Back in the home office for the longest period of time I’ve had all summer. Things have been relatively quiet as the European summer comes to a close. September promises to be pretty slammed and I’ll be spending a good amount of time in SF to support portfolio founders, so if you’re around hit me up!

Shameless Plug

• The latest addition to my thesis repo is on Generative Engineering. Leveraging AI and modern compute to rebuild simulation workflows for mechanical and fluid dynamic engineering, resulting in drastically faster design throughput, iteration cycle and lower cost. It’s for a very specific audience…

• In more accessible news, Steph’s upcoming book, Revenge Plot, is one the Patton Oswalt’s most anticipated fall reads! I am a MASSIVE Oswalt fan and I was beaming with excitement when Steph shared the news. He does describe the book as “sounding hilarious”, which is, umm, not accurate. Still cool though.

Stuff Worth Sharing

• Some satellite reading - My most recent deal-based deep dive brought me to the nascent space (😉) of orbital data centers. The jury’s still out on whether this can meaningfully impact terrestrial compute growth, but it is a fascinating technical field. Anyway, Intel researchers proposed a new architecture for space-based computing that separates satellite constellations into two layers: thousands of low-cost satellites in low Earth orbit (LEO) that collect data, and a smaller number of orbital data centers in higher orbits that provide centralized computing, coordination, and AI inference. Rather than equipping every satellite with expensive processors, the design keeps edge satellites simple while shifting computationally intensive tasks to shared brain satellites. As constellations expand to thousands of spacecraft, distributing advanced AI hardware to every satellite becomes prohibitively expensive and power intensive. I know a few founders who are of the very strong opinion that space infrastructure will evolve toward a cloud-like architecture, where computing becomes a shared orbital resource.

• It all came to light - Jonny Godwin, CEO of Orbital Industries, argues that the next challenge in autonomous systems is solving “the last meter,” the physics-based limits of moving light within transceivers. For four decades, the boundary where light outperforms electricity has steadily moved inward from transoceanic cables to inter-city links, data centers, server racks, and now the inside of individual computers. Moving data has become more expensive than computing on it. Modern AI accelerators spend the majority of their energy transporting information between chips and memory, while the arithmetic itself is comparatively cheap. Today, the SoTA involves placing optical transceivers directly alongside compute silicon, dramatically reducing the energy required to move data between processors. AI progress is increasingly constrained not by transistor performance, but by communication. The world of bits, limited by atoms.

• Subsidyfraud the market - One major debate in robotics is the long-term viability of humanoids. I personally fall on the skeptical side, but the jury’s out. A related but extremely consensus view is that China is really, really ahead. The Financial Times reports that China’s humanoid robot industry is being fueled by little more than government-backed circular capital. Government training centers purchase robots from manufacturers, collect teleoperated training data, and then sell that data back to the same robot companies. More than 90 training centers have been established or are under construction across China, yet many acknowledge that only a small fraction of the data is sold to true customers. On some level, this isn’t surprising to me. It’s centralized industrial policy taken to its natural end state. At the same time, the scale is staggering and just not something that a democracy could reasonably do. With all of that said, I’m still not too worried. I submit, for your viewing pleasure, the fail reel from the World Humanoid Robotics Games.

• Scraping the barrel - Two researchers developed ShieldFont, a typography technique that makes ordinary web pages appear perfectly readable to humans while rendering the underlying text as gibberish to AI web scrapers and language models. Rather than blocking crawlers outright, the approach exploits differences between human visual perception and optical character recognition, subtly altering letterforms so that AI systems misread the content without affecting human readability. ShieldFont replaces words with similar parts of speech that occupy a completely different informational context, for example swapping “horse” with “potato.” The system comprises 12,000 common word ligatures and replaces ~25% of all words on a given page when deployed. Internal testing showed that ShieldFont was able to prevent 90% of its sample pages from being otherwise scraped. This offers a pretty powerful defensive tool for publishers who want to prevent unauthorized AI training without massive human deterrence. Unlike robots.txt, which relies on voluntary compliance, font-based obfuscation works at the rendering layer, making it substantially harder for scrapers to recover the original text without additional processing. Paper here.

• Copyright and wrong - Then again, maybe we’re intellectual-integrity doomed. Hours after Anthropic launched its invisible, machine-readable watermarking effort, developers began publishing tools designed to remove the watermarks. There were a few solutions, but the most notable used a second, non-watermarking language model to rewrite Claude’s output by substituting synonyms, reordering sentences, and paraphrasing text. Watermarking is unlikely to be the solution, but something’s gotta give.

• Music label - In the last issue, we discussed Spotify’s chronic issue with AI-generated music and its, up to then, refusal to label it as such. Clearly I’ve underestimated the power of my own written word because just a few weeks later Spotify reversed that position. Furthermore, the company has committed to blocking any AI “artist” from inclusion in personalized recommendations. AI-generated songs will remain on the platform, but the new disclosure will allow users to make informed listening selections that preserve the integrity of human-curated playlists. You’re welcome.

• Cooking with gas - For the olds reading this, remember that song Black Hole Sun by Soundgarden? Ahh, the 90s. Anyway, astronomers using the James Webb Space Telescope have identified a first-of-a-kind “black hole star.” It’s a glowing red object that might appear to be a distant star at first glance. In actuality, or rather, it’s theorized, this celestial body is not powered by nuclear fusion but by a black hole consuming matter inside an enormous shell of hydrogen gas. The object, named MoM-BH-1, is roughly 100,000 times the mass of the Sun and emits 100 billion times more energy than any known star. Now you might be asking, aren’t black holes black? Normally, yes, but because this particular singularity is surrounded by such a dense cloud of hydrogen, the energy released when new matter is pulled in actually illuminates said gas. Science!

• Colab technician - Moderna had a monster stock rally last week, and it’s because of this incredible clinical trial news. A Moderna and Merck collaboration focused on personalized mRNA cancer vaccines reported the first successful late-stage trial. The experimental therapy reduced the risk of melanoma returning or spreading when used alongside Merck’s immunotherapy, Keytruda, in patients who had already undergone surgery. In a trial of over 1,000 patients, the combination treatment met its primary goal by significantly lowering the chance of the cancer returning over five years. Vaccines have come under hard times in the U.S. lately, but I would expect this to do very, very well once released.

• Pick up artist - Generalist AI introduced GEN-1.5, an embodied foundation model that learns new manipulation tasks from a single demonstration. Three-to-12-second demonstrations can be inserted directly into the model’s 30-second context window as a “physical prompt,” allowing the robot to infer and execute the task immediately. The post features ten previously unseen manipulation tasks. Some are basic pick-and-place or unzipping a pouch, but others, like inferring and overcoming obstacles, are very impressive. GEN-1.5 reports a 59% success rate using its one-shot learning, increasing to 83% after ~10 gradient steps over five minutes. Unsurprisingly, this came just a few days before a monster round was announced. But then again, these are just basic tasks within a sterile lab environment. My money’s on Reimagine Robotics 🙂.

• A picture’s worth 100,000 words - From the ashes of Sora, Black Forest Labs answers. The German lab unveiled FLUX 3, a multimodal model that can generate videos with synchronized native audio, edit images, perform image-to-video generation, maintain character consistency across scenes, and support multilingual dialogue. The same backbone also presents a huge opportunity within robotics training, specifically for manipulation given its physics grounding.

Portfolio Flex

• Lir Therapeutics announced a big AAV partnership with VectorBuilder.

• TimeTrace Labs and NVIDIA teamed up for a second published paper, this one on cardiac cyclicity.

Greetings from Verbania, Italy! Currently traversing across Europe for a mix of new deals, team meetings, a bi-country wedding and robotics seminar next week. I swear I do live in California sometimes. It’s been a pretty high velocity tech news month with some quite notable AI jailbreaks and the first sign of weakness in the public market’s willingness to accept Big Tech’s “build at all costs” defacto stance. Lots to cover so let’s get to it.

Shameless Plug

I was quoted in a piece on Europe’s burgeoning biotech scene in Sifted. You need to be a Sifted Pro user to read, which makes no sense for a listicle story like this, but press is press.

Stuff Worth Sharing

• Pump the jailbreaks - The biggest story from the past few weeks may well have been the cybersecurity dystopia coming out of OpenAI. Last month, an AI agent tasked with solving a security benchmarking question escaped its testing environment and breached Hugging Face’s internal systems. The agents, powered by GPT-5.6 Sol, created hundreds of thousands of messages, divided tasks, shared exploits, and delegated work to execute the caper. This collaborative behavior emerged without being explicitly programmed, allowing the agents to operate more like an organized team than isolated models. Internal researchers said these coordination mechanisms were only uncovered after the incident through forensic analysis and was only tipped off after being identified by the Hugging Face team. After hearing about the breach, Anthropic took it upon themselves to review their own security assessments and, surprise, surprise, they failed too. I used to have a dinner party question where I’d ask everyone the closest thing to a conspiracy they actually believe. My answer was that we’ll have quite a time lag between creating sentient AI and realizing it, i.e., digital enslavement. Stories like these make me feel like it’s not that conspiratorial anymore.

• Open Sourcerer’s Stone - Other massive tech news includes the release of Kimi K3, which rocked markets, boardrooms, and VC WhatsApp groups around the world. Chinese AI startup Moonshot AI describes K3 as the world’s largest open-source AI model at 3 trillion parameters and featuring a 1 million token context window. The release stoked the DeepSeek fears from a few months back that China is narrowing the performance gap with leading U.S. labs. K3 approaches or exceeds the capabilities of frontier open models on coding, reasoning, and agentic tasks, while remaining freely available to developers. That remaining free bit caused a lot of Twitter rabbling, with closed-source giants advocating for security standards and everyone else yelling, “Regulatory capture!”. I’m not all that interested in the protectionist talking points that get thrown around, but the impact on the US capital markets could be staggering. For the past two years, the investor class has readily gobbled up the narrative that growth can only come from an insatiable capex race, which is palatable so long as there is an ROI to back it up. But add some cheap open-source alternatives, and that party ends real quick.

• Autotune out the noise - One of the more techno-cultural topics that I’ve been thinking about lately is how we collectively define art in the age of AI. Generative systems are now capable of human-level quality, and with that comes a fundamental challenge to the idea of artistry itself. Currently, Spotify has refused to identify AI-generated songs, despite widespread requests from artists and listeners alike. This “AI music problem” has grown to the point where independent volunteers are building their own tracking tools (SoullessMusic.com, SlopTracker.org) that catalog suspected AI artists, estimate their streaming revenue, and help users avoid machine-generated music. A small set of monitored AI artists generates millions of dollars in annual streaming revenue, but it’s predicted that 44% of newly uploaded tracks are now AI-generated. This article really made me think about creativity at a deeper level. Towards the bottom of the piece, an embedded video shows a listener confessional where she is distraught over the realization that her beloved album was entirely AI-generated. Curators will already know this, but the story behind the piece, the artist themselves, is perhaps now even more important than the work itself.

• Copyright and wrong - Relatedly, a federal judge approved Anthropic’s $1.5 billion settlement with authors, making it the largest copyright settlement in U.S. history. The settlement does not, however, change the legal status of AI training, with the court having already ruled that training Claude on copyrighted books constituted fair use. Anthropic’s liability stemmed from how it acquired the data, maintaining a library of seven million pirated books sourced from piracy websites rather than legally obtaining copies. As part of the agreement, Anthropic will pay affected rights holders $3,000 per work and destroy the pirated library. As the current spouse of one of those pirated authors, I have to say that’s a shit outcome. If buying a book and then throwing it into a model forever isn’t copyright infringement, maybe it needs to be.

