high sampling rates. I can take a thousand frames a second, and each of those frames will have encoded a trillion frames per second.

So you have got these three aspects: very high frame-rates, single photon sensitivity, and this trillion frames per second capability. So what else can we do with that?

Then I picked up on some work that Ramesh Raskar was doing at the MIT Media Lab. He was trying to look around corners. A lot of people are working on this now, but back then it was quite a revolutionary concept and they were using a different kind of technology. It was very clunky and they had to take data for a full day to be able to image something behind a corner. I think, with these cameras, we can do this in a second! We can do it in real time. And so that’s what we set out to do.

And I think we gave the first demonstration that non-line-of-sight imaging wasn’t just a curiosity. It is something that could have real world impact. We showed that we could track an object moving behind a wall in real-time. And then we started looking a bit more carefully, and you realized that the same maths that you need, the computational imaging techniques involved, the same computation needed to retrieve the image of something behind a corner or around a wall, is the same that you would need if you wanted to just look directly through a wall.

And specifically, when I say through a wall, what I mean is that many objects that look opaque, including the human head, clouds, the snow, sugar, salt, essentially, anything that’s white, is opaque not because it’s absorbing, because otherwise it would be black. It’s opaque because it’s scattering. And so if you look at a cloud, it’s white at the top, and then it does become black at the bottom. But that’s because all the light in this diffusion process and the scattering has been back-reflected out to space, and so there’s no light at the bottom. But essentially, to the first approximation it is transparent, just very highly diffusive. It looks opaque to us, but using these time-resolved techniques we showed that it is possible.

If there is one common thread in my research all the way from 2004 to today, it is time-domain. Even back then when I was looking at strong field physics, you have these ultrashort 30-second laser pulses, you need very evolved time-domain techniques to be able to capture what is going on. Same with the black holes and optical fibers, and now the same with the brain imaging. It’s that common thread. I know how to do time-domain imaging.

Then the real trigger I think for me was one of those family stories you often hear from neuroscientists. My mother had a stroke, and of course it’s a tragic moment, but as a scientist what really struck me was observing the obvious fact that the brain was intact and functional, but my mother couldn’t articulate words. And that got me hooked. What is going on? How is that even possible? I had as a non-expert in brain lesions always thought that if you had a brain lesion, some kind of malfunction, that would impact all of you and wouldn’t have this sort of separated effect, right? But everything was intact, she just couldn’t speak. Very weird. So I gradually got really drawn into the problem.

So then you look around, what have people been doing and what do they know? And we know a lot. Thanks to functional MRI, neuroscientists now know what ADHD or depression looks like in the brain. Adrian Owen has been able to show that people who have locked-in syndrome, you can ask them to reply yes or no by asking them to imagine that they’re playing tennis, and that’s activating their motor cortex. You can see that. And these are remarkable things in terms of the impact that they can have on people.

But at the same time, the richer countries have the most access to these MRI machines. The US has about 40 MRI machines per million people. The UK, we’re doing less well, we’re doing quite badly, less than 10 per million people. And then the question is, what if I could have a wearable MRI maybe at one hundredth, or why not one thousandth the cost?

In 2024 you had this paper with an amazing title, “Photon transport through the entire adult human head”. What gave you the idea to do to try to do something like that?

I guess it’s the physicist’s sort of taking up the challenge of “You can’t do that”.

There’s a famous handbook in TCSPC, time-correlated single-photon counting, and at some point they say, “Can we use these single-photon sensors and TCSPC for photons that have gone all the way through the head?” And they use the words that you should never use: they said “This is impossible”. I said, “Okay, is it now?”

What we’d been trying to do is, we were taking slabs of polystyrene, this yellow insulating foam that builders use. Just by magic, it turns out that it’s got very similar scattering absorption coefficients to brain tissue. We’re just trying to say how thick can we make this foam and still be able to see inside it. So it’s two and a half centimeters, and then it was working really well. They say, “We should stop here.” I said, “Why? Keep going.” Now five centimeters. It’s still working. I said, “Is it now? So, make it go thicker, 10 cm.” So, how thick can we make this stuff and keep going? And then you see this claim that through the whole head is impossible. So, okay, that may not be true. Let’s give it a shot.

They weren’t wrong. It’s very hard. It was many years, a lot of pain, people getting very frustrated and saying after two years I don’t want to deal with this anymore, I’ve got to work on something else, but we kept on insisting and didn’t give up.

fNIRS today is currently limited to probing just the outer surface of the brain. At the moment, this is the one big advantage MRI over fNIRS: MRI will give you a 3D volume of the full brain, and fNIRS can’t do that yet. So that is one of the reasons for trying to do that: can we see deeper?

Where do you place yourself in the world of quantum technologies?

Anybody really working in this area, you sort of sit in this weird zone where you’re overlapping a little bit with many areas. You overlap a bit with neuroscience, you overlap a bit with engineering technology. You overlap a little bit with the quantum technology, and I was playing around with fundamentals.

This is why in the lab we still have people working on very fundamental questions in quantum mechanics. Can rotation or gravity induce entanglement between photons that otherwise start off their life from completely independent places? Do they become entangled thanks to the action of rotation, or gravity? And one reason we’re still pushing that forward is, the technology that you learn how to build and develop there is actually of interest for what we’re doing.

So I’d say strictly, trying to develop a wearable MRI, this healthcare technologist-neuroscience space is probably not quantum technology, but you need the backdrop of the quantum technology and that drive to understand fundamental questions to feed into what you’re doing with the BCI work. I think what I’m trying to hold on to is I don’t want to forget that the fact that the reason I’m here today is because of that fundamental work that we were doing, and therefore I think it still has a place.

Let’s talk more generally about brain imaging. What’s exciting right now?

For me it’s definitely the time-domain aspect. There’s only a couple of companies worldwide commercializing that technology. You’ve got PIONIRS in Italy. It’s a small company building a fairly small device, but with a fairly high-quality signal-to-noise ratio. Then you’ve got Kernel in the US, that instead has gone down the full-head wearable direction. We have both devices in the lab, and in a sense I think the fact that this technology has got to the point that it’s got sufficient interest that people are willing to fund that, that it can become commercial, that I find exciting because now you got this loop where the fundamental research has led to commercial activity, and thanks to that commercial activity we can now go back and do more fundamental research. So that commercial push, that then loops back into the research, and that at the moment is the thing that’s exciting.

As an advisor for e184, you bring your expertise to advance our research. What attracted you to work with e184?

It was the visionary approach, and the openness. It was clear what they wanted, and they were open to discussing. Money wasn’t really the objection, but the fact that maybe a technology doesn’t exist yet also wasn’t an objection. There was just that clear vision of, sure, the technology doesn’t exist, and we are here to develop it. Because we believe that wearable neural interfaces are the future.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

The two events may seem quite different. Foundation Models for the Brain was a gathering of established experts, with a wide-ranging set of concerns. Global NeuroHack’s teams were young: many were undergraduates. And in the few days they had to build, they were laser-focused.

But between the two events was one core mission: supporting the brain decoding community. We got senior academics and neurotech leaders talking about some of the biggest challenges in an incredibly promising area. And we got young talent learning from mentors, refining concepts, and exploring exciting ideas. By getting the community talking and thinking, we’re accelerating progress, getting everyone closer to the scalable solutions we need.

Foundation Models for the Brain

At our evening symposium in NYC, one approach to scalability, foundation models, took center-stage.

Foundation models represent a new approach to understanding and decoding the brain. The hope is that with enough data, researchers can train models that are not just powerful, but generalizable: able to go beyond their training data, and work in new situations. It’s an approach that may be enormously important for developing truly versatile BCI.

After a few opening remarks, our first panel focused on the bottlenecks holding back this approach. Moderated by Patrick Mineault of the Amaranth Foundation, the panel gathered Vinay Jayaram from Alljoined, Andreas Tolias from Stanford, John Crary from Mount Sinai, Liam Paninski from Columbia, Cole Hurwitz from IBM, Marcelo Mattar from NYU, and David Moses from UCSF. One of the biggest themes of the discussion was data. Foundation models are data-hungry, but on top of worries about acquiring enough data, participants wanted to make sure we gathered the right data. That could mean data that takes into account the context the measurements were taken, to make models that generalize better to the real-world, but it could also mean data that avoids hand-made labels that could build in limiting expectations. Overall, there was a feeling that being smart about what data the field uses can go a long way.

The second panel of the evening brought together members of the funding ecosystem. Moderated by Sean Escola of Protocol Labs, the panel included Dimitris Sakellariou from Piramidal, Mariam Khayretdinova from Brainify, Qingyu Zhao from Weill Cornell, Joe Futoma from Oura, Eric Trautmann from Meta Reality Labs, and Surya Ganguli from Stanford and General Catalyst. Panelists discussed the unique funding challenges faced by neurotech, which often needs more time than the 5-10 year timeframe where most VCs want a return, while going beyond the scope of traditional grantmaking, leading to a frequently mentioned “valley of death” effect. In practice, the field’s funders are almost always true believers, who see neurotech’s potential in a way that others don’t.

A few rapid-fire polls read the room. Participants favored better sensors over more data as a way to advance the field, and were more split about whether it was best to optimize models across individuals, or across contexts for one individual.

A keynote from Andreas Tolias, titled “A Less Artificial Intelligence,” focused on the other side of the question: how better models of the brain could inform AI, by uncovering the principles behind natural intelligence. It was a prime example of the wide-ranging nature of the event, bringing together people with a variety of goals into productive conversation.

As the evening closed, conversation was in full swing. The event ended with a reception, where groups met in breakout sessions to discuss more specific topics. Neuroethicists and law experts talked about what governance should look like in the NeuroAI ecosystem. Researchers discussed the signal requirements for decoding motor control on the scale of the whole body, covering a variety of tradeoffs in the needed training data, including whether they should aim to measure intentions or actions. One group focused on EEG, where a key question was whether it was possible to gather enough consumer data, which depends in turn on whether there are BCI applications useful enough that enough people will be eager to adopt them. Data quality was a major topic, and there was broad agreement that better data was more useful than more low-quality data, and that it was important to have data that could generalize across tasks.

As night fell over NYC, the conversations continued. Protocol Labs sponsored post-reception drinks in a rooftop bar around the corner, and afterparties stretched into the early morning, with free-wheeling 5am chats.

The Global Neurotech Hackathon

Early-morning chats were a feature of Global NeuroHack too, though of a busier kind.

The event took place over three days, facilitated by organizers from not just Imperial College London, but universities around the world. Over one hundred students participated, attending workshops from expert mentors before they plunged into a 48-hour rush to build something extraordinary.

In our track, we challenged the students to build something that’s been on our minds lately: a system to take brain data gathered from non-implantable sensors and get out structured communication, like text. The students could use public datasets, but they also had the opportunity to gather data directly, with EEG and fNIRS headsets that they had access to thanks to the Hackathon’s sponsors. We helped recruit mentors, like David Moses and Anshul Kaashyup, to give the students the best start we could.

The results exceeded our expectations.

The winning team, CereBro, brought together students from the École Polytechnique Fédérale de Lausanne and the National University of Singapore. Showcasing the power of brain foundation models, they built a fine-tuning pipeline that linked together multiple EEG backbones into a single end-to-end system to decode inner speech. Their project impressed us with its high technical depth, an amazing accomplishment in such a short amount of time. The hackathon was great for the team as well, with David Zhang, the team’s NUS member, commenting that it was an “absolutely amazing experience, it’s a privilege to have this opportunity to compete alongside some of the most talented students and receive guidance from the best experts in the neurotech field.”

The second-place team, FABLE, displayed stunning creativity. Composed of students from UT Austin’s Longhorn Neurotech, FABLE prototyped a BCI-powered fairytale storybook, using EEG and narrative context to decode a story in real-time. They envision their system as a tool for aphasia recovery and language acquisition, helping those who struggle with language to grasp at what they mean to convey.

The rest of the event was full of amazing ideas as well. One team proposed vagal nerve stimulation for cows. Another, a safety app using EEG wearables, that got the team out into the street asking if passers-by wanted to join their “hivemind”. Others built anti-anxiety glasses, envisioned genetically engineering light-sensitive neurons, and built a neural spike-based guitar hero clone.

Overall, we were impressed by the students’ eagerness and drive to learn. None of the teams were scared to ask questions, and they took full advantage of the opportunity to meet with mentors and representatives from neurotech companies. It was an immensely dynamic and rewarding weekend.

Fielding-Building, not just a phrase

We have a tab on our website called Field-Building. It’s a term that may be unfamiliar to people outside of the worlds of funding and activism. But it’s more literal than it looks. Field-building is what we do to build research fields. Not from scratch, of course…but bigger, and better.

That research is the first step. We get experts talking, and students learning, because we want the world’s most talented minds focused on the problems that matter most. We want a world full of people who address tough challenges and find ways to make progress. We want to encourage the kind of people and ideas that shape an environment where new things can be built.

David Moses was one of the expert panelists talking at Foundation Models for the Brain, and mentored students at Global NeuroHack. He had this to say about the events:

“It’s a pleasure to be part of the community that e184 is fostering. I expect great things to come from this interdisciplinary group as we each contribute to the advancement of neurotechnology and neuroAI!”

We’re proud to have supported so many great conversations in this community, both at Foundation Models for the Brain and at Global NeuroHack. We can’t wait to see what you all build next.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

One of life’s most universal needs is the movement of fluids. From single-celled organisms in the waters of the primordial seas to blood and lymph in multicellular creatures, fluids bring in nutrients and carry away waste. In the human body, each cell and organ is bathed in a regular flow of blood from a network of thin capillaries, a process called perfusion.

Perfusion is life’s solution for how to keep cells alive. These days, it’s often technology’s solution as well. Perfusion engineers have revolutionized multiple areas of the life sciences, advancing beyond older methods to keep living tissue refreshed in a life-like way. These exciting developments are ongoing, as new applications show just how far life’s lessons can take us.

Perfusion for Cells

When the average person imagines a biologist, they often picture a petri dish. For over a century, biologists have used petri dishes to grow cells in what is called static culture. The cells multiply in a nutrient-filled medium, which must be occasionally replaced as it runs out of nutrients and fills up with the cells’ chemical waste products.

Static culture works fine for a biology class in school, but if your aim is to keep cells alive for weeks or even months, for example to observe how they build structure over time, then the method has real limitations. Each time the medium around the cells is replaced, there is a risk of contamination. The dramatic changes also shock cellular systems, causing problems for more delicate cells.

These days, perfusion systems offer another way. Instead of replacing the liquid around the cells once in a while, microfluidics are used to keep liquid flowing continuously. The result is much more lifelike. Cells can be viewed growing and living without disruption, giving a much clearer picture of how they behave in real organisms. The flow of the fluid can be adjusted to resemble blood flow, and because the chamber never needs to be opened to swap the medium, there can be much tighter control of temperature and pH while reducing potential contamination.

These capabilities have proved crucial for some types of research, enabling ongoing breakthroughs. Researchers interested in treating neurological disorders need to understand how drugs can cross the blood-brain barrier, but cell cultures without realistic flow cannot mimic the core way it functions to let different chemicals pass to the brain. Microfluidic perfusion has been a game-changer for this research. This approach has also given rise to new techniques, by revealing how fluid flow can prime stem cells to differentiate into the different types of cells that make up different tissues in the body.

Perfusion for Tissues

Up from the scale of smaller cell cultures, perfusion has enabled progress in engineering tissues like muscle and bone. As on the micro-scale, this progress has built upon how life itself supplies tissue with fluid. Here, the key insight comes from imitating the body’s interstitium, a layer that lies between skin and organs. Supported by a honeycomb-like web of collagen, the interstitium is filled with flowing fluid. This interstitial flow drains into the lymphatic system, the network of channels that coordinates flow of nutrients and waste, and immune defense.

By imitating this rate of flow, researchers have dramatically improved their ability to culture tissues at life-like densities. Without flow, tissues often develop with just a thin layer of living cells around an empty interior. The cells can have dramatically different metabolism than their counterparts in living creatures, for example muscle cells that only metabolize anaerobically (without oxygen), not aerobically (with oxygen).

By instead culturing the cells on a sponge-like structure supplied with realistic interstitial flow, researchers have been able to create dense tissue samples several millimeters across, with life-like structure throughout. This approach grew in popularity in the 2000’s. Scientists continue to find new applications, with recent developments highlighting the importance of flow for culturing bone. Now, groups are developing systems to make the technology portable, enabling more widespread use.

Perfusion for Organs

Organ transplantation is a miraculous medical procedure in some ways, but it is also something doctors have been doing for a very long time. The first organ transplant, of a thyroid gland, occurred in 1883, and other organs followed shortly after. Since then, procedures and logistics have modernized, and nationwide networks connect donors to recipients.

These vast logistical networks are needed, in part, because organs do not last long outside of the body. Human bodies are interconnected machines with every part serving a role, and an organ outside of the body does not stay alive long. For decades, the only solution was static cold storage. Organs are kept as cold as possible to slow down their metabolism, keeping death and damage at bay. The result is a few hours of grace: six or eight hours for organs highly dependent on blood flow, like hearts and lungs, with a bit more time for more resilient organs, especially kidneys. Some organs survive this window, others don’t, and it can be difficult to tell whether a transplanted organ will thrive.

Increasingly, machine perfusion has offered a better way. The idea is an old one, with experiments in animals as far back as the 1930’s and explorations in humans in the 1960’s-1980’s, but only really became feasible in the last few decades. Now, instead of keeping organs on ice, it is increasingly standard to use machines to supply fluid continuously to keep organs alive. This can be done hypothermically, keeping the organs cold for the same reasons as in static cold storage, or normothermically, with organs operating with almost the same metabolic rate and temperature they would have in a living body.

Normothermic perfusion systems require substantial engineering, but come with equally substantial advantages. By keeping organs under lifelike conditions, it becomes possible to assess their potential, checking for damage and observing how they function. Perfusion also keeps organs alive longer than static cold storage. In combination, the result is that more donations can be used in time, saving more lives. The method has continued to show its value, and ongoing studies are assessing the procedure in new contexts.

The Insights that Keep on Giving

By building perfusion systems, engineers have taken life’s lessons to heart with modern technology. Whether it’s maintaining realistic flow in small cultures of cells, mimicking the body’s interstitial layer to grow tissues, or keeping whole organs alive and pumping, perfusion engineering has created new options for researchers and clinicians in the life sciences. Every day, this technology advances, and we all can do that much more.

So if you know an engineer who works with perfusion, give them a shout-out. We’d love to hear what they’re building.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

That progress builds on a rich history. For a long time, biologists tested ideas as they had them, in painstaking experiments carried out by hand. That changed in the 20th century, as broad searches for cancer drugs showed the power of a more systematic approach. Companies picked up on the idea and began to mechanize the process, creating what is now known as high-throughput screening.

Now, high-throughput screening is a core technique, speeding up research across the life sciences. Advances in fields like robotics, computer vision, and microfluidics let scientists carry out millions of tests in a span of time when their predecessors might have only done one.

In the last few years, the game has started changing in remarkable ways. High-throughput screening is being used in more lifelike conditions, with new techniques and technologies emerging to approximate the variable, messy, 3D world of the human body. It’s a time of remarkable challenges, and equally remarkable opportunities.

The First Screens

In 1960, when the US’s National Cancer Institute started testing plants for cancer drugs, many scientists were skeptical. Surely, if there were promising treatments out there, people would have already found them?

What critics didn’t reckon on was the sheer scale of the program. Between 1960 and 1986, the institute scanned 114,000 samples from 35,000 plants. Most were useless, as one might expect. But a few could properly be described as miracle drugs. Taxol may be the most famous, a compound derived from yew bark that radically improved chemotherapy for breast cancer and ovarian cancer. By testing on a massive scale, the researchers were able to find solutions scientists would never have thought of in the old approach.

The National Cancer Institute screen was extensive, but it wasn’t fast. Companies that wanted to replicate that success would need to figure out a way to speed up the process.

Starting in the 1980’s, they figured out how. Originally, experiments were performed in test tubes, each holding one mililiter of liquid, limiting labs to perhaps 50 tests a week. Companies moved from this bulky setup to “microplates”, plastic rectangles with 96 wells that serve as tiny test tubes. They developed new systems that could distribute drops of liquid measured in microliters. When these systems started being put into production in the early 90’s, labs jumped to being able to test thousands of samples per week, and the era of high-throughput screening was born.

The Throughput Gets Higher

As technology improved, labs were able to handle more and more tests in parallel. Microplates themselves changed, subdividing the 96 wells from the original into double, then quadruple the number of tiny test tube-like spaces. Now, companies even use 6144-well plates, with sixty-four times as many spaces as the originals.

Each well contains its own little experiment. Sometimes, the results of these experiments can be measured in simple ways: for example, by shining light through each sample and seeing how much is absorbed. More complicated measurements used to need to be done by hand, with researchers looking at each sample in a microscope to identify features of cells. Increasingly, though, this can be done by machines too: using AI, a technique called high-context imaging can work from images of cells, taking account changes in their shape and layout and boiling the information down to numbers scientists can use.

The experiments themselves are carried out by robots. Robot arms take plates from station to station, adding compounds, incubating, and reading out the results. Already in the mid-2000’s, these robots were capable of screening 100,000 compounds per day, around the same number that the National Cancer Institute spent twenty years analyzing a few decades before. As more advanced AI leads to more coordinated and versatile robots, the speed of progress will keep increasing.

Other advances push the pace further. Microfluidics, engineering processes that manipulate increasingly tiny amounts of fluid, have led to even faster machines, including some that can process hundreds of thousands of drops of liquid per second. Other techniques can sometimes let multiple compounds be tested in the same well, by arranging multiple tests so that scientists can infer which compound made the difference.

Lifelike Conditions, Warp-Speed Solutions

In the world of biotech, we don’t just want to find out how an experiment behaves in a test tube. Biotech companies develop treatments and techniques for human beings, that must work in, or become part of, human bodies. That means for many screens, the more lifelike conditions can be tested, the better.

An important shift was a move from 2D, to 3D. The plates used in traditional high-throughput screening were like old-school petri dishes, hosting a thin layer of cells. In human bodies, though, cells don’t just lie on a plate: they pile on top of each other, exchanging nutrients and chemicals. A treatment that works in a dish might not work in a real organ. To address that, researchers moved to new types of plates that work in 3D instead. Each well became a compartment capable of holding a 3D “organoid”, fed by tiny channels that carry droplets where they need to go.

Working in 3D brings new challenges. The cell aggregates have to be carefully coaxed to a uniform size and shape, so that the tests don’t give misleading results. 3D cell cultures are also tough to image, unlike 2D which can be treated like a slide in a conventional microscope. This requires using a technique called confocal microscopy, which blocks out-of-focus light to collect images of tissue layer by layer. This upgrade has costs: confocal microscopy can be more laborious and time consuming, making use in high-throughput contexts more challenging.

Increasingly, groups are addressing these challenges in systematic and automated ways. An impressive paper by a Connecticut-based group in 2024 showcased an automated screening pipeline, which they used to test cancer drugs on organoids. They managed to automate almost every step, using robot pipetting to set up the organoids, stain them with fluorescent dyes, and set up assays, and a confocal microscopy-based high-content screening system system to automatically gather images. They had to develop new workflows and setups, figuring out how to keep the Matrigel used to grow organoids at the right temperature so it could be dispensed smoothly from automated pipettes. The result was a procedure that was not just consistent and efficient, but sensitive, able to detect subtle effects of low concentrations of drugs in a way that would not have been possible with more traditional chemical measurements.

When confocal microscopy is used in a high-throughput context, the system collects layers and layers of images, and that creates its own challenge, in turn: data. In order to usefully analyze pictures of layer upon layer from thousands of wells of cell cultures, scientists need advanced AI tools, capable of processing huge amounts of information to identify key traits. As AI gets better and better at analyzing images, this goes from a daunting challenge to a real possibility.

As an example of recent progress in this area, a group at the US Naval Laboratory published a new approach last year that promises to identify the boundaries of cells much more efficiently than previous approaches, a task called cell segmentation. Other AI approaches to cell segmentation tended to use a deep learning architecture called convolutional neural networks, which can require over a million cell images to train, and demand human input to both label images and choose parameters for the algorithm. In contrast, their approach uses a technique called self-supervised learning, which allows the algorithm to learn traits of images without labels. It focuses on traits of the images they can characterize mathematically, like one called optical flow vector fields, creating a more transparent model that lets users check why it chose to put the cell’s boundary in a particular place. The result was versatile, working for both live and fixed cells, different microscope types, and over a wide variety of scales. It was also fast and efficient, running in only a few minutes on a laptop. The authors expect their approach to be much easier for scientists to use, saving time and not requiring extensive machine learning expertise.

Realistic cells bring another challenge: they don’t always do what they’re told. When scientists perform experiments on cells with gene editing techniques like CRISPR are, the cells’ genes drift from generation to generation. This can cause problems for high-througput screening, as a test in a single well of a microplate may not actually reflect the intended genetic perturbation. In the past, the only way to deal with this was through statistics: test hundreds of cells with the same modification, and look at how they behave on average. This is already expensive if each well contains a few cells, but when the wells contain 3D organoids, or when the cells need to be grafted into living animals, you introduce additional noise, as cells randomly grow or thrive in living environments.

A 2024 paper proposed an interesting solution. A collaboration between groups at Vienna BioCenter and Mount Sinai Hospital, Toronto used unique molecular identifiers as a kind of single-cell barcode to track cells in a tumor grafted into a mouse from generation to generation. Using a method they had developed called CRISPR-Switch, they were able to modify cells in a way that would turn on or off randomly when triggered. This let them create a kind of internal control group, where each part of the tumor that contained cells with a specific modification also contained similar cells with the modification turned off. By checking cells with each barcode, they could better distinguish which cells died out due to random chance, versus which had their growth suppressed by the CRISPR modification. They call the new method CRISPR-StAR, where StAR stands for Stochastic Activation by Recombination. They build a versatile molecular library to implement their method, which gives much more reproducible results than older noise-prone approaches.

High-throughput screening is tremendously powerful. Making full use of that power has given rise to new challenges, as scientists find ways to adapt more and more types of experiments to function at greater scale and speed. It’s always inspiring to see how different groups around the world bring their own creative approaches to these problems. Science moves forward one step at a time, but a clever idea or an elegant experimental design can move a whole field in ways that few other things can.

At e184, we are excited for what the future holds in store.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

Cutting-edge science happens across the country, not just in the biggest research centers. A small city with the right infrastructure can be the best of both worlds: a place where a company or scientist can carve out their own niche, where a strong ecosystem of suppliers and researchers coexists with high affordability and solid public amenities.

Portland is that kind of place. Stretching from Intel’s global R&D headquarters in Hillsboro to Vancouver, Washington on the other side of the Columbia River, the Greater Portland Area has attracted biotech companies like us with its highly educated workforce and prominent universities. We weighed several factors, including operational costs, regulations, and the presence of key resources and expertise. All together, Portland was the clear choice.

If you’re a scientist, you may be weighing your own factors, deciding where to make the next step in your career. Our CSO, Aleksei Mikhalchenko, faced that same choice. In this piece, he explains why he chose Portland, and why it might be a great step for you as well.

Let me start with something you probably didn’t expect: I love the nature here.