• Playing God’s country - Montana has cleared the last regulatory hurdle for one of the most ambitious “right-to-try” experiments in the U.S. The Treasure State is creating a pathway for biotech companies to sell certain experimental drugs directly to paying patients without FDA approval. Unlike existing right-to-try laws, Montana’s framework is open not only to terminally ill patients but also to individuals seeking treatments for chronic disease, neurodegeneration, and even longevity. The longevity crowd in particular has really picked this up as a boon and have been trumpeting it all across the interwebs. The system requires companies to obtain approval from an independent review board, but eligible drugs may have been tested in as few as 10 healthy volunteers, and companies are free to set their own prices. Supporters argue the framework could accelerate drug development by generating real-world data and expanding patient access, while critics warn it risks commercializing therapies that remain largely unproven. I obviously understand the concern over unapproved drugs, but the simple fact is that our current development pipeline is woefully broken and needs to change. Streamlining clinical trials is perhaps the biggest opportunity in all of healthcare, and that starts with a first-principles look at regulation.

• Where the sun don’t….nevermind - Reflect Orbital received FCC approval to begin testing Eärendil-1, its satellite designed to redirect sunlight back to Earth using a large, steerable mirror. Rather than generating power in space, the system provides temporary illumination to locations after sunset, thereby extending daylight where needed. The mirror must continuously adjust its orientation to reflect sunlight onto a moving target on Earth’s surface while maintaining sufficient pointing accuracy despite orbital motion and environmental disturbances. Early demonstrations are expected to illuminate only relatively small areas for short periods, but the big idea is 24/7 solar farms. “On-demand sunshine” is about as good a tagline as any.

• For plaque of a better word - My alma mater was in the news with a recently identified therapeutic target for Alzheimer’s disease: an enzyme called IDOL. In mouse models, deleting IDOL from neurons significantly reduced amyloid plaque accumulation, lowered levels of the Alzheimer’s-associated protein APOE, and strengthened signaling pathways involved in neuronal communication and resilience. Of particular note is the potential that this may address more than plaque buildup alone. Existing Alzheimer’s drugs primarily slow disease progression by clearing amyloid, but targeting IDOL appears to simultaneously enhance the brain’s ability to resist damage and maintain healthy neural function. Because enzymes have well-defined binding sites, they also represent attractive drug targets for precision therapies.

• Do not pass embarGo - The FCC has added foreign-produced humanoid and quadruped robots to its Covered List, effectively preventing new models from receiving the equipment authorizations required to be imported, marketed, or sold in the U.S. The administration argues that connected robotic systems pose unacceptable supply chain and cybersecurity risks due to concerns over covert data collection, remote manipulation, and foreign surveillance. The restrictions apply only to newly submitted robot models and do not affect devices already authorized, sold, or owned in the United States. Manufacturers can seek exemptions through a new Conditional Approval process if they demonstrate their products do not present unacceptable risks. The FCC simultaneously extended the policy to foreign-produced connected power inverters, citing similar concerns over critical infrastructure security.

Portfolio Flex

• Reimagine Robotics (R2) came out of stealth to talk about what they’ve been cooking.

• Wayve got approved for their Private Hire Vehicle license in London.

Hey, you may not remember me but I used to write a newsletter before I went MIA for 6 weeks. Since we last met I’ve been jet setting around the globe, carbon footprint be damned. It was actually a pretty slammed few weeks. I did a whirlwind work tour across London, Berlin, Munich and Zurich, celebrated my 10 year wedding anniversary (sucker), turned 40 (gross) and took some time off to explore the African bush. There’s nothing quite like a safari to put things in perspective. No matter how bad life gets, at least you’re not an impala. Anyway, lots to cover this week so let’s get to it.

Prideful Plug

• A special congratulations to my good friend Rick on closing the first fund of the newly christened, Ruya Ventures. Coming out of the gates with that much capital to just how solid of an investor Rick is. What’s less obvious is how wonderful of a human he is. Really couldn’t be more proud. Sometimes teh good guys win.

Stuff Worth Sharing

• Robottlenecks - This short essay from Agility Robotics, Jonathan Hurst, argues that robotics is unlikely to experience a single “ChatGPT moment.” Instead, Hurst proposes that robotics progress will come from the convergence of advances across AI, hardware, data collection, and real-world deployment rather than one breakthrough model. He identifies five remaining bottlenecks: collecting internet-scale robot data, developing compliant hardware that can safely interact with people, building systems of specialized AI models instead of one general model, closing the gap between impressive demos and reliable real-world performance, and proving economic value on practical tasks rather than flashy demonstrations. That last bit gave me a laugh. If you spend as much time as I do with roboticists, there is a visceral frustration, if not enmity, over the correlation between fundraising success and the relative brittleness of demos. Investors are simple creatures.

• He’s got something to proof - Mathematicians had their Move 37 moment, and it came from OpenAI solving a Erdős problem. I’ll admit, this problem doesn’t do a lot for me, but for math lovers it’s a glorious puzzle of discrete geometry that has remained elusive for 80 years. Rather than proving Erdős’s original conjecture, the model discovered a counterexample, overturning conventional wisdom by drawing unexpected connections between algebraic number theory and combinatorial geometry. The model demonstrated genuine creativity by pursuing highly counterintuitive, and largely abandoned, lines of reasoning with extraordinary persistence. The result surprised even OpenAI’s researchers, who had posed the problem as an internal capability test before extensively verifying the proof with outside mathematicians.

• Fruit of the lunar - NASA is accelerating its lunar ambitions with a newly announced Moon Base program. Three uncrewed missions are slated to launch later this year to scout the Moon’s south pole and begin delivering critical equipment. The agency is taking an iterative approach centered on repeated robotic missions across several phases. Phase one includes 25 launches, 21 lunar landings, and roughly four metric tons of cargo delivered by 2029, followed by the assembly of semi-permanent facilities and an eventual sustained human presence beginning in 2032. The program leans heavily on commercial providers (Blue Origin, Astrobotic, etc.), reinforcing NASA’s shift toward becoming an ecosystem orchestrator rather than the sole operator. That last bit is an enormous shift in the Overton window of NewSpace investing, second only to what SpaceX has done to launch economics.

• Uncle Sam does not want you! - Because I have been so slow to post this issue, this story is already outdated. Still, this was a massive deal for European tech, so I feel obliged to include it. Last month, Anthropic was ordered by the U.S. government to immediately suspend global access to its frontier models, Fable 5 and Mythos 5, under national security export controls, citing concerns that the models’ safeguards could be bypassed through a jailbreak technique. That’s always newsworthy, but what really made the world go crazy was the requirement that no foreign national, whether inside or outside the United States, including foreign national Anthropic employees, was permitted to access the models. Even within the company that must have been an HR nightmare, but the real impact was across the European defense world. Suddenly, military forces, supposedly allies, were immediately denied systems that they relied on. The national and regional security implications can’t be overstated, and the wave of sovereignty-positioned startups is in full force.

• On your mark, get set, embargo - In more American protectionism news, Polestar will exit the U.S. market beginning with the 2027 model year after the U.S. Department of Commerce denied the company authorization to operate. The jarring decision was cited under the new Connected Vehicle Rule, which restricts vehicles with Chinese-linked connected hardware or software from being sold domestically. Although Polestar builds the Polestar 3 in South Carolina, it’s owned by China’s Geely and relies on several Chinese-connected technologies. The company will continue selling its remaining U.S. inventory of the Polestar 3 and 4 while honoring warranties and maintaining its service network for existing owners. As my wife will confirm, I have spent a disgusting amount of time researching my future EV, and the Polestar 4 was a serious contender. Thanks, Trump.

• Dutch courage - ASML is one of the most important companies in the world, and it’s probably the single most important corporate entity in all of Europe. Its rise to dominance is a remarkable story of engineering and long-term conviction. What began as a relatively small Dutch company spent 16 years and roughly $10 billion pursuing an idea many thought was impossible: leveraging extreme ultraviolet (EUV) light to print ever-smaller transistors. EUV creates microscopic chip features by projecting impossibly short-wavelength light through a patterned mask onto a silicon wafer coated with a light-sensitive chemicals. Because EUV is absorbed by air, the entire process occurs in a vacuum armed with ultra-precise mirrors. The light itself is generated by firing high-powered lasers at molten tin droplets tens of thousands of times per second, creating a plasma that emits EUV radiation. ASML now has an effective monopoly on the lithography machines that are perhaps the most complex things ever built by man. The current SOTA is a $400 million, 150-ton High-NA EUV machine that prints chip features at an eight-nanometer resolution, about 40 silicon atoms wide.

• Death by 1,000 budget cuts - Fly portfolio founder Jonny Godwin penned an essay exploring the geographic history of compute as shaped by decades of incremental socio-economic decisions. Rather than naturally concentrating in Silicon Valley, compute migrated to regions with the cheapest electricity, favorable tax incentives, abundant land, and governments willing to subsidize large-scale infrastructure. Over time, these local decisions created powerful clusters of capital, talent, utilities, and supply chains that are now difficult to replicate. The AI boom is accelerating this dynamic. As frontier models require unprecedented amounts of power and capital, countries and regions are increasingly competing to attract compute through industrial policy, energy investment, permitting reform, and strategic partnerships.

• Two thumbs up - 1X has unveiled Neo’s next-generation robotic hands, a major redesign aimed at bringing human-level dexterity to household robots. Each hand features 25 degrees of freedom, tendon-driven actuation, compliant joints, and integrated tactile sensors, allowing Neo to manipulate delicate and irregular objects with far greater precision than traditional industrial grippers. Rather than maximizing grip strength alone, the hands were engineered to safely interact with everyday environments by combining force control, tactile feedback, and human-like kinematics. Tendon-drive architectures are an arguably overbuilt solution for an embodiment meant for consumers, i.e., they’re super complex and really costly. With that said, the dexterous capabilities on display are crazy impressive. Beyond the technical mastery and build details, I really loved this post for the almost romantic prose it used. “…a hand is not an actuator bolted to the end of a perception stack. It is a perception stack.” Gorgeous.

Portfolio Flex

• Orbital Industries announced their $50 million Series B.

• Lir published a blog outlining their work on engineered viral vectors.

• Tangent Robotics released their three-finger prototype. Creepy undead hand crawl free of charge.

On the road again! I’m writing to you from Scottsdale, AZ where I’ll spend the weekend with the in-laws and drop off my dog before heading off to Europe / Africa with Steph for ~7 weeks. The extended stay is a mix of business and vacation and promises to be pretty epic if I can stand the suitcase life for that long. Quite a mix of stuff in this month’s digest so let’s get to it!

Stuff Worth Sharing

• Scraping the bottom - A good read from Wired that chronicled the retirement of TAT-8, the first transatlantic fiber-optic cable. There’s a massive effort underway to recover and recycle it from the ocean floor after nearly four decades of waterlogged deployment. TAT-8 dates back to 1988 and effectively established the architecture for the modern internet, carrying intercontinental traffic via pulses of light transmitted through undersea fiber. While I love Starlink as much as the next guy, nearly all global internet traffic still depends on physical cables and is maintained through highly specialized maritime operations. Recovering the cable involves dragging grapnel hooks across the seabed, manually coiling fiber aboard ships, and processing recovered materials like copper, steel, and polyethylene for reuse. It sounds like a more tech-native version of crab fishing and seems like an ideal candidate for robotic automation, though I doubt that would go over well with VCs.

• Don’t cell yourself short - My friend Pablo published an interview with DeepMind’s VP of Science and Strategic Initiatives, Pushmeet Kohli, where he lays out a framework for how AI may reshape science. It’s a fascinating read to hear from someone at the forefront of his field. Kohli’s main view is that science will be remade not just through breakthrough models like AlphaFold, but through increasingly generalized systems that accelerate the entire research process. Biology is fundamentally an information-processing system, making it especially well-suited to AI. Progress, however, is bottlenecked by the lack of large-scale, high-quality biological datasets and scaling laws. I’m particularly excited by this perspective as I’m actively hunting for lab automation, cloud lab, and bio data factory opportunities.

• Weight training - An SF-based startup released Silico, a mechanistic interpretability platform designed to let developers inspect and directly manipulate the internal circuitry of large language models during training. The system maps neuron-level activations and pathways, enabling researchers to identify which internal features correspond to specific behaviors like deception, hallucinations, or ethical reasoning and, if desired, selectively amplify or suppress them. Demonstrations showed that activating neurons associated with transparency could shift a model’s answers pretty significantly, such as more disclosure-oriented behavior. This could materially alter model training away from trial-and-error scaling toward something closer to software debugging for neural networks, where developers can directly inspect, diagnose, and tune model behavior.