I know. I’m a scientist. I’m supposed to open with a landmark paper, a funding milestone, or at least a sentence about translational opportunity. But when I think about why I stayed in the Portland area - why I’m still here after seven years, why we’re building e184 Repro here - the landscape is genuinely part of the answer. And I think that’s worth explaining.

I came to Portland from Boston, where I did my PhD. The Pacific Northwest is extraordinary in a way that takes a little time to absorb. Portland sits at the edge of something vast: you are an hour or two away from Mt. Hood, an hour from the Columbia River Gorge, two hours from the Oregon Coast. Forests begin where the city ends. The rivers run through the city itself. And what has made all of it meaningful to me - in a way that the Rockies or the Sierras would not have been - is the climate. Mild winters that don’t strand you indoors for months, summers that are warm but not punishing…but at the same time, actual seasons that change and give the landscape texture throughout the year. For someone who needs to step outside and feel something shift - and I think more scientists need that than we admit - Portland makes it easy to do that on a Tuesday evening after work, not just during a planned vacation.

But I’m not writing this post to talk about nature. I’m writing it because we’re an early-stage biotech startup doing serious work on human fertility - developing treatments using pluripotent stem cells and cellular reprogramming - and when talented scientists see our job postings and notice the location, I imagine many of them pause. Portland? Really? The biotech map most scientists carry in their heads has San Francisco, Boston, San Diego, Cambridge, and maybe Seattle on it. The Portland metro is not on that map.

But there’s more to Portland than you may know. I came here for specific reasons, and I’ve stayed for specific reasons - and those reasons aren’t mine alone. They’re reasons that I think should matter a great deal for any early-career scientist, thinking carefully about where to build their life and work.

The regulatory reality that most people don’t talk about

I moved to Portland specifically because of the science I wanted to do.

Throughout my career, my research has involved human gametes and embryos. When I was finishing my PhD and looking for a postdoc, I spent a long time mapping which states in the US would actually allow me to do this work. The regulatory landscape is more restrictive than most people realize. Massachusetts explicitly prohibits the creation of human embryos through fertilization for research purposes under state law. That’s not a minor bureaucratic hurdle - for someone working in reproductive biology and embryology, it’s a ceiling on the science itself. New York and Oregon were among the very few states where this work could proceed without those kinds of restrictions, operating under federal guidelines rather than additional state-level prohibitions.

That’s why I came here. And it’s worth saying plainly: if you work in reproductive biology, embryology, stem cell biology, or any adjacent field where access to human gametes and embryos is central to your research, the state you work in is not a background detail. It is a scientific constraint, one that can unexpectedly block you from finding what you need to know. Oregon removes that constraint.

For companies in our field the regulatory ceiling is real. A company that achieves promising proof-of-concept results in Boston may find itself needing to relocate before it can take the next step, for example if it discovers it needs to create embryos to move forward. We don’t have that problem. We’re building here on purpose.

The infrastructure you don’t expect to find here

Portland’s biotech ecosystem grew around a few key institutions. The infrastructure they gave rise to makes it an especially good place for companies in our field.

Just outside Portland, in Beaverton, sits the Oregon National Primate Research Center - the first of the seven federally funded National Primate Research Centers established by the NIH, founded in 1962 and affiliated with Oregon Health & Science University (OHSU) since 1998. Its Division of Reproductive and Developmental Sciences has maintained one of the most specialized primate reproduction research programs in the world: gamete function, embryogenesis, IVF, implantation, oncofertility, and assisted reproductive technologies in non-human primate models. This legacy has seeded the area with relevant know-how.

ONPRC is just one part of OHSU’s scientific footprint. I spent six years at OHSU - first as a postdoc, then leading bioinformatics research, and most recently as an Assistant Professor - and the ecosystem that institution anchors is real. OHSU runs hundreds of active clinical trials and functions as much as a hospital as a research university, creating a concentration of clinical expertise and human research infrastructure in the area that is genuinely rare. Tapping into that infrastructure means that we have access to all of the on-the ground expertise needed to support good science, from doctors experienced in clinical trials to suppliers that know how to outfit a world-class lab space.

And while Portland’s biotech scene is small, it has real substance. The area is home to over two hundred biotech companies, employing over seven thousand people, including outposts of major players like Thermo Fisher and Genentech. The Oregon Bioscience Incubator works to grow that community, and has supported over sixty early-stage biotech startups, at least one of which has since gone public. Other companies in the area, like Twist Bioscience and Intel, have facilities with precision requirements that demand reliable local supply chains and engineering infrastructure, shaping the kind of environment startups need to build fast.

None of this makes Portland metro equal to Boston or the Bay Area in scale. I’m not arguing that it does. But the specific things that matter for our work - regulatory conditions, primate reproductive biology expertise, a growing biotech community, deep engineering infrastructure - are here.

What your salary actually buys here

When I came to Oregon as a postdoc, I ran the numbers. Postdoc salaries at most biomedical institutions are anchored to NIH NRSA stipend guidelines - a national benchmark that does not adjust for local cost of living. In practice, institutions in high-cost cities sometimes pay above that floor to remain competitive, but the gap in nominal salary between a postdoc in Portland and one in Boston or San Francisco is far smaller than the gap in what that salary actually covers.

As of 2025, the median home price in Portland is roughly $540,000. In Seattle, it’s around $790,000. In San Francisco, it’s over $1.2 million. Rent for a one-bedroom apartment runs about $1,700 per month in Portland, compared to $2,300 in Seattle and more than $3,000 in San Francisco. Oregon has no sales tax.

For an early-career scientist - someone finishing a PhD or a postdoc, taking their first industry position, maybe thinking about starting a family - these numbers are not academic. They determine whether you can afford to buy a home within a reasonable commute of your lab. They determine whether one salary is enough to cover childcare while a partner finishes their own training. The same compensation package buys a fundamentally different life depending on where you live.

What the big hubs won’t tell you

There is something that happens to scientists in Boston and the Bay Area that I don’t think gets talked about enough: many of them are exhausted. The density of competition, the cost of living, the churn of a hyperactive biotech market - it creates an environment where people move from company to company every eighteen months, where visibility is hard to earn, where burnout is common and normalized.

The Portland metro is not like that. The people who come here and like it tend to stay. There is less turnover, which means more institutional memory, more stable teams, more room to actually build something over time rather than perpetually restart. As a scientist early in your career, the question isn’t just which city has the most companies. It’s where you will have the most room to grow, to contribute meaningfully, to be seen.

In a startup here, you are not one of ten thousand scientists doing similar work within a ten-mile radius. You have real ownership over the direction of the research. What you discover matters to the trajectory of the company in a way that is legible and immediate. That kind of environment is where careers are built, not just managed.

What Portland is not

I want to be honest about the trade-offs, because a scientific audience deserves that.

The Portland metro is not one of the largest-scale biotech hubs. The network density of Boston or the Bay Area - the informal conversations, the conference visibility, the sheer volume of peer companies and investors - is not something Portland can currently match. If you are at a stage in your career where that density is what you need, I won’t pretend otherwise. Some people need to be in those places, and it makes sense for them.

But if you are at a stage where you want to do work that matters, in a place where you can afford to live, surrounded by nature that actually restores you, on a team where your contributions shape the science rather than disappear into it - this corner of the Pacific Northwest is worth a serious look.

I have lived here for seven years. I chose it for scientific reasons. I stayed for the same reasons, and more. If this place sounds like somewhere you could thrive - scientifically and personally - we should talk.

As we build the technologies that matter most, we look forward to welcoming more talented people to the Pacific Northwest. Check out e184.com/careers to find out how you can get involved.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

The road to artificial wombs is a long one, demanding broad scientific knowledge. One area of research that could make a difference is that of stem cell embryo models: assemblies of cells grown to imitate key stages of embryo development. As researchers pursue models of later stages, stem cell embryo models have deepened our understanding, giving us a clearer picture of what it takes to support an embryo as it becomes more complex. By generating embryo-like structures that can be used in the place of actual embryos, embryo models have allowed researchers to directly study development while reducing the need for human embryos.

The growing impact of stem cell embryo models is why we supported Keystone Symposia’s conference last week, Stem Cell Models in Embryology. It’s also why we sent Joseph Owen, the Director of our Program for Artificial Wombs. We wanted to take the pulse of the field, to get an impression both of where the science is, and what researchers feel are the next steps forward. In this post, we’ll talk about what we learned.

The State of the Field

As a research tool, embryo models continue to be powerful. The conference highlighted ways that embryo models have contributed to progress on questions in biology, particularly in the area of gastrulation, where the embryo makes its first steps towards developing the three germ layers necessary for body plan development.

On a more direct level, though, progress appears to be slowing. As blastoids and embryoids have improved, researchers have been able to replicate later and later stages of embryo development with higher and higher fidelity, and it was natural to wonder how far these models could go. Now, though, the field seems to have hit a wall. Blastoids model pre-implantation stages, and are limited in how well they can do beyond implantation. Embryoids, modeling post-implantation stages, can progress to gastrulation, but there is little progress beyond that stage. Organogenesis is still out of reach. Participants discussed why, with several explanations suggested but no common consensus.

While progress in embryo models has moved very little over the past five years, endometrial models have begun to make impressive strides. Three blockbuster results were published in the last month alone, in which several multi-lab teams independently managed to build a 3D co-culture system for embryo and endometrial models, recapitulating key developmental milestones. One group was even able to use their setup to test compounds with the potential to treat implantation failure.

The poster session featured encouraging work by young researchers. One impressive example was the poster, “The Regulation of Human Presomitic Mesoderm Differentiation by the Hippo Signaling Pathway”, which explores how presomitic mesoderm differentiates into what will become the vertebral column, regulated by the HIPPO pathway, and how abnormalities in this pathway could be responsible for severe congenital spine malformations. This work was one of those highlighted by attendees, and we selected it for our $500 Poster Award.

The State of the Conversation

Not all of the discussion at the meeting focused on scientific topics. Any field touching on something as fundamental as reproduction faces ethical questions. How to approach those questions was a major focus of the scientists in attendance, who are well aware that the field faces some tough decisions.

Some of the discussion centered on concrete ethical questions. While some stem cell models use induced pluripotent stem cells, somatic cells from adults reprogrammed into a pluripotent state, others use cell lines derived from human in vitro fertilized embryos and fetal placental cells from aborted tissue. Some members of the public have mixed feelings about this, but it’s not obvious that the alternative, destroying the material rather than using it for research to advance reproductive medicine, is actually an improvement.

For embryo models, a recurring question is how long, and how far, they should be allowed to develop, a topic where there is currently no real consensus. In essence, the question is that of how “realistic” embryo models should be, whether there ultimately ought to be firm dividing lines between them and natural human embryos. This depends crucially on how the models are used, and the motivations for building them. While it is intuitive as a scientist to be curious as to how close embryo models could be to the real thing, the question of whether there is scientific value in reaching a particular stage is one that needs to be actively asked.

There were questions even tougher than these, though. The hardest question is not about any individual ethical issue, but about public understanding of the field.

Fundamentally, the field is conflicted over how these conversations should be carried out. When does it make sense to involve stakeholders from outside the field, or to involve the general public in the conversation? On what timeline do ethical issues need to be addressed, and decisions reached?

There is a worry, a quite justified worry, that the political climate will make these conversations difficult. The field is vulnerable to regulatory shifts or public backlash. For the most part, researchers at the conference seemed pessimistic, expecting a future backlash and not expecting to be able to defuse it. For now, many seem to hope in something akin to security by obscurity, hoping that using the right terminology will forestall public opposition.

We, in turn, are skeptical that obscurity is the right path forward. We see parallels to the early history of IVF, when Robert Edwards had trouble finding funding due to his colleagues’ antagonism to his engagement with the public. Ultimately, the public formed their own opinions, and the fertility community was unprepared for the debate that followed.

Conclusions

There were worrying things about the overall mood at Stem Cell Models in Embryology. Instead of new advances in blastoids and embryoids, much of the discussion focused on comparing and hashing out traits of older models, establishing which existing models are more true to nature rather than working to advance them. This negative focus, coupled with pessimism about conversation with the public, make it natural for researchers to feel that the field is in a difficult place.

At the same time, we see a way forward. The surge of progress in endometrial models is enormously encouraging. It shows not only that science continues to march forward – but that it does so by exploring under-explored questions, in a way that will directly and clearly contribute to reproductive health. In particular, the work identifying compounds to treat implantation failure shows that the areas that are making progress are those with the most potential benefit, the areas that can help families find solutions they desperately want.

In that potential, we see the seeds of a modus vivendi. We can approach ethical questions not with obscurity and worry, but with an eye to the future. We can value research, and decide what we do or do not need to accomplish, on the basis of its potential benefits. And we can bring the public in, tentatively, on that basis: not as opponents whose backlash is inevitable, but as people who can be reassured if they see us focused on goals and values they share, on their survival and flourishing and their hand in the future.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

We’re just starting our journey. We will be talking with experts to hammer out our approach, and hiring talented researchers who can make our dream a reality. At the same time, we’re laying groundwork for our device in other ways. We need to think about how we build our technology: its shape and layout, what regions of the brain we need to measure and with what resolution. And we want to start thinking about data, about the measurements needed to train a machine learning model for the decoding tasks we need.

Focusing on speech, these two questions have a shared core: what, exactly, will we decode? Speech isn’t a single phenomenon. It’s a process, one with different levels, from our earliest ideas of what we want to say to the concrete movements we make when saying it. As we plan, we investigate which of these levels give us the best path to a brain-AI interface.

The Articulatory Level

At the most concrete, we could try to measure articulation. When we speak, we move muscles in the tongue, lips, and jaw, and exercise control of the larynx. A decoder could try to represent those movements directly, inferring their kinematics, or could aim at phonemes, the most basic sounds we use to distinguish words.