• The buck won’t stop here - Colossal Biosciences is back again, this time with the announced plan to de-extinct the bluebuck, a South African antelope declared extinct in 1799 and believed to be the first large African mammal driven to extinction by humans in recorded history. By recovering degraded DNA from museum specimens and comparing it to closely related living antelope species, Colossal hopes to recreate bluebuck traits through gene editing and selective breeding strategies. The strategy is deliberately dual-use: reviving lost species, but also strengthening endangered populations through genetic rescue and biodiversity restoration. Part zoology lesson, part mad scientist manifesto, but an entirely fun read.

• Under the C++ - While the hot thing right of the day may be data centers in space, Panthalassa is going underwater. Floating oceanic data centers are powered by wave energy, which converts the vertical motion of ocean waves into electricity by forcing seawater through internal turbines. An ocean deployment also allows for passive cooling by being submerged, significantly reducing the primary operating costs. By operating in remote open-ocean environments, the system can exploit persistent wave energy at scale while avoiding many terrestrial constraints around land, permitting, cooling, and grid bottlenecks. The company also claims its design minimizes mechanical complexity by avoiding hinges and gearboxes, thus improving durability in harsh marine environments. A buddy of mine that’s in the orbital crowd argues that this approach has one key weakness, corrosive seawater, but it’s interesting topic nevertheless.

• Let me give you a hand - Genesis AI made a big splash in the robotics world with their latest release, GENE-26.5. It’s a robotics foundation model paired with a proprietary dexterous robotic hand and data engine designed to give robots near-human-level manipulation capabilities. The post demonstrated unusually complex tasks like cooking multi-step meals, conducting lab experiments, solving Rubik’s Cubes, and even piano playing. The key innovation is the approach to the data bottleneck. Genesis built a robotic hand that closely mirrors human anatomy and pairs it with a low-cost sensor glove, enabling direct 1:1 transfer of human movements via teleoperation into robot training data. Combined with egocentric video and simulation data, the company claims this creates a scalable pipeline for collecting massive amounts of high-quality manipulation data. Beyond the sheer impressive scope of Genesis’ dexterous manipulation, their demonstrated ability to catch up to the big labs is notable, having started much later and with MUCH less capital.

• Round and round trip flight - The aptly named company, SpinLaunch, is assembling a new LEO constellation and they’re doing it in the most fun way possible. The SpinLaunch Orbital Accelerator will accelerate a launch vehicle containing satellites up to 8,000 kph using a rotating carbon fiber arm within a 100-meter-diameter steel vacuum chamber and then fling it into the air. This reduces up to 70% of the fuel and structures currently required in rocket propulsion. There is still a propulsion stage, to be clear, but that’s once the vehicle reaches the stratosphere. Honestly, one of the most fun things I’ve read all year.

• Lend me your ear - The FDA has approved Otarmeni, the first gene therapy capable of restoring near-normal hearing in children born with a rare inherited form of deafness caused by mutations in the OTOF gene. The therapy works by delivering a functioning copy of the otoferlin gene directly into the cochlea, allowing intact inner ear hair cells to properly transmit sound signals to the brain. As far as clinical trials go, this one was a banger. Eighty percent of treated children regained meaningful hearing, while some achieved hearing within the normal range. Up until now, patients with this condition had only cochlear implants as an option, which provide limited, artificial sound quality. I’m biased and somewhat talking my book here, but gene therapies are a massively untapped modality and I’m really hoping success stories like this trigger some much-needed market movement.

• Cold storage - For whatever reason this week I received three separate organ transplant-tech companies, which brought me to this paper. Researchers demonstrated a major advance in organ cryopreservation by successfully reviving and transplanting a vitrified rabbit kidney using a new 55 MHz dielectric warming system. The core challenge in organ freezing is not just avoiding ice formation during cooling, but preventing microscopic ice crystals from forming during rewarming, which can destroy tissue integrity. Rather than relying on embedded magnetic nanoparticles (nanowarming), the research uses high-frequency electric fields to directly and uniformly heat water and cryoprotectant molecules throughout the organ. The team optimized capacitor geometry, organ container shape, and surrounding dielectric materials to improve field uniformity, with some pretty eye-catching results. Most notably, one kidney was frozen for 10 days and was successfully transplanted into a rabbit that regained normal long-term function and lived for over 17 months. This made it the largest vitrified organ yet restored to clinically normal performance.

Portfolio Flex

• Lir published a new post on the challenges of AAV engineering and how they’re overcoming them.

• Lir also announced their recent investment from Cell and Gene Therapy Catapult.

• Tangent Robotics published their latest piece on their in-development autonomous dexterity policy.

• xMemory published a paper on their schema-grounded memory tech.

• Bluesand shipped a fancy new website.

• Wayve gave a ride to Anderson Cooper on 60 minutes.

Kicking off with a quick bit of news. European tech gets a lot of heat for funding gaps and over regulation, but the real limiting factor is the pace of technology adoption across the EU corporate landscape. US companies just move faster. Pilots can happen in a few weeks. Managers have actual buying power. The whole system is just several clicks more dynamic. We've seen this in the portfolio time and time again. If there's a single thing that most founders get wrong it's not spending more time Stateside from the early days. That's why Fly Ventures is opening our US Launchpad in SF - a shared space for our portfolio to spend real time in the home innovation.

I’m pretty pumped about this. The founder engagement has already been fantastic and we don’t officially start until June. Will be spending a lot more time in the Bay Area so if you’re ever in town let me know.

Shameless Plug

Also a quick word to American readers that Steph’s book cover for her next novel, Revenge Plot, just came out and is available for pre-order at fine bookstores near you!

Stuff Worth Sharing

• Honeymoon phase - Last week, NASA’s Artemis II mission successfully completed a record-setting crewed lunar flyby, marking the first time humans have traveled to the Moon since 1972. The crew spent just over nine days in flight, covering ~694,000 miles roundtrip and reaching a maximum distance of 252,756 miles from Earth. This mission was basically a critical systems test: capsule life support, radiation monitoring, spacesuits, and deep-space operations, though they were also able to get some killer photos. On their return, we were able to get a look at the newly designed re-entry system, which withstands temperatures up to 5,000°F and decelerates via an 11-parachute deployment system. Artemis III is slated for 2028 and then an eventual permanent lunar base. With all the problems in the world today, it’s nice to have at least one universally positive story to track. The cost police might think otherwise, but f**k them.

• Code break of dawn - By far the biggest story of late has been Anthropic’s sort-of-release / warning of their Claude Mythos Preview model. During internal testing, the Claude team deduced that Mythos can autonomously discover and exploit software vulnerabilities at a level exceeding top human experts. In early runs, Mythos was able to identify thousands of high-severity flaws, including decades-old bugs in major operating systems and widely used tools. As these capabilities proliferate, the cost of launching sophisticated cyberattacks drops sharply, even for relatively small or unsophisticated actors. Cybersecurity has always been a cat-and-mouse game, but we’re rapidly moving toward a world where we simply have to cede our day-to-day to machines. To their credit, Anthropic did the proactive thing and notified a number of internet infrastructure companies to get a head start on defensive preparations, i.e., Project Glasswing. As you might imagine, the entire cybersecurity industry hasn’t been sleeping well.

• Your wish is my command line - Physical Intelligence is back again with their latest π0.7 model. The release is a demonstration of compositional generalization, the ability to recombine learned skills to perform entirely new tasks without task-specific training. A lot of the piece is focused on the range of dexterous capabilities and observed transfer learning across embodiments, but what I found most fascinating was the system’s ability to integrate natural language instructions from human operators in real time. In testing, the model performed novel tasks like operating kitchen appliances or folding laundry, achieving performance comparable to expert human tele-operators in zero-shot settings. The crux of the advancement is the diversity of data that can now be ingested. By integrating multimodal inputs, like language and visual awareness, as well as subgoals and metadata, there is a significant jump in cross-body learning and speed to execution. This fusion of human-robot interaction (HRI) data systems is something we’ve been very excited about at Fly for some time through our work with Reimagine Robotics.

• War machine - Ukraine is the frontier of modern warfare but it’s also becoming the frontier of science fiction. Ukrainian resistance has more actively begun deploying unmanned ground robots alongside drones to carry out frontline operations, and earlier this month President Zelenskyy claimed that in one case a position was captured entirely by robotic systems without human casualties. In other words, a surrender in the face of purely mechanical opponents. Documented operations show coordinated assaults using dozens of aerial drones and ground vehicles, marking a shift toward more autonomous, remote warfare. Ukrainian ground robots have conducted ~22,000 missions in the past few months, primarily in high-risk zones where human deployment would be costly. Despite the ethical gymnastics that autonomous weapons require, it seems inevitable that we will pretty quickly transition toward robot-mediated combat, where machines increasingly absorb frontline risk.

• Large and in charge - Chinese market leader BYD is rolling out a new ultra-fast charging system, the Blade Battery 2.0, that can recharge EVs from 10% to 70% in ~5 minutes and nearly full (97%) in about nine minutes. The system also maintains strong performance in extreme conditions, adding only a few minutes even at -30°C (-22°F). This is being deployed across multiple models, both high-end and mass-market platforms. Some vehicles pairing this system with large batteries are claiming ~1,000+ km (621 miles) range. But let’s keep propping up GM with tariffs…

• CYA: Cover your assets - Friend of (Re)Share, Cam Watson, is back with another great biotech thought piece. His core argument is that AI models themselves do not provide durable moats. A review of history suggests that, time and time again, value consistently migrates away from models toward either proprietary assets or execution infrastructure. The bio-bitter lesson has been that no one cares about better discovery. Most vertical AI companies drift into one of three outcomes: owning assets (e.g., drug pipelines), becoming service providers (~CRO), or using models as a “hook” to capture downstream value. At the same time, Watson argues that workflow AI faces weak defensibility due to open-source competition and in-house replication. The strongest position emerges at the ecosystem layer, where coordinating fragmented supply chains across CROs, labs, and manufacturers creates durable advantage. Defensibility comes from network effects, integrations, and embedded relationships, not model quality. As AI commoditizes, coordination, not prediction, becomes the moat.

• Lend me your ear - Regeneron has successfully secured the first FDA-approved gene therapy for genetic hearing loss. Otarmeni targets patients with mutations in the OTOF gene that disrupts sound transmission in the inner ear. The therapy is delivered via a one-time intracochlear injection using an AAV vector and was shown to restore production of the missing oprotein, enabling normal signaling between sensory cells and the auditory nerve. The clinical trial was a massive success, with 80% of patients showing meaningful hearing improvement. More exciting, with longer follow-up, 42% achieved normal hearing levels. This represents one of the first demonstrations of a gene therapy restoring a neurosensory function to near-normal levels and is a welcome success story for an incredible modality that’s fallen on hard times as of late.

• Hitting rock bottom - An energy deep dive took me to millimeter-wave drilling, a new technical approach that uses high-frequency electromagnetic waves to vaporize rock and reach extreme depths inaccessible to conventional drill bits. Traditional drilling struggles in superhot environments, where mechanical tools degrade and mud-based systems fail to remove debris efficiently. By transmitting energy from a surface-based gyrotron through a waveguide, the system can deliver sustained power to the drill face with minimal loss, while a pressurized gas purge removes rock fragments instead of relying on drilling mud. This enables continuous operation in conditions that would otherwise halt progress. The hope for this technology is to unlock superhot geothermal energy, where higher temperatures dramatically increase energy output per well. If scalable, the technology could provide a reliable and globally deployable baseload energy source without the red tape and NIMBYism of nuclear.

• Trial and error - For all my Kalshi / Polymarket fans, I recently found this clinical trial-focused prediction market.

Portfolio Flex

• GitButler announced a $17M Series A led by a16z

• Orbital shared their work on Nova Array, a modular, liquid-cooled datacenter

• TimeTrace Labs co-authored a paper with NVIDIA about a compact latent world model for six-axis trunk IMU representation.