Aiming for articulatory data would mean a focus on those brain areas that most directly relate to the muscle movements of speech. The precentral gyrus sends movement signals to the spinal cord, and its ventral region contains the orofacial primary motor cortex, which controls core functionality for speech, including loudness. The adjacent premotor cortex is involved in planning movement, and its signals correspond to more complex patterns. Successful approaches that target articulatory data generally measure signals from these regions. This can involve looking for analogues to the signals observed from muscles in electromyographic measurements, or by examining the high-gamma brain waves that have been observed to correlate well with the kinds of articulatory movements sought.

Targeting the articulatory level is intuitive for a device for paralyzed patients, who may send nerve signals intended to produce speech and be unable to produce it. It seems less useful for our goal of a device that can be used silently by healthy users. It also may be more vulnerable to muscle artifacts, especially if we wanted to use it in a non-implanted approach, as current approaches rely on features with high frequency and high spatial specificity. This may make this approach inherently limited to the implanted approaches currently used to target it, and thus less desirable from our perspective.

The Acoustic Level

Another approach would be to target the sound itself, mapping brain states to a waveform or spectrogram representation. This involves targeting the representation of sound in the auditory cortex, with the superior temporal gyrus involved in processing higher-level features of sound (turning phonemes into linguistic data) and Heschl’s gyrus appears to have an important role in imagined speech, particularly dialogue with oneself.

While some research studies acoustic representations in order to understand how the brain processes auditory information in general, other studies, based on successes decoding imagined speech, see the potential for a brain-computer interface. Typically focused on paralyzed patients, interfaces along those lines create synthesized speech, for example with a vocoder representation or using a self-supervised approach to identify key auditory features.

As we prioritize enabling users to compose text and communicate with AI, the specific vocal waveform is less relevant for our use-case than it would be for paralyzed patients who may be interested in speaking in something approximately like their own voice. It is also harder to evaluate and model, being a very high-dimensional representation of speech data.

The Linguistic Levels

Moving up in abstraction from movement and sound, we can ask about the actual words someone intends to convey (lexical-level speech), or work at the level of parts of words like syllables or even individual phonetic elements (sublexical-level). Speech at this level is spread more widely in the brain, and different approaches have targeted areas of the sensorimotor cortex and the perisylvian cortex.

These approaches tend to use features of language models as features to structure data, like word logits or n-grams, and to aim to produce likelihoods for classifying phonetic elements. The end result, unlike the previous two layers, can be interpreted directly as text, which is an appealing trait for our goal of a brain-AI interface.

While linguistic approaches line up well with our goals, they have so far been challenging targets. The most successful approaches still heavily depend on particular conditions: they have trouble transferring between subjects, or they work only for a limited vocabulary or a particular task structure or protocol. This has especially been an issue for non-implanted approaches.

The Semantic and Sentence Levels

Finally, one could imagine targeting speech at its most abstract, at the level of meaning. This could mean trying to reproduce whole sentences, or even broader concepts. Neural data at this level is more widely spread through the brain, and studies have found useful signals in large-scale networks like the default mode network, the frontotemporal language-selective network, and the visual network. Targeting this level would thus require fairly broad brain coverage.

Approaches to this level can train on data gathered from reading stories, and classify the input with computational tools that capture semantic content like contextual language model embeddings and semantic vectors. Some work at this level results in a ranking of candidate sentences, but other work yields a paraphrased text.

Such a paraphrase would be more relevant to our goals, but would come with clear downsides for users, who might feel that their words are not being accurately represented. It also raises privacy concerns, as a device trained purely to pick up meaning could pick up meanings that the user does not intend to communicate.

Conclusions

Our analysis suggests that the articulatory and acoustic levels, while important for medical devices, will not be as well-suited to our goals. Some mixture of linguistic and semantic content, instead, seems ideal. Data at the semantic level could patch over the difficulties purely linguistic-level approaches have in non-implanted modalities, while lexical or sublexical data is needed to make sure that the device accurately conveys what the subject actually intends to say.

With that said, at this stage, we shouldn’t rule anything out. We want to build a versatile device, one that can target a wide variety of brain regions. We don’t just want to reproduce existing successes of implants in a non-implanted modality. While that would already be a great success, the field is already showing some progress in that direction. What we want, instead, is a platform to build on, one that gets us closer to our ultimate goal of extending human capabilities, fluidly incorporating AI into our own cognition to safeguard our voice in the future. Our best path at the present time may still be a familiar one, growing out of current-day successes…but it could also be something radically new.

As we investigate the way forward, we especially want to discuss different modeling approaches: different sources of data, different embeddings, different targets and architectures. Soon we will be reaching out to expert contacts, getting multiple opinions to help us find the best way to begin. If you’d like to join that conversation, or to work with us on other frontier BCI applications, contact Peter Zhegin at p@e184.com.

And if you’ve got your own thoughts on the topic more broadly, we’d love to hear them. What levels should we prioritize for speech decoding, and what methods should we use? Let us know in the comments!

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

Right now, there is a technology that can already enable some women with absolute uterine factor infertility to give birth. Just as kidney and heart transplants give new possibilities to patients, uterine transplants offer a second chance to women who lack a functioning uterus, whether due to hysterectomy, congenital conditions, or other conditions that impact uterine functioning. Uterine transplantation is relatively new, and still quite rare and expensive. Its presence despite its cost and difficulty shows an unmet need. If we can broaden the range of childbirth options available without major surgery, we will brighten a huge number of lives.

History of Uterine Transplantation

While kidney transplants were first attempted in the 1930’s and succeeded only in the 1950’s, and heart transplants became possible in the 1960’s, uterine transplants were only pursued much later, with the first documented human attempt in the year 2000. It would be over ten years of research progress before a grafted uterus was able to survive long-term in the recipient, and the first successful birth from a transplanted uterus occurred even later, in 2014 in Sweden.

That success opened the way, and transplant centers began to establish clinical trials for the procedure. The first trials were in the US, and the four US centers that perform the procedure (Cleveland Clinic in Cleveland, University of Pennsylvania, University of Alabama at Birmingham, and Baylor University Medical Center) process the majority of cases worldwide.

Brazil followed in 2019 with another significant milestone, the first successful birth from a uterine transplant with a graft from a deceased donor. Now, some procedures use grafts from brain-dead donors, while others use living donors who volunteer for the procedure. Both approaches can be logistically challenging.

Over the last few years, uterine transplant programs have been founded in several European and Asian countries, as well as Australia. The technology is still young, and is gradually expanding in availability.

An Unmet Need

That gradual expansion has not met the demand. While most of the general public is not even aware the procedure is possible (for example, a survey in Australia found only a third of respondents who were otherwise aware of organ transplantation knew that uterine transplants had been performed), of those that are aware a significant number are interested.

Between 2015 and 2022, 5194 women applied for uterine transplantation in the US. The majority had undergone hysterectomy, while a substantial minority had a congenitally absent uterus. Of these applicants, only 37 received uterine transplants in that timeframe. While some recipients had undergone hysterectomy, the vast majority had a condition called Mayer–Rokitansky–Küster–Hauser, or MRKH, syndrome. In this condition, affecting roughly one in 5000 women, key structures in embryonic development fail to form, resulting in a total lack of a uterus from birth. This difference, between an applicant population where comparatively few had MRKH and a recipient population that mostly had the condition, may be ultimately due to different health indications in those with MRKH versus those who had undergone hysterectomy. It does suggest, however, that doctors view the procedure more as a treatment for congenital conditions and less in terms of its broader potential to restore reproductive choice.

One might guess that the procedure is so rare due to a limited number of donors, but while fewer prospective donors reached out to US uterine transplant centers than prospective recipients, the numbers are not as different as one might think. There were 2217 applicants for uterus donation between 2015 and 2022, almost half the number who applied to receive a transplant, and sixty times the number of transplants that were actually performed.

Rather, the difference is likely due to the simple fact that this is a new procedure, and still faces a number of challenges. It has substantial costs: estimates vary due to the difficulty of tracking costs in the US medical system, with lower estimates coming at slightly greater than the overall cost of surrogacy and higher estimates exceeding US $1M. It has meaningful health risks for patients: while less risky than transplant of life-critical organs like the heart or lungs (no uterine transplant recipients have died as part of the procedure), it still represents a significant surgery with all of the risks that involves. Finally, it poses substantial challenges to doctors, with only a few centers so far having the expertise to perform it.

Defying the Odds

In the face of both risks and costs, why do women still choose uterine transplantation?

Interviews and surveys in the UK, Australia, and the US show the incredible motivation and dedication that lies behind the decision. The women who opt for uterine transplantation are well-informed, and have considered many options, often including both adoption and surrogacy. Some even have prior experience with one or the other: of the UK cohort, 63% had considered adoption, 5% had attempted to adopt and 1% (2 participants) had already adopted children, while 76% had considered surrogacy, with 10% attempting to gestate with surrogate assistance and 1% having children in this way.

Nonetheless, for all of these women, uterine transplantation represented a unique opportunity, one valuable enough to pursue despite awareness of the risks.

For some the desire was deeply linked to their body, a wish to have the experience of gestation or to “be whole”.

Others were motivated instead to defy the odds, accomplishing something many had believed their whole lives was impossible. As one of the participants in the US study said, “It wasn’t that I’m doing this ‘just because’. It was because I was told I would never be able to do this.”

Others wanted to avoid surrogacy for cultural or ethical reasons, or because of uncertainty over its legal implications in certain countries. Ethical issues can be raised with uterine transplantation as well, including questions of how to fairly allocate scarce medical resources and questions of the rights of families of deceased donors.

Overall, every approach to enabling reproduction for women with absolute uterine factor infertility has its own complexities. It is indicative of just how much the choice to reproduce matters that so many people have pursued the opportunity for uterine transplantation, even with these questions and uncertainties in place.

Looking Forward

In developing artificial wombs, we are building another solution to absolute uterine factor infertility, a new perspective that offers many potential advantages. Artificial wombs will not require the complexity of organ transplantation surgery, let alone repeating this complexity twice for the donor and recipient. They will be based on technology that can be iterated and developed in tests on animals and stem cell embryo models, and produced under replicable conditions. That development process tends to lower costs, and there is a decent chance that, while artificial wombs will still likely be expensive, they will not be as costly as uterine transplantation.

It is clear, from the history of uterine transplantation, that such a technology will be valued. Women with MRKH may be the majority of uterine transplant recipients today, but a more accessible technology will also benefit women who have undergone hysterectomy, trans women, and more generally a wide variety of families who cannot or will not carry a child, but still want to reproduce. It will not deliver the physical experience of carrying a child, for those for whom that is the primary motivation. But it will enable more people to defy the odds, and do something they may have been told was impossible. It will enable more choice, the opportunity to make an immensely personal decision in what for many will be a more personally appropriate way.

We have a long road ahead of us. Over a decade passed between the first attempt at uterine transplantation and the first success. It may well be a decade before the first families can use our technology as well. But looking at the history of uterine transplantation, we know just how much benefit the end of that road can bring.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

In this piece, we interviewed Thorold Theunissen, an Associate Professor of Developmental Biology at Washington University in St. Louis, Missouri. Theunissen’s career spans these developments, from biologists’ first appreciation of the versatile power of stem cells to the development of stem cell-based embryo models that can shed light both on the hidden first weeks of embryonic development, and the crucial role played by extraembryonic cells. We asked him about the history of key advances in the field, his own lab’s work modeling the implantation of embryos in the uterus, and what he expects in future.

What follows is an edited and curated version of our discussion.

To start, why don’t we go back to the beginning? What first motivated you to study stem cell biology?

This goes back to my freshman days at Harvard. Doug Melton was there. He was one of the early pioneers in stem cell biology, and he had just launched the Harvard Stem Cell Institute. This was the first organization of its kind, really, a dedicated institute at a major research university focused on stem cells. The Harvard Stem Cell Institute had the resources to derive embryonic stem cell lines in the Boston area from IVF clinics, that would then be distributed to researchers around the world. So they were leading this early stem cell wave, and they just had phenomenal scientists associated with that institute. I was successful as a young undergrad to get Doug Melton to be my advisor, and I asked him, “Can you recommend any good lab opportunities?”

He very kindly wrote to a scientist in the Netherlands (which is my home country), Christine Mummery. She’s also a pioneer in this area. She works on cardiomyocyte differentiation from embryonic stem cells. And I was very fortunate to spend a summer in her lab, that was such a great experience. There was just so much buzz around stem cells, being basically a method to generate any cell type of interest in the body. And the cool thing with the cardiomyocyte is you get these beating areas. So it’s something you can actually tangibly see under the microscope, there’s contractions, there’s a beating heart almost.

Then I returned to Boston for, I think it was maybe my junior year, and Doug kindly arranged for me to talk to Stuart Orkin, who is a famous blood stem cell biologist. They had this ongoing project mapping the protein interactions in embryonic stem cells, and I was actually able to contribute to that. It was a Nature paper, just really exciting for a young scientist to get an opportunity to contribute to that level of science, and they got me hooked.

Then, as a graduate student in Cambridge, I was fortunate to join the lab of Jose Silva, a pioneer in induced pluripotency. That experience solidified my interest in stem cell biology and motivated me to pursue postdoctoral training in this field.

What was the path that led to the first stem cell embryo models?

The big question at the time was, “Can we create human stem cells that are more similar to mouse ES cells [embryonic stem cells]?” This was one of the fundamental problems in the field: mouse ES cells are really good at controllably differentiating them into different fates. If you put them back into a mouse embryo they will give rise to many different tissues. Austin Smith published this key review in 2009 with Jenny Nichols, where they argued you have these different stem cell states, naive and primed, and that was a nice conceptual framework to think about these different types of stem cells that we see in mice and humans. And the obvious missing piece in all of this was the human naive state.

What ended up happening is in 2014 I was able to publish in Rudolf Jaenisch’s lab these defined conditions to culture naive cells, and Austin’s lab published right around the same time. If you look at epigenetics, a cool sign of the naive state is they undergo X chromosome reactivation. Males have a single X, females have two, so you have this process of X chromosome dosage compensation where one of the X’s is inactivated in females. And this normally happens in a random manner. That’s how you have the spotted coat color of the calico cat. But what happens in very early development in the early embryo, the pre-implantation naive epiblast, is that in females both of those X’s are reactivated. And what’s really cool is that the typical primed human embryonic stem cell lines, they tend to show one inactive X. But when we apply these naive media, they undergo X reactivation. So they end up with two active X chromosomes.

But there was still kind of this nagging doubt in the field if these are truly naive stem cells. And that brings me right back to what e184 is interested in. In the past five to ten years, a number of groups have shown that you can really use these naive stem cells to not only make embryonic cells but also extraembryonic cells, and we were one of the first to do that.

After I started my lab at WashU in 2018, my first graduate student found that when we apply media for placental cells, remarkably these naive human stem cells can actually make placental cells. This was very surprising because in the mouse system, that’s not possible.

Mind you, there are different kinds of extraembryonic cells that support the fetus. You have the placenta, you also have the yolk sac that plays a key role in blood supply, and there’s another thing called the amnion, which forms an amniotic sac around the embryo. And subsequent work, not from us but by other groups, also showed that these naive stem cells can make yolk sac precursors. They can make primitive endoderm, they can make extraembryonic mesoderm, so it was really another exciting wave of papers showing that these naive cells are a starting point for extraembryonic differentiation.

The ultimate test of these cells came with the embryo models. Full credit to the people that led the wave here. First, Nicolas Rivron showed that you can do this in the mouse system. You can combine embryonic stem cells with trophoblast stem cells and make a blastocyst-like structure, a blastoid. Magda [Magdalena Zernicka-Goetz] did really cool work, in again the mouse system, combining embryonic stem cells, trophoblast stem cells, and primitive endoderm cells to make these first synthetic embryos, what she called ETX embryos. They’re post-implantation-like. So we realized that if the human naive cells are able to make placental cells in vitro and they are themselves like the inner cells of the embryo, the ICM, then theoretically you should be able to make a whole embryo-like structure just from the naive cells. And we were playing around with this. I had quite a few people working on this in the lab. But as we were working on that, there were actually a couple of papers from other groups showing that you can make what are called human blastoids starting from naive cells. One of those first papers came from Jun Wu, Austin’s former postdoc Ge Guo had a similar paper, and then Nicolas Rivron as well later that year. So those were really the first papers showing that yes, you can take naive stem cells and make a blastoid.

One impressive recent result from your lab is the development of a stem cell embryo model into the post-implantation stage. What were some of the key insights that made that possible? What were the biggest challenges?

With the blastoids, I obtained permission at WashU to take them up to 21 days. I was a little concerned whether we would get permission, but I explained we’re not going beyond gastrulation. We’re going right up to the stage where these specialized germ layers begin to form. There’s no heartbeat yet. There’s no neural tube. And we have this committee, the ESCRO committee, that oversees these types of experiments. They have all sorts of stakeholders. They have clergy, I think there’s a rabbi and a priest, and all sorts of administrators and scientists that actually look at our proposal. And they decided 3 weeks is a fair limit.

Here I need to give credit to my former postdoc, who’s now an independent PI at the University of Colorado, Rowan Karvas. We tried a 2D protocol but it didn’t work very well at all, the blastoids were disorganized. So at that point Rowan tried these 3D thicker matrices, essentially a thicker ECM gel. The idea is, the three-dimensional surface is more mimetic of what actually happens when an embryo implants into the uterus. We tried a bunch of different substrates, and we ended up settling on a combination of Matrigel and Geltrex.

Using that matrix we were able to culture these blastoids up to gastrulation stages, and we could see very good development of the placental structure. The outside of the blastoid on that gel forms these beautiful invasive projections, and it really resembles early placental development when you get these villi that sprout out from the embryo and go into the uterus, and we see that very nicely. On the inside of the blastoid we start to see the embryonic compartments organizing, but they’re still pretty disorganized.

What is clearly still a challenge for us and others is to allow that blastoid to form structures with the right spatial organization. You’ve got the cell types, but you don’t have the structure. And we’ve been honest about that. I think that’s an issue for the field. We do see this early sign of a primitive streak forming, which is where you start to get this asymmetry within the embryo where you’ve got the specialized cells forming on one end of the epiblast, it’s where the early mesoderm, the muscle cells, arise. So, we do see that happening, but it’s clear that the blastoids don’t have the right organization yet.

Exciting. While there are still challenges ahead, it sounds like the field is making great progress. Can you tell us a bit about how these models can lead to improvements in reproductive health?

Yeah. I think it’s unlimited in terms of the potential ways in which this can develop. To give a specific example, just recently in the past few weeks there were three studies that came out using blastoids to model implantation. And I’ll give these guys full credit, they sort of scooped us on it. It’s really exciting work from Peter Rugg-Gunn’s lab in Cambridge, his former postdoc Matteo Molè’s lab at Stanford, and a group in China led by Leqian Yu. They showed that you can take blastoids, combine them with endometrial organoids and endometrial stromal cells, and essentially recapitulate early maternal-fetal interactions.

And what’s really nice, the group from Beijing showed that you can not only model interactions between trophoblast cells, blastoids and endometrial cells, but you can even use patient-derived endometrial cells to recreate the conditions that lead to implantation failure. They took endometrial cells from women who experience recurrent implantation failure and they found that the blastoids actually attach to those much less frequently. And then they went even further. It was a Cell paper, to me it was like three papers in one. They even did a whole chemical screen to identify FDA-approved compounds that can overcome this implantation barrier. A lot of couples go through IVF, and even if the embryos are genetically normal, they still don’t implant. This is really the first study to suggest that you may be able to fix that with the right drugs.

The other key point, though, is they saw a very high degree of patient variability. Some of these patient cells required certain compounds to overcome the implantation failure. Others require different compounds. It’s really a need for patient-specific tailored precision medicine.

So I think the sky is the limit.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

For this piece, we chatted with two of our advisors: Alejandro Aguilera Castrejón, a Group Leader at Virginia’s Howard Hughes Medical Institute Janelia Campus, and Guojun Sheng, a Professor at Kumamoto University in Japan. Each contributes a different kind of expertise, from Aguilera Castrejón’s pioneering work with roller-culture machines that can keep mouse embryos alive at the very earliest stages to Sheng’s broad perspective on evolution of embryo development informed by decades of experience working with bird embryos.

What follows is an edited and curated version of our discussions.

Let’s start with a question for both of you. What drives your interest in embryology?

Sheng: I like seeing things change from simple to complicated. I can play a movie of how embryos develop for any organism, and I can make sense out of it, from one single cell, a fertilized egg, to a complicated organism. I’m interested in trying to figure out, essentially, how to make an embryo survive from simple to complicated, and how each different tissue and organ type evolves and connects to each other. So a sort of holistic view of embryogenesis.

Aguilera Castrejón: When I started, I studied biology because I like animals. I didn’t know that you can be a scientist on the bench, inventing technology. I was thinking more of watching animals in the jungle, something like that. But then during my bachelor’s degree, I discovered molecular biology, and then developmental biology, in particular, the potential of stem cells. The biological mechanisms of how the cell has the ability to form a body is very interesting for me.

Professor Sheng, can you say a bit about how the research you do in your lab differs from the research of other labs in the embryology field? What makes your lab special?

Sheng: We study a process called gastrulation, that’s the transition from pluripotency. When you have fertilization you have cells which are actually pluripotent, meaning they can become every part of your body, but at some point they have to decide to not stay pluripotent, so they have to differentiate, and there is a sort of stereotypical way of differentiation, so that the whole thing can evolve as an organism rather than a cell in the cell culture. So gastrulation is the most important kind of developmental process, transitioning from pluripotency to three germ layers: ectoderm, mesoderm and endoderm, and then from each germ layer you have more specific lineage differentiation, eventually giving rise to the lung, kidney, eye, brain, whatever.

Dr. Aguilera Castrejón, the pluripotency of stem cells is also very important to your work using mice to study embryology. In particular, it enables a powerful research tool, stem cell embryo models. Can you say a bit about what they are, and what kinds of research they are enabling that couldn’t be done before?

Aguilera Castrejón: In general, to get an embryo you need to put together a male and a female mouse, they mate, and then the egg is fertilized and you can study this fertilized egg. But of course the source is always the mice, right? So there is an ethical limitation on the number of mice that you can use.

In the case of stem cell embryo models, by aggregating different populations of stem cells that you grow in a petri dish under specific conditions, these cells will talk to each other. Then they will self-assemble into a structure that really looks very similar to an embryo. The advantage of this is that you can study the process of how the embryo self-assembles just from cells in the dish.

You had an important role in the development of the roller culture machine, used to grow both mouse embryos and stem cell embryo models. Tell us about the impact that has had on the field.

Aguilera Castrejón: Before, it was basically impossible to test whether stem cell embryo models can grow to more advanced stages, because there was not a stable protocol to grow even natural embryos. Now we know that we can culture these natural embryos to very advanced stages, which open a window for mechanistic studies of live mammalian embryo development during gastrulation and organogenesis. So this allows us to test whether the stem cell embryo models have the same capacity to grow as a natural embryo.

Curiosity-driven research is important, as is helping patients. What are some of the ways that a better understanding of how embryos develop can advance reproductive medicine?

Aguilera Castrejón: Now, if you see that there is a fetus or a baby with a specific disease, it is impossible to correct it, especially while the embryo is developing. In the future, if we create systems that allow you to maintain the fetus alive until term in vitro, if you detect a fetus or an embryo that has some genetic defects, then you can take it out of the womb and treat it to correct the disease. We know that it’s impossible to put it back in the womb, but maybe you can keep it alive in this ex utero system and then this will allow you to have a healthy human being in the end. So, I think that’s a very long-term goal, but it could make a big impact. Besides, the basic research for understanding organ formation during development that ex utero culture allows will certainly help for regenerative medicine purposes.

Professor Sheng, while Dr. Aguilera Castrejón’s lab works with mouse embryos, your lab studies bird embryos. Birds can do something that most mammals currently can’t: gestate outside their mother’s body. How well do we understand that process?

Sheng: That’s a very interesting thing.

Even in chickens you have the interface between embryo and environment. You need oxygen the same way the embryo needs oxygen, and the organ that you use to get oxygen is called the chorioallantoic membrane. Then you would imagine that whatever signal is being secreted from this tissue for an oviparous animal like a chick, there’s no maternal tissue to stimulate, right? So it can only stimulate the embryonic growth itself. And if it’s in a viviparous animal [an animal, like a mammal, that bears live young], then you have signals which can influence both the fetal side as well as the maternal side. So if you look at the genes involved in some of these, signaling is actually quite conserved there. Which means that what we think of as essential for fetal-maternal interaction in humans or in mammals, it’s probably built upon some system which was already present in the ancestor of mammals, and we’re building upon this already-existing kind of signal to strengthen the fetal-maternal interaction.

From a comparative embryology point of view, the placentation process, the fact that development needs maternal input, looks like it’s a novel thing for so-called eutherian mammals, placentals like us. But if you really think from a comparative embryology point of view, it’s actually not new. There are all sorts of intermediate things. Many reptiles have so-called live birth, and many reptiles also have nutritional absorption from uterine secretions from the mother, which is more or less like what we know of mammals. And you have mammals which have a minimal amount of the nutritional acquisition, like monotremes for example, and also not a very robust kind of fetal-maternal interaction. You have different ways of getting the maternal nutrition, uterine histotroph, which is to get the nutrition from uterine secretions, or hemotroph, meaning that you get nutrition from maternal circulation, but in a way that is different from the human cases, so-called hemochorial placentation mediating the fetal-maternal interaction. For example in pigs and horses they have very superficial implantation, they don’t have invasive implantation. So it’s a whole spectrum of things. Meaning that everything is okay as long as you meet some basic needs. Essentially, the embryo needs food and the embryo needs oxygen. The embryo needs an aqueous environment, and the embryo needs some way of getting rid of metabolic waste. And amniotes [a clade containing mammals, birds, and reptiles] can do it in a more or less conserved way.

As advisors for e184, you bring your expertise to advance our research. What attracted each of you to work with e184?

Aguilera Castrejón: I think that e184 is very realistic. There are many companies in the ex utero field but a lot of them are already trying to sell that they have machines for humans. And I think in order to see something like this in humans we need to do a lot of tests first in animal models. So I like that from e184, that it’s setting goals to establish a solid background for the field first in animal models, in order to be able to translate it later to humans. I think that’s something that attracted me to work with e184.

Sheng: For me as a more holistic view, I tend to be skeptical. “Interesting, but can it really work?” right? But thinking about it another way, a lot of things can work. You can have a cell from plant tissue and grow the whole plant out of it. So it’s not impossible: multicellular systems can start from a cell and then can recapitulate the whole thing. Then the challenge is, is it really possible for humans, or for some primate, or for eutherian mammals to achieve that…and I feel that the time is kind of right.

It takes a combination of the right mixture of people, that’s very important, and I think most individual labs or individual universities or individual schools of thinking, it’s probably not enough. It takes something that can lift a little bit up, from a slightly higher perspective.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

Such a device will need high-quality magnetometers or other advanced sensors, to resolve signals from neuronal currents as close to crystal-clear as possible. But high-quality signals are only one part of the puzzle. We will also need to be able to decode them.

Brain decoding has seen enormous advances in recent years…and faces unique challenges. Like many other scientific fields right now, it is an area where machine learning has made an outsized contribution, with powerful new models that map brain data to language, images, emotions, and movement. At the same time, there are obstructions that make quality models much more challenging to build than in fields that have seen dramatic public progress, like language processing and protein structure.