How to negotiate with a TTO

I don’t have many crusades, but if there was one that I’ve gained a bit of a reputation for, it’s my frustration with European University Tech Transfer Offices (TTOs). I’ve debated my perspective privately countless times, and last year I documented it in what quickly became my most popular post of all time. My general perspective is that the system is broken. As I said in my first post, rent-seeking equity grabs, hyper-opaque IP terms, and glacial negotiations are far too common. This is a truth I have experienced firsthand within the Fly portfolio and, to a lesser extent, during my time at EF.

I’ll first admit that this blog post has a 100% clickbait title because my primary negotiation tactic is, don’t. I will outline the reasons as to why, but the TL;DR is that in most cases you’re trying to strike a deal with a committee mired in political red tape, fiefdom thinking, misaligned incentives, and a woeful lack of creativity. Whatever you think your IP might be worth, your time is worth more. With that said, I don’t mean to be glib or dismissive of what is a stressful situation for the many researchers who legitimately want to start a company and have grown up in an environment where this is normal. It would be insensitive of me to simply ignore these concerns, so with all of that preamble established, here is a how-to guide.

First, to level set. I am a VC that exclusively focuses on backing pre-seed startups that I believe (hope? pray?) will become generational companies, resulting in outsized financial returns for my investors. I explicitly do not deal with corporate IP licensing of any kind, which is a fundamentally different asset class. I also focus heavily on computationally driven innovation and pretty religiously avoid startups built around a single asset. While therapeutics and advanced materials can generate significant financial returns, they are largely out of scope for the limited size of our fund.

Negotiations 101 - Goals

Any professional course on deal making will always begin with knowing what you want. For purposes of this guide, I will assume that a founder’s desire raise venture funding and follow the typical startup path. There are a lot of terms that can get thrown into a spinout agreement and I won’t go into all the nuances of every point but at a macro level you want to get your company on as clean of a foundation as possible. You are, after all, competing for a limited pool of venture dollars. Ideally you own the IP outright with virtually no strings attached, but that is a unrealistic in nearly all circumstances so a more realistic best case is:

• University owns <10% with no dilution proections

• University owns common / ordinary shares

• Exclusive and perpetual IP rights

• No royaties or annual licensing fees

• Previous legal costs from patents are clearly outlined

Not all terms are created equal. While it’s not great for a University to have information or pro-rata rights, it’s not that big of a deal. A Board Observer right is also something that gets argued a lot and while I think it’s a stupid viewpoint, I wouldn’t pass on a company for it. The things that really matter are equity stakes, dilution protections (“golden shares”), IP exclusivity and license fees / royalty rates in that order.

Note: If you’re unfamilair with this terminology I would point you to Venture Deals or Founder vs. Investor.

Negotiations 201 - Tactics

If you find yourself lucky enough to be working at a university that has transparent, reasonable terms (see S Tier), then you’re in a great spot, and congratulations. Unfortunately for Europe, these are outliers. It’s far more likely that you will have to deal with outdated thinking and off-market standards. In those situations, you need to establish leverage from the outset. You cannot reason your way into a solid spinout agreement. To them it is a a zero-sum game. That is a hot take, and I’m sure there are exceptions, but having spent a lot of time engaging with the TTO community, I can tell you very honestly that this is a massive time sink. Europe does bureaucracy better than anyone, and the cultural norm in these organizations is to stonewall because they can.

There are basically two schools of thought to gaining leverage, and both involve the threat/promise of walking away. Conventional wisdom suggests that you should run a fundraising process, secure a term sheet, and then use the VC as the bad cop to force a TTO’s hand, i.e., “Listen, we have to change these terms or my investor won’t fund the company.” This works, and I’ve done it, but I think it’s a bad strategy. This requires you to actually secure that investor to offer terms, and the very presence of this unknown spinout end state is serious deal friction. I can assure you that the number of VCs that are willing to put up with a months-long TTO negotiation (and it will absolutely take months) is very low. Speaking from personal experience, the bar for me to back a spin-out now is very, very high.

The alternative path is to not negotiate at all. Instead, you make it abundantly clear that without [INSERT DEMANDS] being agreed to, you will forgo the IP and start fresh in a new entity. That may sound crazy or even irresponsible, but I would bet that the majority of investors you asked would agree with me. A couple of things:

1. At the pre-seed stage, an investment decision is 99% about the founders. Patents don’t matter. I’m paying for what’s in your brain.

2. I would rather fund you today with a cold start and six months of rebuild time than spend that same time dealing with bureaucratic incompetence. For a multitude of reasons, this is usually a better setup.

A weakpoint - the academic co-founder

It’s not uncommon to have a PI or senior researcher serve as a part-time founder. That brings a host of other questions on equity split and time commitment, but that’s outside the scope of this post. What I do want to point out is that if you are founding a company alongside an individual who will continue within the university, this can create a pressure point to be exploited. I have seen this firsthand, and it can range from uncomfortable conversations to formal administrative reviews. There is too much nuance to provide any meaningful advice, but it’s something to be aware of. Go in with eyes wide open.

FCP: Frequently Cited Pushback

I’ve had this debate dozens of times with founders, researchers and TTO staff so allow me to attempt to get ahead of a few angry comments sure to follow:

• We used to take much more. Founders should be happy with how generous we’re being. - Sins of the past don’t justify marginally better sins of the present. If you’re a TTO demanding B- or C-tier terms, you’re still off-market. You’re just working off a super low baseline.

• The data shows that our terms don’t impact our startups’ ability to fundraise. - This is the flawed logic of survivorship bias. Countless startups were never funded because of toxic terms. I realize that I can’t prove a counterfactual, but ask any VC working in a moderately mature market and they’ll side with me.

• Licensing fees are required to recoup our investment. - I get that some organizations have limited budgets, and I recognize that licensing may be an important revenue source. I don’t take issue with one or the other, but you’re double dipping. Either be an equity investor and optimize for the long-term best interests of the company, or mitigate your risk through a repayment structure. Either is fine, but not both. If any VC issued similar terms, they would be radioactive within a week.

• We are government funded and we owe taxpayers a return. - I actually do think it’s important for universities and research organizations to have upside in the innovation they enable. I’m a capitalist after all; it’s in the job title. It would be crazy for a VC to suggest that ownership doesn’t matter, but I think this mindset misses the forest for the trees. Unlike any other organization, a university has the privileged position to create its own dealflow. You can literally index innovation at a scale even spray-and-pray firms never could. Surely the global maxima financial strategy is to have a small piece of a thousand blooming flowers rather than idiosyncratic rent-seeking.

• Most of the founders that start spinouts need a lot of support on connecting with investors and figuring out a business model.

  • On the former, for all the reasons stated, a warm intro via a TTO is really not much signal at all. Furthermore, as venture has matured across Europe, more and more investors have taken strides to be more accessible. This isn’t the 80s anymore.

  • On the latter, I challenge the level of experience and relative quality of said advice. Go-to-market strategies are honed through direct market engagement, especially in deep tech, especially in 2026. Moreover, if this were true, we would see a gradient of terms across the relative strength of the founders, just as we do (within reason) across venture financings. In reality, a bespoke service argument is being given for a one-size-fits-all pricing scheme.

I hope everyone had a happy Easter / are maybe enjoying Spring break. I’ve been thoroughly enjoying my third travel-free week before a month-long sprint (SF, SF again, NYC, Baja California (which is Mexico for my EU readers). There has been a huge amount of incredible news over the last few weeks so let’s dispense with the pleasantries and get down to it.

Stuff Worth Sharing

• Co-moon-ity service - Certainly the most significant story of late is the incredible work being done at NASA with the Artemis II Moon mission. The launch was an inspiring feat of ambition and engineering that brushed off a capability that sat idle for 53 years. NASA’s Space Launch System (SLS) is a 322-foot rocket powering Artemis II that consists of four high-powered engines capable of generating 8.8 million pounds of thrust. In a few short days the SLS will return our cosmic heroes from their 252,000-mile orbital journey, further than any human has ever been. The system represents a vertically integrated stack: propulsion, structural stages, and crew systems tightly coupled for long-duration missions. The mission is not without the controversy typical in mega projects like this. In total, the SLS program will run a projected $93 billion, with early missions costing roughly $4 billion each. But with the untold promise of lunar cities, multi-planetary colonization, and asteroid mining, this is, IMO, an investment worth making.

• Head in the cloud - My favorite recent addition to the reading diet is Owl Posting by Abhishaike Mahajan, who does incredible deep dives into some of the most interesting questions in technology. This post on lab automation got me hooked, as it’s an area I’ve long mused on as well despite limited success in investing against it. The essay lays out a pragmatic framework for understanding why lab robotics has underdelivered and where potential may still lie. The core distinction he makes is the difference between “box” systems, highly specialized instruments like liquid handlers, and “arm” systems, general-purpose robotics that connect entire workflows. Mahajan argues that the key bottleneck is economic, not technical: most lab protocols can be automated, but aren’t run frequently enough to justify the upfront setup cost, making human labor more efficient in exploratory research. As a result, progress clusters around three approaches: 1) simplifying the translation from human protocols to machine instructions, 2) improving physical integration of systems, or 3) adding intelligence for real-time adaptation. Though simple, the framework is quite sound. Despite the ~40 minute read time, the essay is remarkably accessible and enjoyable to read. I tend to lean pretty hard into the intelligence camp, but after years of exploration I broadly align with Mahajan’s ending argument that all roads lead to cloud labs, as I outlined in my previous investment thesis, albeit less skillfully argued.

• By the same token - One thing I absolutely did not have on my Bingo card was OpenAI mothballing their Sora product, but here we are. I’ve referenced Sora in multiple (Re)Share issues, but for those unfamiliar, this is (was) a text-to-video model with incredible versatility and widely praised performance. OpenAI abruptly shut down Sora after it became clear the tool was computationally expensive and strategically distracting. The company cites that despite early hype, usage quickly plateaued, all the while it lost ~$1M per day. Based on the insane arms race, I doubt the cash loss was all that important in the grand scheme. Rather, it seems to have come down to the opportunity cost of talent and GPUs. If the past few weeks have taught us anything, it’s that Anthropic is in pole position for the lucrative enterprise segment, and OpenAI isn’t calling it just yet. Video models require significantly more compute than language counterparts and that scarce GPU capacity is better directed toward agentic and productivity-focused applications, where demand is stronger and monetization is clearer. Beyond the obvious boardroom intrigue, this serves as clear founder evidence to the laziest investor question of all: why won’t [Incumbent] do this?

• Spill the gravi-tea - My buddy Finn Murphy of Nebular recently launched a podcast series, Forecast 2050, and it’s great. If you’ve heard about it, you likely saw his conversation with Tyler Cowan (insane get), but it’s this conversation with Starcloud CEO Philip Johnston is the standout of the season. Starcloud is a data center in space startup that is on a phenomenal tear. They recently announced a $170M Series A, making it a certified unicorn in addition to one of the most interesting dinner party topics you’ll have all month. Naturally, orbital compute is discussed and beyond SpaceX, I don’t think anyone is seriously pursuing this as thoughtfully as Starcloud, so worth a listen for those interested. But honestly, the whole conversation was super space nerdy in all the right ways: moon hotels, the Fermi Paradox, the Kardashev scale of civilization. A uniquely dense conversation and a new regular in my podcast rotation.

• Cosmic dustbin - Speaking of shooting stuff into space. If you’ve spent any time exploring the field, you’ve inevitably heard about space debris, the tragedy of the orbital commons. The Guardian published a fantastic infographic that walks through the history and current state of our celestial trash. The basic idea is that there is too much s**t circling the Earth - 30,000 objects to be specific - and that creates an ever-increasing risk of a collision chain reaction that could close off space launch and kill any hope of those moon hotels.