We have identified some of these challenges, and found candidates for promising solutions. We are just beginning our journey. We don’t claim to have all the answers yet. Our aim in this post is to establish a dialogue. If you work on brain decoding, we’d like to hear your thoughts. Are there key requirements we missed? Better methods than those we’ve identified?

We are establishing our hardware lab soon. We plan to launch an AI lab next, to start building the model stack we will need. We will need skilled researchers from day one, to plan and lead the project. If you’d like to get involved, or to collaborate in another capacity, we’d love to hear from you.

Key Challenges

Compared to other paradigm applications of machine learning, brain data is uniquely challenging to model. Brain data varies significantly in the same individual over time, and it can vary dramatically between individuals. The brain is in many respects expected to be an non-ergodic system, one which does not return to the same behavior over time, and as such with behavior that cannot be captured with a finite statistical sample. As one of the most important metrics for the success of a machine learning model is its generality outside of its training data, this variability is concerning. It isn’t a show-stopper: there clearly are general dynamics in cognition, not just across humans but even shared between species.

Our goal, then, is to find a model architecture and gather training data that lets us leverage universals and adapt to changing brains over time. We do not need to do this in perfect, scientific detail, but we need to do it well enough for our initial product goal of a mental smartphone: a device that lets a subject produce text and images with the power of their mind. We want that device to be a foundational technology, one that lets us build toward the full potential of a brain-AI interface for enhancing human cognition, and we know our data and architecture demands will grow and change along the way. But for now, we want a model stack that can cross that first hurdle: text, and images. The question then becomes what such a stack would need to accomplish.

If consumers are going to buy mental smartphones, they need to be comparable to, or better than, ordinary smartphones. That means comparable speed for text, 80 words-per-minute or better, and ideally to speech, 150-200 words-per-minute, as well as a low error rate, under 2%. For images, the question of error rate is inherently subjective: do people feel that the device better captures what they imagine than simply writing a text prompt for an image model? One could imagine measuring this through a ranking leaderboard, like those used for chatbots today.

A mental smartphone should also not be too difficult to set up. As anyone who has helped an elderly relative use a conventional smartphone can attest, the more time needs to be spent learning a new technology, the fewer people will be able to adopt it. In a brain-computer interface, the challenge is that not only will users need to learn the system: the system will need to learn the user. Each brain is different, and brains change over time. Calibration will be needed, and it is important to keep the burden on the user of that calibration as low as possible.

Finally, in a project of this scale, it is important to coordinate resources. That means keeping track of costs in data gathered and computational resources used, finding methods that make efficient use of both. It means building models that can be reused, transferred to different sensors, different signals, or different populations. And it means keeping track of feedback cycles, how long a research avenue will take until it can be assessed and how much interplay needs to occur between hardware and model development.

Promising Leads

With these challenges in mind, we can begin to ask: what kind of model “stack” would address them? What pieces will we need?

Foundation Models

One avenue that is driving a lot of excitement at the moment is foundation models for the brain. Much as large language models are trained on text completion but can serve as foundations for models that perform a variety of reasoning tasks, the expectation is that a large model trained via self-supervised learning on brain signals (using masked prediction or contrastive learning, for example) would be able to serve as a pre-trained foundation for models to decode a variety of types of brain data.

In EEG, LaBraM is an example of what such foundation models are capable of so far, with impressive performance on multiple types of tasks. Currently, no model of this scale exists for MEG, where state of the art performance on language so far has had to leverage substantially smaller datasets. These results do indicate that performance continues to improve as datasets get larger. Between those initial hints, and the general success in foundation models at generalizing across a wide range of applications, we see a foundation model as a strong avenue to achieving our performance goals in words-per-minute, error rate, and subjective rankings of images. Creating such a model will require a substantial investment: in recording new MEG data, in computational costs, and above all, in time. But we can expect that investment to pay off, creating a flexible foundation from which we can more easily accommodate new decoding tasks, calibrate to new subjects, and adapt to developing sensors.

Lightweight Personalization

It will be crucial to be able to calibrate to new users efficiently, without asking users to spend too much time in calibration or accumulating too many errors when varying between users. To address this challenge, we will need a lightweight way to personalize models without needing to change deeper features. Ideally, we would want to keep a high-quality universal encoder, for example one built on a foundation model as described in the previous section, and then make small modifications for each user, adding a small number of additional layers or a brief fine-tuning. A lightweight approach of this kind would also allow us to make more efficient use of data, as performance typically improves more by adding more data per subject than by adding more subjects, so keeping the need for training data for inter-subject calibrations low will allow for a more performant model from the same total recording-hours.

A variety of approaches to do this are actively being explored in the literature, and several look promising. Broadly speaking, one wants a form of transfer learning, in order to reuse data from similar subjects on a new subject. One interesting approach would be to learn a “subject embedding”, analogous to the word embeddings used in language models, to capture differences between subjects in the embedding space. There are a number of distinct methods for aligning data between subjects, several of which have yielded substantial performance improvements. Personalization approaches are already showing their worth for several of our goals of interest, with an approach similar to participant embeddings used for language and with tuning to subjects playing a crucial role for images. Overall, while it is not immediately obvious which method in this area will be most useful, it is clear that there is already an ample foundation to build from.

Continual Learning

Brain data changes not just between subjects, but in the same subject over time. A useful device will need a way to adapt to its user over a variety of timescales to remain effective. While one could do this by periodically asking the user to run new calibration sessions, this would represent a substantial burden on the user and would lower adoption of the technology. Ideally, we would instead be able to gradually adapt over the course of everyday use, shifting the model appropriately without a dedicated calibration session.

This kind of “continual learning” has been proposed in the literature. An impressive recent example is EDAPT, a framework that achieves update times of 200ms on current consumer hardware. This approach will likely require gathering additional training data for the initial model, in the form of data from the same subject over multiple days, but these resource requirements don’t appear excessively onerous.

Affective Feedback

A smartphone doesn’t have to know your opinion of it to function. Generating images from mental data, though, will be a different story. We do not imagine images in nearly as much detail as we perceive them, so the results of image decoding are on some level necessarily subjective. In order to achieve our performance goals for images, we will need to capture that subjectivity.

The most straightforward way to do this would be a manual rating system, but for a brain-computer interface, a more powerful option is possible: rating images directly from brain data on the user’s emotions. Emotion decoding is already a fairly robust subfield, and has even been applied to image classification already. Affective feedback may also help refine accuracy for language decoding, though this is likely to be a more minor effect.

Generative Models

Ultimately, brain data will need to be “fleshed out” on some level to produce specific images or words. An interesting possibility is to work generative models into the mix for selected tasks, cleaning up language and clarifying imagery. There has already been some success along these lines in the literature. Generative models have been used to improve the fluency and plausibility of decoded text. Their role in imagery is even more essential, and essentially every successful approach in that domain uses generative models. We expect a generative model layer to be a useful feature in any approach that can reach our performance goals for speech and for language.

Artifact Removal

Under real-world conditions, brain data is nearly always contaminated by artifacts. This is especially true for EEG, but it remains true for MEG, especially as new sensors make MEG viable in more varied conditions. In order to reliably extract meaning from brain data and meet the lower error rate we aim for, we will have to reliably filter out the effects of eye blinks, head muscles, movement, and magnetic noise from nearby devices.

There are several techniques in the literature to remove artifacts. A recent interesting example is the Artifact Removal Transformer, which trains on several types of artifacts in order to remove them from the data. A method like this looks to be an efficient way to remove particular kinds of expected data artifacts, yielding more accurate and robust systems.

What do you think?

We welcome feedback. Our method tends to be one of specialization: finding a few high-priority challenges and approaches, and focusing on understanding them better. Are there important challenges we aren’t considering? Powerful methods we’ve overlooked? If you have something to add, please comment here, or reach out to the Director of our Neurotechnology Program, Peter Zhegin, at p@e184.com.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

We would like to thank Blake Richards for discussions which sharpened the ideas in this piece, including helping us reach a better understanding of the unique challenges of brain data and the space of promising approaches to working with it.

Magnetoencephalography, or MEG, measures magnetic fields produced by coordinated currents in dendrites of pyramidal neurons. When currents in the cerebral cortex are lined up tangentially with the skull, they lead to measurable, if small, fields outside the head. These fields are on the order of femto-Tesla, much lower than the Earth’s magnetic field, which is around fifty micro-Tesla.

Measuring these tiny fields is the fundamental challenge of MEG. The approach has evolved dramatically over the years from David Cohen’s first experiments with copper coils in 1968, and is now able to reach greater sensitivity in more varying conditions. We plan to take it even further.

SQUIDs: MEG 1.0

While Cohen’s first experiment used an induction coil, he soon switched to what at the time was a brand-new technology, the superconducting quantum interference device, or SQUID. To date, SQUIDs still form the basis for the large majority of MEG devices in practical use. They can detect extremely tiny magnetic fields, down to several atto-Tesla in ideal cases, making them capable of registering the extremely tiny fields generated by the brain.

Due to their use of superconductors, SQUID MEG sensors must be contained in dewars of liquid helium. Hundreds of sensors are typically arranged in arrays to achieve quick readout from different parts of the brain, with each sensor registering signals with a single orientation. The volume of cryogenic equipment needed positions these arrays at around 2cm from the subject’s head. The result is a bulky, rigid setup, one that requires patients to sit still and have a head of roughly the right size and shape, making it difficult to impossible to use on children.

The SQUIDs used in MEG have a dynamic range of twenty nano-Tesla, high for a magnetometer but still too low to span the ten orders of magnitude between signals of interest and the Earth’s magnetic field. As such, all SQUID MEG is performed in magnetically shielded rooms, sophisticated installations with layered metal to filter out ambient sources of magnetic noise. These rooms are a feasible investment for hospitals and universities, but obviously not viable for a device meant for everyday use.

OPMs: MEG 2.0

SQUIDs use loops of superconducting material to create systems that are extremely sensitive to magnetic fields. Over time, it has become increasingly feasible and practical to create atomic and solid-state systems with this kind of sensitivity, a research field broadly referred to as “quantum sensing”. In recent years, quantum sensing approaches have leapfrogged ahead, finding advantages in a variety of settings from astronomy to biology.

Optically pumped magnetometers, or OPMs, are a quantum sensing technology that has made remarkable strides in magnetoencephalography. By using light to “pump” atoms into particular quantum states, OPMs control the spin state of atoms in vapors of alkali metals like rubidium, creating a system that is highly sensitive to magnetic fields without relying on the cryogenic cooling needed by superconducting systems.

The 1957 observation that optical pumping can create a magnetically sensitive state occurred even earlier than the prediction of the Josephson effect that lies behind SQUIDs, and by the 1970’s researchers had dramatically improved sensitivity and started the path towards miniaturization. Still, it took almost thirty years for the technology to become competitive with SQUIDs for applications that required compact sensors, with a resurgence of interest driven by the surprising observation that high-density vapors can avoid reductions in sensitivity caused by colliding atoms exchanging spin. The discovery of this spin exchange relaxation free (SERF) regime led to a expansion of neuroscience applications of OPM technology, starting in the 2010’s and continuing to today.

With these sensitivity gains, OPMs’ lack of need for cryogenics constitutes a potential massive advantage over SQUID systems. It means OPM MEG sensors can be placed much closer to a patient’s scalp, and can move flexibly to account for heads of different shapes, for a three-to-five-fold improvement in sensitivity. The systems can be made mobile, with comparably light headsets linked by cables to an apparatus comparable to a large backpack, allowing for MEG studies of subject in motion, as well as medical uses on children and others who would have trouble sitting still enough for SQUID-based MEG. OPMs are also vector magnetometers, able to detect magnetic fields in multiple directions at once, potentially gathering a richer signal if noise challenges can be overcome.

OPMs typically still need magnetic shielding, however. For OPMs these needs are not just due to the extreme difficulty of detecting fields from the brain against the Earth’s much larger field, but also due to tradeoffs in the system’s sensitivity in a noisy ambient field. As a result, many current OPM-MEG systems need almost an order of magnitude more shielding than SQUID systems. With that said, there is substantial potential to overcome these limitations. Due to the sensors’ ability to be placed closer to the scalp, active shielding is a realistic possibility, decreasing the need for passive shielding. By using arrays of OPM sensors to build gradiometers, researchers are able to reduce shielding requirements even more. There are even cases of OPMs being used to measure MEG signals under ambient conditions, albeit with strict restrictions on the environment to make sure that there were no nearby electronic devices or large metal objects that could introduce noise.

OPMs also still lag behind the state of the art in SQUID devices, although they are enjoying rapid progress. Currently, they have lower dynamic range than SQUID MEG, and a higher level of noise. The individual sensors are also bulkier and heated, and there are interference effects between nearby sensors, making it more challenging to make a headset with a large number of them. Still, there has been substantial progress in this area, and at least one system with 384 channels – more than typical SQUID MEG – is in late stages of development.

OPMs can now be routinely used to detect responses to sensory input, and have been used to detect the brain’s tracking of speech rhythms, and even to construct a rudimentary brain-computer interface in which patients spelled words based on looking at letters on a screen.

Next-Generation Magnetometers for MEG 3.0

To get beyond the need for shielding and special-purpose environments entirely, we expect to need a different quantum sensing approach.

Finding the right approach has been a significant journey. We investigated a long list of sensors, including dozens of magnetometers. Some were ruled out quickly, others merited a closer look. In the end, we identified technologies with the potential to go the distance.

Some of the most promising approaches involve solid-state sensors. With dramatically higher density than atomic gases, solid-state systems can pack more sensing power into smaller spaces, while their customizability allows systems to be tailored for a number of desirable properties. Many such systems can work at room temperature, avoiding both the cryogenic requirements of SQUID and the heat of OPMs. Many can also work in large bias fields without the nonlinear effects that degrade OPM performance, allowing them to work in ambient conditions without need for bulky magnetic shielding. These systems can be engineered by implanting defects in a material, as in nitrogen vacancy centers in diamond, by producing artificial crystals like yttrium iron garnet spheres, or by making use of novel properties of layered materials, as in sensors making use of the tunneling magnetoresistance effect or magnetoelectric laminates. While there has been some preliminary progress towards use of these systems for MEG, none of these technologies are currently sensitive enough to reach to the needed femto-Tesla range in a multi-channel system. However, many of these technologies have only recently been explored in this context. We see substantial potential for improvement.

What’s Next?

At e184, we’re building foundational technology for a brain-computer interface for all. Right now, our clearest path forward is to invest in sensor development, and build MEG 3.0.

We’ll be hiring soon. We’ll have physicists and engineers developing sensors, as well as experts in machine learning, building models so we can interpret input from our new sensor technology. We’re also building a board of advisors, looking for experts spanning multiple fields who can keep our project on-track. If you see yourself in those descriptions, send an email to p@e184.com to find out more.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

For some, brain-computer interfaces are a medical tool, surgically implanted devices justified by their enormous benefit to patients locked behind paralysis and neurodegenerative disease. For others, brain-computer interfaces are an opportunity, a way to take advantage of increasingly cheap sensors to gather health data that consumers might want to use.

We have a third perspective. For us, BCI isn’t just a tool to help a comparatively small number of patients, or a way to tempt the public with cute applications of current technology. It’s a path to the future, to our place in the future, for all of us. It’s the most promising way forward, to break free of the limitations of our evolutionary past and become something more. To keep up with artificial intelligence, to collaborate with it, to make it a part of us. It’s something that we want to benefit everyone, not merely a few, and something we want to be transformative, not merely nifty.

That perspective comes with fundamentally different priorities. It means we’re looking for a technology that can develop like cell phones: something that may begin bulky, expensive, and impractical, but with development and iteration can grow cheaper and lighter. It means working on the fundamentals, finding key technologies, like modern semiconductor chips and batteries were for cell phones, that make ambitious goals radically more feasible. And it means building a device not for one application or another, but something that can cover most of the use-cases we can imagine, and uses we can’t yet begin to predict. In fifty years, cell phones went from limited toys for the wealthy to essential work tools for some of the poorest people in the world. As technology moves faster, we expect BCI to follow an even quicker path.

We don’t expect this to be easy. Taking these priorities seriously means pushing for a kind of holy grail, a combination of traits companies have dreamed of and none yet have managed. Over the last year, we’ve sifted through the approaches on offer, and thought through what we’ll need. The technical demands are stringent, we won’t deny that. But they’re worth pursuing. After all, the future is at stake.

Evading the Surgical Bottleneck

In science fiction, brain-computer interfaces tend to be implants. Cyberpunk novels are filled with hackers with ports on the side of their heads, ready to plug in to a 1980’s vision of the internet. Right now, the most prominent players in brain-computer interfaces are following suit. Neuralink inserts electrodes surgically, close to key neurons in patients’ brains, while Synchron uses blood vessels to gain entry in a less invasive approach.

Neuralink and Synchron are building medical devices, and in that context, this approach makes sense. While researchers are still discussing how best to measure the benefit of these technologies, it is clear that, for patients unable to communicate or navigate the world on their own, the technology could be life-transforming. With benefits like that on the table, both patients and doctors are more than willing to accept the risks and costs of surgical intervention.

Those risks, and costs, are much less reasonable for the healthy. Surgery takes expertise, even with surgical robots. It takes time to heal, even for the simplest procedures. And it takes a risk, even for less delicate operations than modifying a human brain. There are roughly 1.5 million plastic surgeries in the US each year, and a bit over half a million LASIK surgeries. It is hard to imagine any procedure involving brain surgery to be anywhere near as common, and hard to imagine even a lighter surgery for minimally implantable tech to cover the thousandfold gap to universal adoption.

At e184, we are working to preserve our voice in the future, and we view brain-computer interfaces as essential to that work. We need a solution that will not just be accessible to a few, but that has the potential to benefit everyone in the world. That means we cannot let ourselves be bottlenecked by surgical techniques. We need a non-implantable technology.

Our First Goal: A Mental Smartphone

We won’t achieve the symbiosis we aim for in one shot. On the other hand, we don’t want to get sidetracked. We want to build foundational technology, technology that doesn’t just address a current need but paves the way for what we aim to build in future. That means we need to make sure our system has the right capabilities to be that foundation.

What are those capabilities?

Right now, the most general AI systems available interact via text and images. As a baseline, then, we want an interface that lets a user, with a thought, communicate a text or an image. We want users to be able to think through a message word by word, and send that message as a prompt to an AI, or a message to another user. We want users to be able to imagine an image, guide an AI through fleshing it out, then print it or post it to social media. In short, we want a device much like a smartphone, operated by the user’s mind. If we can manage that, we will have a solid foundation to go farther, and meaningfully enhance human cognition.

Already, researchers have made progress towards some of these baseline capabilities. Using magnetoencephalography in a magnetically shielded room, a team at Meta have managed to predict what volunteers are typing based on their brain activity with reasonable accuracy. Invasive approaches have gone farther, and using brain implants researchers are able not only to measure direct attempts at speech, but even to pick up on speech that their patients only imagine. Decoding images is harder, but there has been steady progress in the last few decades, going from identifying small black and white grids to using diffusion models to generate appropriate images from fMRI and even EEG data. These methods focus on decoding images that are seen, not imagined. Reconstructing imagined images will be more challenging, but there has some progress in reproducing imagined images from fMRI data.

In contrast, research on interfacing in the “other direction” – that is, stimulating a user’s brain to convey speech or images – is still in its infancy. While there are researchers and companies that stimulate patients’ brains with brain-computer interfaces, they generally focus on specific medical outcomes, like treating epilepsy. At the moment, it is unclear how this side of a general-use brain-computer interface would work, and this is a question we postpone for the future. Initially, our mental smartphone will be one-way: a tool to send messages and images, not an additional way to receive them.

That still adds up to a substantial challenge, as we are well aware. It will mean reading signals from multiple regions of the brain, to decode both imagined speech and images: ideally, we want to get as close to whole cortical coverage as possible, to decode signals from the entire cerebral cortex. And it will mean doing so without using an implant, to an unprecedented resolution.

The Measure of a Brain

In the ideal case, we would want to detect the activity of individual neurons in the cerebral cortex. This would require distinguishing objects a fraction of a millimeter in size, corresponding to the length of dendrites of human pyramidal cells. If that proves infeasible, then we expect we will at least need to distinguish the activity of small groups of cells, on the scales that invasive techniques probe via implanted electrodes, meaning resolution of a few millimeters at worst. Currently, this is the scale that medical BCI systems have needed to probe in order to achieve communication rates on the same order as those of speech with a realistically large vocabulary and acceptable error rate. Non-surgical techniques have yet to achieve this threshold, and we suspect this is due in part to the absence of any technique that probes the same resolution as implanted BCI. If we can achieve that few-millimeter resolution, we have a chance.

In addition to achieving good spatial resolution, we will need to achieve good temporal resolution. Neuron voltage spikes can be spaced hundreds or even tens of milliseconds apart, with peaks often lasting only a few milliseconds. That means we will need to resolve and distinguish signals on those timescales. If we measured with a lower temporal resolution, we would not be able to see activity of individual neurons, and would be restricted to collective measurements.

How do existing non-implanted methods compare?

Some methods detect brain activity indirectly, via its impact on blood flow. This includes functional Magnetic Resonance Imaging (fMRI), which detects movement of oxygenated blood in the brain via inducing resonant effects in hydrogen nuclei in a strong magnetic field, functional Near-Infrared Spectroscopy (fNIRS), which distinguishes oxygenated blood in the outer brain through its absorption of near-infrared light, and functional Ultrasound (fUS) which uses the Doppler effect to measure the relative speed of blood to an ultrasound probe.

All of these methods rely on neurovascular coupling to track brain activity via its dependence on blood oxygen. The enormous magnets required by fMRI make it purely useful for medical and research applications. Portable devices need to be based on fNIRS or fUS, though both are still quite motion-sensitive, with fNIRS having additional sensitivity to scalp details due to only probing a fairly superficial layer and fUS faces challenges getting a high-resolution signal through the skull. More fundamentally, all three methods, to different degrees, suffer from a shared disadvantage: timing. Blood flow in the brain responds to signals on time scales of seconds, at least ten times too slow for the kind of real-time communication we are aiming for.

Electroencephalography, or EEG, instead detects electrical signals from voltage jumps in the cell membranes of groups of pyramidal cells (along with electrical signals from muscles in the head, eye blinks, and saccades, which must be disambiguated). This is a favored method for many non-implanted approaches at the moment, from devices that gather health data from a small number of electrodes, like ear-EEG, to measurements based on hundreds of electrodes across the scalp.

While EEG is well-suited to current approaches, we do not expect it to achieve the kind of precision we need. The voltage signals detected by EEG are transmitted nonuniformly through the brain and skull, leading to volume conduction effects that make them challenging to interpret with real precision. EEG is best at detecting radially-oriented currents towards the outside of the brain, demanding more powerful interpretive techniques to access the remaining two-thirds of cortical currents. Actually localizing signals often depends on models for the heads of individual patients, an added source of potential error. Even if signals for EEG could be perfectly disambiguated and traced, the technology has a more fundamental limitation: in order for a signal to be detectable at all on the scalp, it must be of sufficient strength. Potentials in the scalp act as a low-pass filter, restricting the frequencies that are able to pass through. This rules out detecting signals from individual neurons or small groups, instead demanding large synchronous patches of the cortex to fire. As such, even the most optimistic estimates expect a resolution of centimeters, likely too low to reach the accuracy we need to decode speech and mental imagery.

(It may be possible to improve on this resolution limit by coupling the brain’s electrical activity with induced ultrasound waves, creating a mixed signal that can pass through the skull and scalp. While this idea is intriguing, it is still too early to tell whether it can overcome EEG’s disadvantages to a sufficient extent.)

Finally, magnetoencephalography, or MEG, detects magnetic signals generated by electric currents within the brain. As currents from axon and synaptic sources cancel out, MEG only detects signals from dendritic currents in pyramidal neurons, which are conveniently exactly the brain activity that is most relevant for this kind of BCI. Unlike electrical signals, magnetic fields pass through the head almost entirely undistorted, probing much deeper and giving a much cleaner interpretation than EEG, and can localize sources to a 2-3 millimeter scale.

MEG does, however, have one enormous limitation: the Earth’s magnetic field. While brain signals lead to magnetic fields outside the scalp in the range of hundreds of femto-Tesla, the Earth’s magnetic field is around fifty micro-Tesla, a difference of eight orders of magnitude. In practice, this limits MEG to clinical and research environments with either passive or active magnetic shielding. Conventional MEG uses superconductors, which necessitates cryogenic cooling, leading to extremely bulky setups. It is also extraordinarily sensitive to motion, so a subject’s head must be kept as motionless as possible.

Optically pumped magnetometers have recently seen use for MEG (referred to as OPM-MEG), and their requirements are much lighter. They still require shielding, but less than conventional MEG, and are able to be used for moving subjects. They do have additional issues with cross-talk between sensors and field drift, and heat and power transmission makes it challenging to pack enough of them next to the head to achieve high resolution. Still, they represent a significant step towards our goals.

The Upshot

We are at a pivotal time in history. It matters enormously what we choose to build. If we are to enhance human intelligence to catch up, and collaborate, with machine intelligence, we need brain-computer interfaces that can be accessible to all. That will mean non-surgical technology, and it will mean technology powerful enough, with high spatial and temporal resolution, to pick up what we want to communicate.

That in turn, means that none of the technologies available today will suffice. Hemodynamic methods are simply too slow, and EEG has too low resolution, while current MEG cannot be used outside of magnetically shielded rooms.

However, while the limitations of hemodynamic and electrical measurements are biological, the limitations of MEG are a matter of engineering. Already, OPM-MEG represents a step in the right direction. A MEG sensor that can do what we need of it, achieving high resolution in everyday life, does not contradict any physical law. It simply does not exist…yet.

Thus, a brain-computer interface company that aims for true human-AI symbiosis…must first be a sensor company.

In an upcoming post, we’ll explain what that means.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

That, in turn, requires a virtuoso performance of cell biology. We need to take cells and change which gene programs they express, leading them to develop in new ways: a process called cellular reprogramming. Academic researchers have made remarkable discoveries in this field since Takahashi and Yamanaka's groundbreaking work on induced pluripotent stem cells in 2006, in areas as diverse as treating Parkinson’s, addressing macular degeneration, and performing large-scale drug screenings. But the next chapter demands something different: systematic exploration of cellular plasticity at an industrial pace.

Academic research excels at discovery and fundamental understanding, uncovering new biological principles, developing novel methodologies, and pushing the boundaries of what's possible. These contributions form the foundation of everything we build. But translating these scientific discoveries into robust, reproducible therapies for human patients requires a complementary approach: one optimized for systematic exploration, standardization, and scale.

From discovery to systematic exploration

Traditional academic cellular reprogramming operates within natural constraints. PhD and postdoc timelines favor projects that can yield publications within reasonable timeframes. Grant structures like R01s support focused investigations rather than massive systematic projects.

These constraints have driven incredible creativity and fundamental insights among academic scientists, but they also leave a niche for a different approach. Where a typical academic project might test dozens of regulatory factors, industrial platforms can systematically explore hundreds or thousands. Where publication timelines favor demonstrating proof-of-concept with single cell lines, clinical translation demands protocols that work robustly across diverse genetic backgrounds.

These platforms don’t replace academic discovery. They build on it. Academic innovations provide biological insights and foundational tools, while industrial platforms engage in the systematic exploration needed to translate these discoveries into reliable, scalable applications.