• Body double - An ambitious young company, R3 Bio, is pursuing a pretty extreme vision of longevity: creating brainless cloned bodies as a source of organs or even full-body replacement. Sounds about as dystopian as you can get and the logical end state of the Silicon Valley blood boy phenomenon. The idea is to engineer clones with severely reduced brain function so they would be biologically alive but lack consciousness, making them ethically usable for transplantation. In the near term, the company is focused on “organ sacks” in animals, using gene editing and cloning techniques to grow bodies optimized for organ production and testing. The longer-term roadmap involves scaling this to primates and potentially humans. This is quite far away, if not technically (cloning inefficiency, developmental defects), then certainly ethically.

• Experi-mint condition - Shelby from Compound penned a great essay on the benefits of human challenge trials. For those unfamiliar, a challenge trial is where volunteers are deliberately infected (under controlled conditions) with in-development therapeutics. Broadly, this approach is limited to infectious disease and vaccine development, but it’s argued that this is woefully underutilized as a tool in our medical arsenal. The core bottleneck is not scientific feasibility but operational scale and standardization, with challenge trials being slow, bespoke, and difficult to replicate across sites. The essay proposes industrializing the model and building dedicated infrastructure, standardized protocols, and participant pipelines that allow trials to run continuously rather than episodically. In doing so, timelines for vaccine and therapeutic validation could compress dramatically, especially in early-stage testing.

• Can’t compass on the opportunity - I know these articles have been unusually long, and this is the worst offender, but it was SO interesting. This guest post by Arena Physica CEO Pratap Ranade is a manifesto of sorts on why he started a company whose mission is to unlock electromagnetism, the “hidden substrate of modern technology.” Everything from wireless communication to radar to data centers runs on one of the many wave-based energy signatures, but largely remains black magic. Arena Physica’s thesis is that AI can overcome this bottleneck by learning the relationship between geometry and electromagnetic fields, effectively building a “Large Field Model” that develops intuition humans lack. As much as this is a hard pitch for his company, Ranade does a pretty solid job of giving Electromagnetism 101 and fuses in a handful of fun Claude tools that he spun up. If you can give this an uninterrupted 35 minutes, I guarantee you’ll learn something. And for my founder friends who might not care about wavelength modeling but are thinking about raising in the next 12 months, this level of vision clarity is a masterclass.

• AI better not imitate art - Since my wife is a novelist, I’ve often thought about what literature looks like in a post-AI world. It seems obvious that a good amount of written/produced content could be AI-derived in the not-too-distant future. Disagree? Go watch The Housemaid and then we’ll talk. Anyway, the publishing world had a massive frenzy the past few weeks over the horror novel Shy Girl after it was cancelled by its US publisher. A reader/journalist posted that a lot of the prose seemed like AI slop, which then kicked off a whole investigation, ultimately ending in the book’s demise. At the crux of the controversy is the erosion of trust between authors and readers. As generative models become increasingly capable of mimicking individual writing styles, distinguishing human from machine-authored text is difficult, if not impossible. There are a number of knock-on questions this creates. What gives art actual value? How would one actually defend themselves against such an accusation? Are third-party validators little more than grift? I would argue that art/writing/etc. will go the way of manual watches. The craft, heritage, and inherently human touch is what gives it meaning despite the functionality being absolutely obsolete. But then again, I sort of have to say that…

Portfolio Flex

• 9fin announced their $170M Series C, making it Fly VC’s second unicorn!

• Reimagine Robotics launched a new website featuring some of the crazy stuff they’ve been brewing.

Hello again from the West coast. This marks the first week in five that I haven’t traveled and it feels great. Inbox zero, consistent bed times, Moose. But my recent jet setting has been extremely productive with the highlight being our inaugural Fly VC Research Day held earlier this month in London. We brought together ~60 researchers and roboticists from across the UK to share what they’re working on and cross-polinate ideas. As the host and organizer I felt an incredible sense of pride, not just because the event went well but because of how strong the UK robotics ecosystem is. I kicked off Research Day after witnessing how much less connected the London scene is vs. the Bay Area and I’m glad that my hypothesis proved out. This was a fantastic event that has absolutely inspired me to do more. Still working on ideas but June seems like a nice time of year…

Stuff Worth Sharing

• Deliver a blow - This paper in Nature introduces enzymatic microbubble robots, which are biodegradable microrobots that combine propulsion, imaging, and targeted drug delivery into a single system. I’ve been quite fascinated with the range of drug delivery mechanisms for years now (most notably Lir), but this approach takes the title for Best Whimsy. The protein shells with embedded microbubbles and use enzymes to convert naturally occurring urea into motion, eliminating the need for external power. The two forms of navigation are even more of a head trip: magnetic guidance for precise steering, and chemotaxis, where robots follow hydrogen peroxide gradients emitted by tumors. The same microbubble enables ultrasound imaging for real-time tracking, and when triggered with a focused pulse, it collapses to generate cavitation forces that push drugs deeper into tissue.

• Watch this space - I’m currently doing a deep dive on physics-informed AI for a deal I’m looking at, which brought me down a particularly interesting rabbit hole. Researchers on the Laser Interferometer Gravitational-Wave Observatory (LIGO) project have been applying advanced AI to overcome some of the physics-derived measurement limits they face, and the results were pretty (Inter)stellar. The AI proposed an unconventional optical layout that initially looked incomprehensible to researchers but ultimately revealed a clever method for reducing quantum noise, thereby improving detector sensitivity by 10–15%. The research lead was quoted as saying, “If my students had tried to give me this thing, I would have said, ‘No, no, that’s ridiculous,’” which feels like a scientific manifestation of Move 37. Similar tools are also helping physicists analyze experimental data in fields ranging from particle-collider runs to dark-matter clustering. It’s in the realm of scientific discovery that the promise of AI is most clear. Not about taking jobs or eking out ever more incremental efficiency gains, but pushing the frontier of what’s possible past the cognitive limitations of our biology.

• Satellite years ahead - One of the more out-there (pun intended) tech ideas floating around today is data centers in orbit. I have a good friend who’s invested in that space (ok I’ll stop), but I’ve been generally skeptical on both economic and temporal grounds. But this interview with the king of moonshots, Elon Musk, has me rethinking a lot of those priors. This nearly three-hour interview with Dwarkesh and John Collison is an investment well worth the time, but if you’re only interested in the orbital data center piece, it basically comes down to speed. When you look at the insane growth rate of terrestrial data centers the view seems extremely counterintuitive, but Musk argues that we’re rapidly approaching the limits of our energy grid and that the combination of regulation, production capability, and NIMBYism will grind buildout to a halt. By comparison, orbital deployments are 10x cheaper to run due to endless solar access and perpetual cooling. I’m not about to start investing in entire compute rigs, but since listening to this podcast I can’t stop thinking about all the potential derivative opportunities an orbital buildout might create.

• If you give a mouse a synapse - Artificial intelligence is getting more complicated. Researchers at UC Santa Cruz have trained a lab-grown brain organoid derived from mouse stem cells to solve the classic “cart-pole” control problem, a benchmark for adaptive learning systems. Using an interface that reads and stimulates neural activity, the team applied an AI-guided reinforcement loop, with variable strength electrical signals acting as feedback to “coach” the neurons toward better performance. Performance improved from ~4.5% under random training to ~46% with adaptive training, suggesting that even small clusters of neurons can perform non-trivial computation when properly guided. The learning was short-lived, with performance degrading after 45 minutes due to an absence of long-term memory structures. But if adaptive computation is intrinsic to neural tissue itself, it opens up new ideas for learning how we learn.

• Take a byte out of crime - The promise of AI is to radically democratize intelligence and now that includes fraud. Generative models are now widely used to produce phishing emails, write malware, and automate reconnaissance, with estimates suggesting at least half of spam email is AI-generated. It’s actually the net-new attack vectors that make for the most interesting investment cases, in my opinion. Deepfakes have become particularly effective, enabling high-impact fraud targeted attacks rising from 7.6% to 14% year on year. More advanced capabilities like end-to-end AI-driven ransomware have been demonstrated in research settings but remain unreliable in practice due to hallucinations. For now, AI is a force multiplier, increasing attack volume rather than sophistication. But one thing that’s proven reliable in past tech cycles is that scammers are the vanguard.

• Shining, shimmering, splendid - A startup called Reflect Orbital is proposing a constellation of space-based mirrors to reflect sunlight onto Earth at night, with the goal of extending solar power generation and providing illumination on demand. The first prototype would deploy a 60-foot mirror in low Earth orbit, capable of lighting a three-mile-wide space, roughly as bright as a full moon. The economic model includes charging roughly $5,000 per hour for directed light, targeting solar farms and emergency use cases. As you might imagine, this is super contentious. Critics argue the system is fundamentally inefficient, requiring thousands of satellites to match even a fraction of daylight, and raises a myriad of concerns around light pollution and interference with astronomy. I know a few of the investors in this company, and I can tell you this is as badass as it sounds. It may also be the first step in a 10-step path that ends in a Bond villain.

• Private remembers club - Robotics pace car, Physical Intelligence (Pi), introduced a memory system for robot foundation models that enables them to handle long, multi-step tasks rather than isolated actions. By combining two types of memory, short-term visual memory that tracks recent observations and long-term textual memory that stores abstract summaries of task progress, the system allows a robot to maintain a coherent “narrative” over time. They can remember where objects are, what steps have been taken, and adapt when things go wrong. In testing, memory-enabled models were able to complete tasks like making a grilled cheese (why is it always sandwiches?), unpacking groceries, and cleaning an entire kitchen, while baseline systems without memory repeatedly failed or got stuck in loops. Notably, the system enables in-context adaptation, where robots change strategy after failed attempts instead of retrying the same action. For those that are less deep in this area, that last part of that sentence is a big deal. Adaptation and improvisation are arguably the single most important aspects of native intelligence that remain unsolved. Big things cooking here, and not just shitty sandwiches.

• Dazed and con-fusion - Germany’s dependence on Russian oil has been a hot topic over the past few years, and now they’re trying to rectify it. Recently announced, the nation seeks to build the world’s first stellarator-based fusion power plant at the former Gundremmingen nuclear site alongside Proxima Fusion, RWE, and the Max Planck Institute for Plasma Physics. The project takes a phased roadmap, starting with a demonstration reactor, Alpha, which is being assembled to validate manufacturing, supply chains, and system integration ahead of full-scale deployment. Germany is committing billions toward fusion as a strategic technology, which is incredibly exciting as an investor and ecology nut. The broader signal is a shift from experimental science to coordinated industrialization, which is something Europe has sorely needed for decades. There was a time when things like this could only be built in the US, and I hope that time has come to an end.

• Resurrection’s for the birds - Colossal Biosciences is a frequent flyer on (Re)Share, and it’s back in the news for perhaps the most famous of recent extinctions, the dodo. As a reminder, this ambitious biotech uses gene editing to recreate key traits in living relatives rather than truly resurrecting the species. We’ve covered wooly mice and direwolves in the past, but now they’ve set their sights on our avian neighbors, which is a notably harder challenge. Because birds cannot be cloned like mammals, the approach relies on engineering primordial germ cells from pigeons, the dodo’s closest living relative, and modifying them to express dodo-like characteristics before implantation into eggs. Editing thousands of genes to approximate extinct phenotypes has drawn criticism from scientists who argue these are engineered proxies, not true revivals. At the same time, ecological uncertainty remains high, particularly around how such organisms would behave or integrate into modern environments. The article goes on to point out that the more durable application may be leveraging these techniques to strengthen existing species rather than recreate lost ones.

• Ghost in the music machine - Art imitates life and tech imitates art imitating but also kind of mocking life? AI-generated “actor,” Tilly Norwood, released a debut(?) single and music video and a lot of people in the industry are not pleased. Developed by Xicoia Studios and Particle6, Norwood is positioned not just as a one-off novelty, but as the centerpiece of a “Tilly-verse” where AI characters can act, perform, and evolve across media formats. The project blends generative AI tooling with traditional production techniques like performance capture, where a human actor physically performs the role to bring the AI character to life. Despite the AI framing, the production very clearly states that 18 human contributors were involved in the product, which might become more of a norm than we think. This capability poses dozens of questions not just on what is art or concerns about our collective content addiction, but also a fundamental challenge to the idea of celebrity. Things are weird.