The combinatorial frontier

Consider the mathematics of cellular reprogramming. Takahashi and Yamanaka began with a screen of 24 transcription factors in order to discover their Nobel Prize-winning cocktail of four. Academic research has gradually expanded this combinatorial scale, with many studies now considering combinations from pools of 50-100 factors.

There are 10,626 different ways to choose four transcription factors out of 24, a space of combinations still within reach of a systematic screen. Choosing four out of 50 requires considering 230300 different possible combinations, which already stretches the limits of the feasible.

Robust protocols for complex cell types like gametes may require exploring much larger spaces. Modern screening approaches systematically evaluate extensive libraries of regulatory factors, not because massive numbers are necessarily required in final protocols, but because discovering optimal combinations often necessitates screening considerably larger pools to navigate complex regulatory landscapes.

In order to achieve this, we’ll need a different approach. We won’t be able to test every combination, so we’ll need to guide our search, using machine learning models to pick out which combinations are most likely to be promising. We’ll be following in the footsteps of companies like NewLimit, building models that predict the ability of different transcription factor combinations to enable transformation to gamete cells with the same precision that their models predict effects on cell age.

The challenge isn't just computational, though. Even screening hundreds of combinations requires standardized experimental procedures, automated approaches, consistent quality control, and sophisticated data analysis. It demands the kind of systematic engineering mindset that naturally follows after academic discovery.

The research opportunities

Our IVG approach creates unique research opportunities that bridge fundamental biology and clinical application. We're building a team to work on problems that span multiple scales:

Fundamental discovery: Constructing causal gene regulatory networks through systematic perturbation experiments. Understanding how epigenetic features influence transcription factor efficacy. Exploring combinatorial control mechanisms where factor combinations achieve what individual factors cannot.

Technical innovation: Developing accelerated differentiation protocols. Creating robust, reproducible methods that work across diverse cellular contexts. Building predictive approaches for cellular reprogramming outcomes.

Clinical translation: Advancing therapies for infertility through systematic understanding of human gamete and embryo biology. Creating protocols that work across diverse patient populations. Contributing to scalable approaches for future cell-based therapies.

More than just scale, our approach is one of integration. Computational insights directly inform experimental design. Experimental results immediately feed back into predictive approaches. Fundamental discoveries connect directly to clinical development.

What we're looking for

This work requires deep expertise across multiple disciplines, and more importantly, researchers excited by interdisciplinary collaboration. We need:

Cell biologists and developmental biologists who understand the fundamental mechanisms of fate conversion, cellular reprogramming, and developmental pathways. Experience with transcription factor-based reprogramming, stem cell differentiation protocols, or reproductive biology is particularly valuable.

Data scientists and bioinformaticians skilled in machine learning, multi-omics data analysis, and predictive modeling. We need people who can work with multi-modal datasets, build interpretable and predictive models of cell fate conversion, and translate computational insights into testable hypotheses. Experience with single-cell transcriptomics, chromatin accessibility data, proteomics, or other high-throughput biological measurements is valuable.

Engineers and automation specialists who can design systematic experimental approaches, implement reproducible protocols, and scale biological processes. Experience with laboratory automation, quality control systems, or bioprocess development is highly relevant.

Researchers at the intersections - people passionate about solving problems that combine computational methods with experimental rigor. Whether you're a computational biologist interested in experimental validation, or an experimentalist eager to incorporate predictive modeling, we value and encourage interdisciplinary thinking.

The advantages of our approach

Beyond pure scale, systematic approaches to cellular reprogramming offer key advantages:

Speed: Well-designed factor-based approaches can dramatically accelerate cell fate conversion research compared to traditional methods.

Efficiency: Systematic screening identifies more effective factor combinations that increase differentiation success rates.

Reproducibility: Standardized, systematic protocols ensure consistent results across experiments and researchers.

Novel insights: Large-scale perturbation studies reveal regulatory relationships and combinatorial effects that smaller studies cannot capture.

Robustness: Protocols developed across multiple cellular contexts and genetic backgrounds are more likely to work reliably in diverse applications.

Career development and culture

We offer unique growth opportunities that few academic or industry settings can match:

Leadership development: Early-career scientists can lead cross-functional projects where biology, computation, and engineering intersect. You'll work directly with team members across disciplines rather than within traditional silos.

Clinical exposure: Direct involvement in translational challenges and the path from scientific discovery to therapeutic application.

Technical breadth: Exposure to cutting-edge approaches across computational biology, experimental automation, and clinical development.

Scientific freedom: Substantial independence to pursue creative approaches within our shared mission. We believe the best innovations come from empowering researchers to leverage their unique insights while maintaining a collective focus.

Rapid iteration: See your ideas tested and refined quickly rather than waiting years for validation.

Join, and build something that matters

We're assembling a team that can bridge the gap between advanced research and clinical reality. If you're excited by systematic exploration at an unprecedented scale, if you want to see fundamental biological insights translated into reproductive technology that helps people become parents, if you're drawn to research problems that require both deep thinking and systematic execution - we'd love to hear from you.

Whether you're a PhD student interested in how systematic approaches can amplify research impact, a postdoc ready to take on interdisciplinary leadership challenges, or an experienced researcher excited by the intersection of computation and experimental biology, there may be a place for you on our team.

The next breakthrough in reproductive medicine won't come from a single experiment or a single lab. It will emerge from systematic integration of prediction, validation, and clinical insight - all working together at scale.

Ready to be part of that future? Connect with us to start the conversation.

Here at e184, we're not just building better experiments - we're building a better approach to biological discovery. One that honors fundamental insights from academic research while creating the systematic capabilities needed for clinical impact.

You can be part of that.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

Over the next few months, we plan to lay out the core of that strategy. We want to explain why we think our goals are realistic, and the road we expect to walk to get there. In doing this, we aim to extend an invitation: if you like where we’re headed and you’ve got the scientific know-how to make it happen, we’d like to find a role you can play.

First, though, we want to say just what those goals are. At e184, we work to give people a voice in our future. We’re addressing infertility, aiming to give hundreds of millions of families a new chance to influence the world of tomorrow. In our view, brain-computer interfaces are also critical to preserve people’s stake in the future, on an even greater scale.

AI, and our role in it

The conversation around artificial intelligence (AI) often defaults to existential risks - endless debates fuelled by fear of a rapidly changing world. If you have checked Google Trends lately, the spike in searches makes it clear: people are nervous. But while others dwell on downsides, we are focused on the upside. The real question is not just, ‘What if things go wrong?’ but rather, ‘What if we get it right?’

We believe in a future of AI systems that can absorb information and generate actions at 10x or even 100x the speed of the human brain, tackling humanity’s grandest challenges - curing diseases, solving hunger, and reversing climate change. But here is the catch: when AI delivers solutions at superhuman speeds, humanity may hit a critical bottleneck - our capacity to comprehend and implement those solutions.

This is not just about social constraints like regulatory inertia or misplaced fears. No, this is about the fundamental limits of human cognition compared to AI’s potential for exponential growth in speed, intelligence, and adaptability. It's an ironic reality that while we’ve made tremendous progress in advancing artificial neural networks, we’ve barely scratched the surface when it comes to understanding - let alone enhancing - our biological ones.

Imagine an AI medical researcher, tasked with addressing cancer. We would want a human expert to be able to check its reasoning: to observe the assumptions it makes in real-time, to step into its chain of thought and correct for distractions, reminding the AI of the constraints of medical ethics and the goals we have for care. As AI gets faster and faster, this goal becomes more and more distant.

Or on a more everyday level, imagine using an AI to design a home. Today, you would need to write down a description, conveying only part of what you had in mind, then laboriously tweak the prompt or modify the final design to fit your goals, leaving most of the potential of the AI wasted. What if instead you could convey what you wanted directly, sending an image directly from your imagination, then stay in control as the AI fills in the details?

This growing disparity between human and AI capabilities isn’t just a technical challenge - it’s a barrier to adoption,to happiness even. It is a roadblock to fully realising the transformative potential of AI in the most critical areas of society.

That’s why we believe the next frontier is clear: the integration of human and artificial intelligence. To truly thrive in the age of AI, we must explore how to bridge this gap and perhaps even merge human cognition with AI systems. Only then can we unlock the full potential of a world powered by collaborative, benevolent AI - one that works faster and in harmony with human ingenuity and intuition.

The Landscape of Human-AI Collaboration

There are other approaches to bridging a future gap between human and artificial intelligence, with different expectations. Some are in progress now, others are hypotheticals for the future. Some expect us to stay as we are, others expect humanity to become something radically new. Here’s how we envision this landscape of approaches, and our place in it:

While there is immense potential for progress purely on the AI side - through enhanced explainability, alignment, or even intentionally slowing AI down to better collaborate with humans - we believe that it will eventually be necessary to change humanity itself.

With that said, we are not chasing the singularity or the full digitisation of humanity. Those ideas, while captivating, pose significant scientific, technological, and philosophical questions that remain beyond the scope of our mission. Instead, we are laser-focused on achievable breakthroughs - solutions that empower humanity to collaborate with AI on equal footing and unlock new horizons of creativity, innovation, and impact.

We see that work as taking place in stages. First, we want to help humans communicate with AI. Initially, this will mean building on our conscious thought, by making it as natural to send a command to an AI in the cloud as it is to think up a sentence or call a memory to mind. In time, we will go further, and approach the speed of unconscious cognition, tapping more directly into our underlying intent. In the even longer term, we want to understand human cognition better, to build a kind of operating system, one that seamlessly integrates AI capabilities. Achieving this work will be no small feat - it challenges millions of years of evolutionary compromise.

BCI AI - augmenting evolution

Communication bottlenecks

Our nervous system and the interfaces we use to communicate were designed for a specific evolutionary context, where every calorie counted and no extra capabilities could survive. The external information transfer rates they support - whether between humans and humans or between humans and machines - are shockingly low. Depending on the method of communication (reading, speaking, eye tracking, mouse movements), these rates range from about 1 bit per second to no more than 50 bits per second. Compared to typical internet speeds of hundreds of megabits per second, this means that a browser is able to communicate with an AI at a data center a million times faster than a human user can deliver commands to that browser. Bridging even a small portion of this vast gap would be an enormous jump in our capabilities.

On top of that, the context-independent nature of our communication with machines further limits efficiency. While our brains operate with complex semantic relationships that naturally embed context, our communication with computers relies on syntactic commands devoid of semantics. Imagine a jet pilot needing to issue a series of commands (e.g., ‘turn left, then accelerate’) instead of conveying a single, context-rich directive like ‘follow that target’.

To improve the performance of machines and user experience, we need computer systems to be able to incorporate more information beyond what we can give in an explicit command. Such information may help to build and train better AI, combining reinforcement learning from human feedback with richer, more meaningful information from BCIs that enhances AI's responsiveness, adaptability and effectiveness. In essence, in order to align AI, we would like it to not merely do what we say in detail, but do what we intend, overall, a combination of our conscious desires and the unconscious processing that underlies them. As brain-computer interfaces make progress from detecting explicit intended speech to inner speech, we expect there to be a path forward to guide AI with this kind of processing.

Processing bottlenecks

Ultimately, our goal is not merely to enable humans to communicate with AI as we are: we want to use computers to improve our internal ability to process data. Here too, there is clear room for improvement. Peripheral processing by the ‘outer brain’ (the retina or visual cortex, among other areas) operates in parallel, handling vast amounts of sensory information simultaneously. For example, the retina produces a million parallel output signals, which the visual cortex processes further through its hypercolumns, a parallel set of 10,000 modules. However,once sensory information reaches central cognition, the ‘inner brain’, for example, the prefrontal cortex, processes it serially in order to commit to a conscious response. As a result, ‘… when faced with two tasks in competition, individuals consistently encounter a ‘‘psychological refractory period’’ before being able to perform the second task’. This serial processing clocks at 10 bits per second, a limit to how quickly we can consciously assess new ideas.

These limits, in turn, appear to affect our mental representations of what we see. Researchers have argued that ‘… items that are attended to and foveated [i.e. related to the fovea, the small central part of the retina responsible for sharp central vision] are perceived at a higher resolution, while items that [are] unattended or are in the periphery are primarily perceived as being part of an ensemble…’, wrapped into a kind of lower resolution sketch.

Now compare these limits to the streamlined logic of computers - machines comprised of billions of transistors, designed for sequential processing, each operating millions of times faster than a human neuron. Thus, in addition to the constraints of our motor system, which slow our ability to communicate with computers via typing or speech, we are constrained by the processing capabilities of our brains themselves.

Possible futures

If humanity is to thrive alongside AI, we need to solve at least three critical bottlenecks: communication, context, and internal processing. Only when we achieve this, will our brain-machine interface be able to evolve into a brain-machine operating system, enabling augmented cognition that transcends our natural limitations.

Imagine an AI seamlessly handling complex sub-tasks for us, using external processing as an extension of our minds. Even without AI, with a brain-reading device we could store results externally, retrieve them, and solve problems with unparalleled efficiency. AI adds additional text-processing capabilities on top of this. While we process text consciously at around 10 bits per second, GPT outputs around 100 tokens per second, with each token carrying around 10 bits of information. Giving people direct mental access to this kind of processing would be a jump of two orders of magnitude in speed of conscious thought.

We’re building this future - a world where humans and AI are no longer disparate entities but seamlessly integrated partners. We pursue this future because we cannot stomach the likely alternative: a world where humanity is entirely led by a powerful AI, where we lose our agency and simply execute ChatGPT’s orders.

Join us in creating an extraordinary future where constraints on human cognition are lifted and AI becomes a partner, not a distant tool. Together, we can unlock a new era of augmented human-AI cognition.

We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:

General Application

We would like to thank Blake Richards for helpful comments on an earlier version of this piece.