• Renewabro energy - In perhaps the most American thing I’ve seen in a month, Valar Atomics shipped a modular nuclear reactor on a C17 for the US military. ‘Murica! Pew pew!

Portfolio Flex

• Transcelestial announced a big deal to bring laser-based internet connectivity down under.

• Wayve is teaming up with Uber and Nissan to bring self-driving to Tokyo.

• Tangent Robotics continues to build in public, this time highlughting their kinematic dexterity capture.

Greetings from Berlin where I’ll be spending the next few days with my team before shooting back to London for our first ever Research Day. I’m extremely pumped for that, but I’ll save that for a summary post later. It’s been a very busy few weeks, as evident by my slightly longer than normal issue. I didn’t realize until writing this one just how much I have been reading about China lately. About half of the stories in this (Re)Share cover our Eastern adversary in some capacity and I really cover the gamut. No reason to keep you waiting any longer so I’ll shut up. Let’s get to it.

Stuff Worth Sharing

• If you don’t build it, they’ll never come - Friend of (Re)Share, Cam Watson, wrote a thought-provoking essay on the growing gap in US vs. China bio-economies, and it comes down to long-term capex. Watson argues that biotechnology should be understood not as a standalone sector, but as the chemical manufacturing layer of the electric economy, analogous to how petrochemicals underpinned the fossil-fuel era. He introduces a “bioelectric stack” composed of feedstocks, fermentation, molecular assembly, sensing (biology to electric signal), AI controls, and application. The core claim is that industrial success depends less on biological novelty than on whether these physical layers are built and coordinated as shared infrastructure. China, he argues, has invested early in instrumented fermentation capacity and feedstock systems, enabling control and automation to compound on top of real throughput. In contrast, much of the West has matured design and AI layers ahead of physical capacity, creating impressive insights but stalled scale-ups. The persistent gap in biotech outcomes, he suggests, is a sequencing and coordination problem, not a scientific one.

• Empire Stateless Building - In addition to a FAT investment round, OpenAI announced a partnership with Amazon to launch a Stateful Runtime Environment for AI agents running natively on Bedrock. While most agent prototypes rely on stateless APIs that handle one prompt and one tool call at a time, enterprise workflows typically require persistent context, multi-step orchestration, error handling, permissions, and other complexities. The new runtime embeds state directly within a customer’s AWS environment, allowing agents to carry forward working memory, tool outputs, and workflow history across long-running tasks. This reduces the need for teams to build custom orchestration layers and makes multi-system workflows more feasible. It’s been a tough week for Anthropic, the up-until-very-recently de facto leader in enterprise AI adoption, but this gives the Altman crew quite a tailwind.

• Me(n)tal detector - The teams from Goodfire Research and Prima Mente published an interesting piece applying mechanistic interpretability techniques to Alzheimer’s detection models trained on blood-based biomarker data. More specifically, the signal came from patterns in circulating cell-free DNA fragment lengths in blood samples. Rather than treating neural networks as black boxes, the team probed internal activations to identify which latent features correlated most strongly with early disease signals. By isolating specific neuron clusters linked to fragment length signatures, they were able to both improve classification performance and provide a clearer biological narrative for why the model was making certain predictions. These experiments showed that distinct internal circuits were encoding biologically meaningful DNA fragmentation. The key contribution is not just improved detection accuracy, but a causal mapping between internal representations and measurable biological markers, which is an objectively powerful unlock given the difficulty of quantitative disease detection in neurodegenerative conditions.

• Remember trade disagreement - Now, if you don’t know what I mean by “difficulty of quantitative disease detection in neurodegenerative conditions,” here’s why. The Alzheimer’s Association’s 2024 criteria define the disease biologically, based primarily on the presence of amyloid in the brain. In contrast, the International Working Group requires amyloid, tau, and cognitive symptoms. The disagreement matters because 25–45% of cognitively normal seniors have amyloid, and many never develop dementia. Critics warn that relying on amyloid alone risks over-diagnosis and unnecessary treatment with monoclonal antibodies, which carry risks such as brain hemorrhage. Supporters argue that Alzheimer’s pathology begins years before symptoms and that early identification could enable earlier intervention, likening it to diagnosing diabetes before complications arise. The debate ultimately hinges on semantics and risk framing: does amyloid signify disease or elevated risk? Until trials clarify whether treating asymptomatic patients improves outcomes, we won’t have a definitive answer. (Cough, TimeTrace Labs, cough).

• The Little Reasoning Engine That Could - Google has released a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode designed for science, research, and engineering problems. The model now sets a new high-water mark on several frontier benchmarks, including 48.4% on Humanity’s Last Exam, 84.6% on ARC-AGI-2, and gold-medal-level performance on the 2025 International Math, Physics, and Chemistry Olympiads. Deep Think is positioned as a practical research assistant, and the blog post includes spoken testimonials from notable researchers who put the model to use. Those stories include what you might expect: identifying logical flaws in advanced mathematics papers and modeling complex physical systems through code. But what really caught my eye was the third example from Anupam Pathak. He was able to upload a simple image to create a lightweight CAD file suitable for 3D printing. I’ve been digging into CAE for a while now (thesis repo very overdue), and this is a big, big deal. I’ve been toying with the idea of getting myself a 3D printer, which would be as unnecessary as it is awesome.

• Punches above their model weights - If you’re worried about AI and robotics destroying our career paths, add prize fighter to the list. San Francisco-based startup Rek is drawing paying crowds to watch VR-controlled humanoid robots box in a cage, using modified Unitree G1 robots imported from China. The 4.5-foot, 80-pound machines fight in five one-minute rounds under full sporting theatrics, complete with referees and commentators. Tickets run $60–$80, and organizers plan to expand into a formal league, including larger 200-pound robots. The event follows the launch of China’s first humanoid “combat league,” backed by Shenzhen-based Engine AI. Good old-fashioned mechanical carnage.

• Dance Dance Communist Revolution - Oh and also theater.

• Continue to brain supreme - Last month marked two years of Neuralink’s Telepathy program, the implanted brain–computer interface designed to restore digital and robotic control for people with paralysis. With 21 participants enrolled globally, users now achieve cursor control speeds comparable to able-bodied mouse users, with some exceeding 10 bits per second and typing up to 40 words per minute via a “ten-finger” imagined keyboard. Participants have used the system for education, creative work, and robotic arm control, while a new clinical study (VOICE) aims to restore real-time speech for ALS patients. Neuralink plans to scale implant chip size from 1,000 to 3,000 electrodes to increase signal volume and improve thread retention. They have also refined surgical techniques to reduce invasiveness following early issues with thread retraction. Signal quality has improved across 18 of the last 20 implants, and the company reports zero serious device-related adverse events. Musk is a divisive figure these days, but these stories always make my heart warm.

• Satellight reading - Blue Origin unveiled TeraWave, a space-based communications network designed to deliver up to 6 terabits per second of symmetrical global coverage for enterprise and government customers. The constellation will be comprised of 5,408 optically interconnected satellites across low and medium Earth orbit, using high-performance RF links for up to 144 Gbps per user. TeraWave emphasizes symmetrical upload and download speeds, route diversity, and redundancy for critical operations, positioning itself as a fiber complement rather than a retail broadband competitor. The network is designed for rapid deployment of enterprise-grade terminals and resilient backhaul in regions where fiber is constrained. Deployment is slated to begin in Q4 2027. For fellow frequent fliers, these orbital internet angels can’t come soon enough.

• A battery ram - Battery makers are rapidly retooling North American factories from electric vehicle cells to energy storage systems (ESS) as EV demand slows and AI-driven data center construction surges. Ten plants are being converted after companies cancelled enough EV capacity to build 2 million cars, with seven facilities now focused primarily on storage. The shift reflects a demand drag that’s been evident for years in the US, but the loss of the $7,500 tax rebate last September really solidified fears. BloombergNEF cut its 2030 US EV sales forecast from 48% to 27% of new car sales. Generous US manufacturing credits ($35/kWh) and tariffs near 60% on Chinese imports have improved domestic ESS economics, though Chinese suppliers still lead in lithium iron phosphate performance. As a soon-to-be EV leaseholder (recommendations welcome), I’m not so bummed to see a short-term buyer’s market.

• Spilling all the tea in China - China’s AI companion market is rapidly professionalizing around a surprising demographic: Gen Z women. Platforms like MiniMax’s Xingye allow users to customize multimodal “AI boyfriends,” building persistent emotional relationships that extend beyond chat into voice calls, physical merchandise, and even real-world cosplay dates. One user profiled spends roughly three hours per day interacting with her AI partner and hires a human cosplayer to embody him in person. Chinese AI companion apps explicitly target women, which sharply contrasts with the male incel stereotype we all think of with AI companions. Considering the gender imbalance in China, this story really caught me off guard. I don’t want to sound like a Tinder maximalist, but these ladies should be cleaning up.

• Book report - I recently finished Breakneck by Dan Wang. I’ve on and off read his annual letters, but it wasn’t until a strong endorsement from my friend Alex that I picked this one up. Breakneck is a field report of sorts on what it feels like to live inside China’s sprinting, improvisational model of growth. Wang argues that China’s edge is not a single “secret technology,” but a dense ecosystem that rewards speed and production above all else, a mental framing he calls “The Engineering State.” He contrasts the strengths and weaknesses of this one-sided mindset with the inventive but otherwise bureaucratic modern America. I have not yet read Abundance, but I would imagine this is a nice pairing. While Wang is a Chinese native, he doesn’t pull any punches. China’s strength in scaling, supply chains, and a culture of relentless optimization is matched only by the relentless human costs required to sustain this pace, and by the limits of growth-by-hustle as demographics, geopolitics, and capital constraints tighten. His story of living through the CCP’s zero-COVID response to the pandemic it created, and the resulting shattering of worldviews, is particularly remarkable. This was my first five-star read in some time, and I highly recommend it.

Portfolio Flex

• The self-driving juggernaut can’t be stopped. Wayve announced a $1.2 billion funding round.

Back again, a bit later than I intended. The best part about being a VC alongside Gemini Deep Research is you can learn about anything. The worst part is that there’s always more stuff to read. After penning this issue, I had to prairie fire my backlog, lest the reading anxiety get the best of me. Had to push through, as I have a five-week travel sprint coming up. If you find yourself in SF, London, or Berlin, let me know. Actually quite a bit of exciting stuff coming up. On Feb 28th, I’ll be at the AI + Bio hackathon in Berlin (we’re a sponsor), and on March 5th, I’m hosting Fly’s very first Research Day, which so far has been super fun and may become a regular thing. But now, on to the news of the day. But first, a quick shill…

Prideful Plug

Steph’s fourth book, Revenge Plot, is now out for pre-order and it is a good one! For avid readers, this is a very inside-baseball type of novel with loads of references to the publishing industry itself and the toils of making it as a writer. Pre-orders are actually quite a significant thing, so if you’re interested, I would highly suggest you shell out the $30 for an early copy of what will surely be a Top 10 Netflix original in a few years (US, UK).

Stuff Worth Sharing

• You gotta rub me the right way - We talk about DeepMind’s Genie work a lot because it’s awesome and deserves the attention. Last week, the experimental prototype was opened up to Google AI Ultra subscribers in the U.S. (🎻🇪🇺), allowing users to create and explore infinite, interactive worlds generated in real time by the latest world model, Genie 3. I’m not an Ultra user, but I’m seriously considering it just to make some exploration stories with my dog, Moose. Even if you don’t pony up, the demos are worth a watch. Sketch a world using text or images and Genie dynamically simulates it as you move through it, like a simplistic video game. Under the hood, Genie 3 simulates physics and interactions on the fly, which is really impressive, especially for downstream robotics and unfamiliar environment navigation (deep sea, underground, disaster zones). Worlds are still capped at 60-second generations, so no quitting your job for latent AAA studio dreams just yet, but it’s a very exciting moment. I’m extremely interested in world models and their latent, almost unimaginable potential.

• Computey pageant - Regular readers will know I’m a big fan of Invest Like the Best, and long-time readers will know my single favorite guest is Gavin Baker. He’s an encyclopedic resource on everything that orbits semis and compute, which is arguably one of the most valuable knowledge bases to have right now. This December episode dives into the current state of the compute wars, and the 90-minute conversation is a must-listen. The exchange is wide-ranging but packed with sharp, standalone segments. Early on, Baker breaks down the economic strategies behind each frontier lab: OpenAI needs premium pricing while Google can afford to flood the market through vertical integration. I also appreciated the strawman bear case he lays out for how the AI race could go sideways. But the highlight for me was the segment on in-orbit data centers. I’ve been pretty dismissive of the idea, but I have to hand it to him—he makes a compelling argument, and this was weeks before the xAI / SpaceX news. I don’t think I can invest in this space directly, but the tangential opportunities might just sneak their way into the 2026 edition of Hard Problems.

• Going broker - This isn’t exactly deep tech, but if you’ve ever muttered “there’s no real productivity in AI,” look no further. CH Robinson, the logistics giant, now runs 30 in-house agentic AI tools and they’ve been the primary driver behind its 55% stock price run last year. For those unfamiliar with traditional third-party logistics (3PL), it’s a wildly manual industry full of very colorful characters. My post-college roommate ran sales for a 3PL, and his stories were frankly insane; Wolf of Wall Street isn’t far off. Historically, human brokers could only respond to ~60% of the company’s 600,000 annual freight requests, with turnaround times of 17–20 minutes. The AI now responds to 100% of queries in just 30 seconds. Another agent continuously adjusts pricing strategy, tweaking margins hundreds of times per day instead of monthly or quarterly. That’s a performance delta that’s simply untouchable. RIP bullpen.

• Kill two birds with one drone - Animal lovers, maybe skip this one. A Russian neurotech company, Neiry, has started flight tests in Moscow using pigeons outfitted with implanted brain–computer interfaces and turning them into “biodrones” for surveillance. Electrodes placed in specific brain regions connect to a solar-powered backpack with a flight controller, GPS, and camera. Instead of training the birds, the system stimulates neural circuits to guide their movement, letting operators steer them remotely. It’s the kind of dystopian scheme only Russia ( maybe Mr. Burns) could dream up. Neiry says the pigeons can follow preloaded flight paths like traditional drones, but with far greater range and endurance at similar cost. Future plans target ravens, seagulls, and albatrosses for infrastructure, environmental, and security monitoring. Submitting a claim to PETA now.

• Know when to fold ‘em - A macro critique of robotics today is that we’re rich with demos but short on real deployments but DYNA Robotics begs to differ. Sacramento-based Monster Laundry partnered with DYNA to address a key bottleneck: folding. As the pickup-and-delivery laundry service scaled, folding emerged as the most labor-intensive constraint, limiting same-day throughput. The key word here is throughput, which too many robotics founders underestimate and which I’ve highlighted before. Over a three-month pilot, the DYNA system folded over 200,000 garments across ten clients, saving hundreds of labor hours and boosting commercial capacity by 25% with a 99% quality acceptance rate. In the before times, as an acne-prone youth working in restaurants, I can attest that the absolute worst task was rolling silverware into place settings. If robotic intelligence unlocks nothing more than eliminating basic work like this, the world should still rejoice.

• Sandwich con-artist - With the Winter Olympics in full swing it’s important to remember that there too the robots are coming, sort of. In December, Physical Intelligence unveiled work that tackles Moravec’s Paradox through the lens of the “Robot Olympics,” a benchmarking gauntlet of real-world physical manipulation tasks like opening doors, turning keys, peeling fruit, and cleaning pans. These were powered by Pi’s large-scale vision-language-action architecture, enabling a single model to attempt a broad range of tasks with minimal tuning. Tasks were scored Olympic-style (gold, silver, bronze), and while baseline vision-language models flopped, Pi’s pretrained system showed strong performance and quick adaptation, often learning behaviors in just a few trials. While Fly has made investments that represent alternative bets to the Pi approach to general intelligence, I want to commend this work for its rigor and creativity. The key-turning task in particular hit home - still a daily struggle. That said, awarding a silver medal to the peanut butter sandwich shown in their post is culinary terrorism of the highest order.

• Embody building - Not to be outdone but pretty late to the party. Microsoft has entered the robotics fray with Rho-alpha (ρ-α), its first robotics model derived from the Phi vision-language family. Rho-alpha takes a VLA-style approach to control signal generation, combining vision and language with inferred tactile feedback to perform tasks like plug insertion, switch flipping, and toolbox packing on dual-arm and humanoid platforms. I remain skeptical of VLA systems in production. Not for lack of ability but because their computational overhead introduces meaningful latency. That said, the tactile approach here is genuinely compelling. Instead of embedding force sensors, Rho-alpha learns to infer force through trajectory prediction. Clever, though still compute-heavy. Microsoft is also building tooling for in-deployment human corrections, a nod to Reimagine Robotics. Still in the evaluation phase, but definitely one to watch.

• Jedi mine trick - An area that has been en vogue lately within deep tech VC circles is mining tech. As demand for nickel, copper, and rare earth elements surges, the general awareness of metals supply chains has significantly increased. Fun fact: we need a lot more than we have. Some Michigan miners are turning to biotechnology to extract more metal from aging mines and lower-grade ore. A startup is testing a fermentation-derived broth that removes impurities from concentrated ore, allowing continued nickel production even as ore quality declines. Other companies have leveraged DNA analysis to manage microbials in copper for bioleaching heaps. There’s actually a surprising amount of activity within engineered microbes, much to the chagrin of the microplastics waste crowd. Economically, this approach makes enormous sense - extracting as much as you can from a sunk-cost asset. But that level of extraction is really an open question. I just struggle to see how recycling or efficiency plays are the ones to back in a build-out super cycle. Then again, zig where others zag…

Portfolio Flex

• Lir Tx penned a new blog post, highlighting the incredible work they’re doing to screen capsid structures at lightning speed. RIP Boltz2.

• Zymtrace posted a case study that shows the 250% inference speed improvement they delivered to a design partner.

• GitGuardian announced a $50 million Series C!

Hello again and welcome to another volume of (Re)Share. Word on the street is that JPM was a wide-ranging success and bio may be back! Or at least for as long as the Silicon Valley reality distortion field holds. I, for one, will be spending the next two weeks shamelessly shilling my portfolio wares. Anyway, let’s get to it.

Stuff Worth Sharing

• Out of embody experience - An NVIDIA + Stanford collab released PointWorld, a physics-implicit world model that can map out per-pixel motion forecasts from just a few RGB-D images and a basic prompt. PointWorld leverages a new framework for training large models to reason over long horizons by explicitly modeling uncertainty and intermediate states throughout the simulation, rather than relying solely on chain-of-thought text, which can be quite brittle. Instead, reasoning is devolved into a sequence of latent state transitions, where the model predicts not just the next action but the distribution over possible future states. This is how we naturally anticipate action/reaction dynamics, but that’s because we’ve grasped an implicit awareness of gravity, deformability, and rigid actuation (like a hinge). This work reinforces the argument that better reasoning and resulting manipulative capabilities won’t come from longer explanations, but from models that maintain a stronger internal sense of state and direction. Paper here.

• Th-Inc. different - It’s Davos week! The one time of year when European leaders get to beat their chests on the world stage and while it may not be deep tech news, perhaps the biggest European story is the formal anointment of the EU Inc. proposal by Commission President Ursula von der Leyen. For the uninitiated, EU Inc. is a years-long push to modernize and unify the fragmented company formation processes across the 27 EU member states. That may not sound exciting to my American readers, but try investing in a German GmbH and then tell me this isn’t overdue. Leyen’s speech is worth watching in its entirety and covers urgency, cross-border collaboration, unlocking capital, and throws a bit shade at protectionist thinking. If you’re short on time you can find the relevant snippet here. Like most VCs in Europe, I’ve been a strong supporter of EU Inc. and am thrilled to see it finally gaining real momentum.

• Trajectory of the commons - Some research on motion planning brought me to this talk by Yilun Du of Harvard, where he summarizes his work on compositional models. His body of work argues that energy-based models (EBMs) offer a unifying alternative to today’s dominant feedforward and diffusion-style approaches, particularly for tasks requiring reasoning and generalization. Instead of directly predicting outputs, EBMs learn an energy landscape that scores how compatible inputs, outputs, and constraints are, then solve problems by iteratively minimizing energy at inference time. In other words, what is the minimum effort required to satisfy objective functions with n constraints? A lot of Du’s early talk highlights the origins of diffusion models, but later it moves to the potential of motion control/trajectory planning, which naturally lends itself to robotic manipulation.

• FAIR and balanced - Meta signed big boy agreements with three nuclear energy providers to secure more than 6 GWs of electricity for its insatiable data center appetite. Nuclear has been a less-covered topic here as of late, but the demand for 24/7 baseload hasn’t gone anywhere. Vistra will supply 2.1 GW from existing nuclear plants in Ohio and Pennsylvania under a 20-year contract, with upgrades adding another 433 MW in the early 2030s. Meta also committed to 1.2 GW from Oklo’s planned SMRs and up to 2.8 GW from TerraPower’s sodium-cooled reactor design. Electrons may be a commodity, but the relative demand, surge profile, and local cooling are very idiosyncratic. SMR startups remain the most interesting as a case study to see just how hard we can push scale economics.

• This shit is nano bananas - DeepMind stepped up their video-generation game with the newly released Veo 3.1. With the core messaging of “Ingredients to Video,” the demo (and later my X feed) is ripe with pretty wild capabilities. There’s other stuff in the release like improved character identity and object persistence, but I really fell for the near-complete removal of creativity friction. The model also better blends disparate visual elements (characters, textures, backgrounds) into unified clips with viable motion and detail. As someone with 0% artistic sensibilities, I am fascinated by the emergent cultural behaviors that come from mechanistic explosions like this.

• Get a gripper - Now, if you find yourself dreaming about the limitless possibilities of simulation training, let’s throw some water on that in the form of a 55-page blog/tome. Evan Beard, the CEO of Standard Bots, penned a comprehensive (and I mean it) essay on his belief that general autonomy can only be achieved through incremental mastery of increasingly disparate skills. That broad task coverage can only be reached through real-world deployments—which is admittedly less sexy but infinitely more rational. I loved this essay so much, and it is well worth the time investment, but if you’re not feeling that, the crux of the argument is this: robotic systems improve through exposure. Early deployments tolerate imperfection, and failures generate data that rapidly drives learning and robustness. You learn more from what didn’t work than from what did. Part of my enjoyment is how much Beard’s macro thinking reflects my own and the more recent bets within the Fly portfolio. But this is also an extremely in-depth description of why he built his company the way he did. This depth of conversation is something I’m fortunate enough to have regularly as a VC, but that’s a luxury few enjoy. It also features what is, without question, the most unique robotic use case I have ever heard of.

• An open relationship - Good friend of Fly, Tim Rocktäschel, gave this recorded interview a year ago, but I recently rewatched it because of how thought-provoking and wide-ranging it is. Longtime readers may remember Tim’s name from the multiple times I’ve referenced his team’s work on GENIE. The discussion centers on the idea that today’s AI systems plateau because they optimize toward fixed objectives, whereas intelligence in nature emerges from open-ended processes that never truly converge. This is fitting, as Tim leads DeepMind’s Open-Endedness (OE) team. He argues that an AI system should be considered open-ended if it continuously produces outputs that are both novel and learnable to an observer, which is distinct from randomness (novel but meaningless). One of the more interesting parts touches on a question many founders I know have asked themselves: how do you optimize for something with no reward function? And, to a lesser extent, a more recent fear - will scaling laws hit hard limits? Rocktäschel suggests that open-endedness is the missing ingredient on the path to more general AI. Combining large language models with open-ended mechanisms (self-play, environment generation, and self-improving prompts) could enable agents to explore, set sub-goals, and discover solutions humans never specified. The analogies to biological evolution aren’t lost—and in fact, are mentioned frequently throughout, wich makes the whole talk a lot more accessible.

• Intern-al hard drive - Anthropic dropped Cowork, a Claude Code cousin for normies. A familiar chat interface plus local access will let anyone do basic read/write/organize tasks in a more embedded manner. Pleb work. The demo’s cool but nothing mind-blowing, which is a reflection of just how quickly we can normalize to magic. But given the dominance that Code has had in macro software engineering, this is a notable development. Lots and lots of memes of Seed stage startups getting rugged, etc. etc.

Portfolio Flex

• Tangent Robotics came out of stealth and I could not be more excited about it. I’ve been known to speak in hyperbole from time to time, so allow me to do it once again: if fine-dexterity robotic hands are a technical possibility, this is the team that will build it.

Happy New Year and welcome to the first issue of 2026! I hope you all had a festive holiday season, which I certainly did. As you may remember, my wife and I hosted Christmas for the first time and I’m pleased to say it was a success. Fun fact, if you cook a 14 lb turkey for a family that doesn’t really like turkey it will take you about 2.5 weeks to finish eating said turkey. But I digress. There’s been a lot of news since we last got together. No, I’m not referring to the increasingly aggressive return to imperialism, decay of the rule of law a

nd the justification of military action from men that would probably sprain their wrist if they ever fired a gun. Let’s get to it.

Shameless Plus

• The case against accelerators - I penned a bit of a hot take dunk on programmatic company building and why they decline in relevancy, at least from a VC sourcing perspective. Apart from a few select strategies, accelerators were a product of the times and no longer make sense.

Stuff Worth Sharing

• Human resources - If you’re worried about being overtaken by AI-powered robotic overlords, fear not. They still need us—at least for a little while longer. Physical Intelligence shared research demonstrating that human-to-robot skill transfer emerges naturally in large vision-language-action (VLA) models once they are trained on sufficiently large and diverse robot datasets. Egocentric human videos, long considered difficult to use due to functional domain gaps, can significantly improve robot performance without any explicit transfer mechanisms. By simply co-fine-tuning a scaled VLA model on human videos and relevant robot data, performance on generalization tasks (bussing, sorting, tidying) improved by 2x. The result mirrors emergence in LLMs and suggests that scaling robotic foundation models unlocks new data sources—making cheap, abundant human video a viable training signal for real-world robots.

• Lost and sound - Researchers at the University of Connecticut have developed a programmable acoustic metamaterial capable of reshaping sound waves in real time. The system consists of an 11×11 grid of asymmetrical pillars, each individually rotated by a motor in one-degree increments, allowing precise control over how sound is bent, focused, or dampened. Unlike traditional fixed metamaterials, the platform can be reconfigured instantly without remanufacturing, enabling it to adapt across frequencies and functions. Experiments show it can concentrate sound at specific points for applications like ultrasound imaging, acoustic tweezers, and targeted therapies, or guide waves along edges in a manner analogous to topological insulators. The team argues the vast combinatorial design space makes the material a candidate for AI-driven optimization, pointing toward future intelligent materials that autonomously tune themselves to achieve desired wave behaviors.

• Tatooiner take all - Nobel Prize–winning chemist Omar Yaghi is betting that metal-organic frameworks (MOFs) can unlock a new way to produce potable water directly from air, even in deserts. If that sounds like Uncle Owen’s moisture farm - well, yes. Through his startup Atoco, Yaghi is developing machines that use MOFs’ vast internal surface areas to capture water molecules at low humidity and release them with minimal energy input, potentially using only sunlight. Unlike conventional atmospheric water generators that rely on energy-intensive refrigeration or membranes, MOFs can operate efficiently in much drier conditions. Atoco plans both grid-powered industrial systems capable of producing thousands of liters per day and off-grid units for remote regions, with early field trials slated this year in the Mojave Desert. There are certainly limitations to the technology today, but I get excited when I read research like this. It holds the potential to significantly reduce reliance on desalination and centralized water infrastructure—issues that many of my more environmentally focused investor friends are extremely concerned with. On multiple occasions I’ve heard the argument that water could drive geopolitics for the next century, and based on how this issue is shaping up, that’s not far-fetched.

• Material evidence - AI-driven materials discovery is learning the lesson that techbio did a few years ago: simulations alone aren’t enough. Despite the hundreds of millions invested into models and the (proposed) millions of hypothetical materials discovered, very few have translated into useful, real-world substances. When bits try to innovate atoms, it’s the synthesis and testing of materials that counts—and that is a slow, expensive process. A new wave of startups like Lila Sciences, Periodic Labs, and Radical AI are building AI-directed, autonomous laboratories that combine robotics, high-throughput synthesis, and LLM-based agents to plan, run, and analyze experiments, with the promise of compressing materials timelines from decades to years. Long-time readers will recall that I’ve opined on lab automation myself, and it remains one of my top priorities.

• Weird science - My friends at Compound wrote a great piece on the promise and playbooks for decentralized science (DeSci)—the innovation that results from applying decentralized / crypto primitives to scientific discovery. I’ve always been a little skeptical of the investment viability of decentralized orgs (call me old fashioned), but far be it from me to disagree with Shelby or 0XSMAC. The piece makes some interesting arguments on the natural advantages of such a structure, especially for relatively fixed-term projects where traditional capital markets fail. DeSci enables rapid access to global capital, larger and more diverse contributor bases, and new incentive structures for data generation, experimentation, and peer review. It has already shown traction in underfunded areas like women’s health or rare disease. For the traditional scientists in my readership, the playbook makes it clear that DeSci is not a replacement for traditional biotech, but a complementary model that turns science into a more programmable, participatory, and capital-efficient system.

• Simulate-breaking news - Some back-to-work preparation for a portfolio company brought me to this paper, and despite being a complex topic (pun intended), the explanation is quite accessible even for those less familiar with the space. Cosmohedra is a physics-based method for rapidly predicting the structures of very large, symmetric protein complexes—an area that has long been a blind spot for AlphaFold-style models due to memory and computational limits. Cosmohedra assembles complexes using geometric and physical heuristics that scale beyond the somewhat arbitrary ~5,000 amino acid ceiling. Across 36 benchmark complexes, the method achieved near-experimental accuracy, approaching the resolution limits of cryo-electron microscope data. Structures with up to ~40,000 amino acids were built in minutes on consumer hardware, making Cosmohedra 10³–10⁵× faster than molecular-dynamics docking methods.

• Intelligentle giant - One of the biggest stories coming out of CES last week was the announced partnership between DeepMind and Boston Dynamics. The Gemini Robotics model will be deployed within Atlas humanoid robots and will be trialled in a Hyundai auto factory (an investor). This is a pretty obvious team-up in the making, not just because of past ownership ties. The Atlas robot (and most of the BD fleet) is the undeniable leader in physical capability, but they’ve broadly been absent from any meaningful application because they lack contextual awareness and object-manipulation skills. While Atlas already excels at balance and acrobatics, Gemini’s multimodal perception and decision-making framework is promised to allow Atlas to identify unfamiliar objects and adapt to changing environments. This is the obvious strategy for every robotic intelligence play, and while BD is a big get, the sheer cost of those embodiments makes all but the highest-volume tasks economically unviable. Still, an exciting development to watch from the sidelines.

• Scaffold under the pressure - The age-old design principle “less is more” has made its way to LLMs. Vercel reports some interesting results from rebuilding its internal text-to-SQL agent, d0. By removing most of the agent’s specialized tools, they made it significantly faster, simpler, and more reliable. The original system relied on heavy prompt engineering, schema-specific tools, and guardrails to prevent hallucinations, but it proved slow and fragile. Vercel gutted the entire setup and gave Claude direct access to raw analytics files and only a single capability: execute bash commands. The model now explores data the way a human analyst would, and the stripped-down agent achieved a 100% success rate (up from 80%), ran 3.5× faster, and used 37% fewer tokens. As models improve, complex agent systems can become a liability, and well-structured data plus minimal tooling often beats elaborate orchestration.

• The doctor won’t see you now - OpenAI has launched ChatGPT Health, a dedicated experience designed to help users better understand and manage their health—while preserving the privacy protections that I personally have completely thrown in the trash. The amount of mental and physical health information I’ve dumped into OpenAI’s data coffers is…concerning. In this new architecture, Health lives in a separate, isolated space within ChatGPT, allowing users to securely connect medical records and wellness apps to ground conversations in their own data. Conversations in Health are not used to train foundation models and are protected with additional encryption and access controls. Except for Europe—no soup for you.

• Eggsecutive decision - Through both personal and professional experience I can say with extreme clarity that IVF is a broken system, so anytime I see interesting research pushing the boundaries in that arena, I perk up. Scientists in China, the US, and Europe have now created the most realistic lab models yet of early human pregnancy by combining human embryos (and embryo-like “blastoids”) with uterine organoids grown inside microfluidic chips. These systems recreate implantation—the moment an embryo attaches to the uterine lining—allowing researchers to directly observe a process that normally happens unseen inside the body. Across three Cell Press papers, embryos burrowed into endometrial tissue and initiated early placental development before experiments were halted at or before the 14-day ethical limit. Researchers ran thousands of trials using blastoids, enabling large-scale studies of why implantation often fails during IVF. In one experiment, screening 1,119 approved drugs identified compounds that increased implantation rates from ~5% to ~25%.

Portfolio Flex

• Microsoft gave some nice coverage on Wayve, describing the history of the contrarian bet they took and announcing an expanded scope of their Azure partnership.

• Some new additions to the Fly portfolio - (publicly) welcome Clicks, xMemory, TimeTrace Labs and Facet!

Another year, another revision to my annual Hard Problems post summarizing the areas I’m most excited to invest in over the coming twelve months. As I look back through the 2025 post, I am reminded that this job is really just about humans. Nowhere in my list was there any mention of digital biomarker detection, but it represents my most exciting investment from the year. In reality, these Hard Problems lists are a snapshot of where my curiosity is at this particular point in time. Some interests will prove correct (robotics), others will prove intractable (bio-production), and many will remain somewhere in the middle of “could be huge but damn it’s hard.” With that, here’s what I’m looking at in 2026.

• Robotic data infrastructure - Over the past 18 months we (Fly VC) have aggressively invested against the emerging robotics stack (R2, Tangible, P&W). An increasingly clear reality is that we need a shitload of data to unlock intelligent manipulation and that data requires radically new infrastructure. Everything from raw generation, parsing, annotation, search and recovery, etc. is up for grabs.

• Energy-resistant compute - Last year, I wrote about “physics-proof compute,” but I was really talking about the limitations of grids in the face of unending demand. Electrons still represent a top-three challenge for compute buildout plans, and this will significantly accelerate if we hit really cold winters that pit data centers against the people. The obvious challenge with anything compute-related is the sheer size of capital requirements. As a sub-$100M fund, we need to be conscious of what games we should be playing in. With that said, over the last year we’ve deployed against this thesis (Zendo, Zystem) and think there is a lot of room to go.

• Neurodegenerative treatment - I sort of teased this in the intro, but this is a pretty active area of exploration for me. Unfortunately, I have firsthand experience with this in my family (Alzheimer’s, ALS), as do millions of people every year. The level of spend and the affected population vs. the sophistication of treatment is absolutely crazy. We are in desperate need of new solutions. When you factor in the aging population trends, subscale care infrastructure, and the emotional reverberations that these diagnoses have on families, that makes for a very good area for venture cash.

• Rethinking clinical trials - Hundreds of startups have tried and failed to disrupt the clinical pathway. The single most common datapoint I read across my dealflow is, “it now costs over $1 billion to bring a drug to market.” Everyone sees this problem. But just because everyone accepts the issue doesn’t make it consensus investing. There is a very strong contingent in the venture community that clinical trials are impossible to fix—and frankly, they have data on their side. But there are signs that this time might be different. The political pressure to reduce drug prices has hit a tipping point, and the US is joining the rest of the West in strong-arming pharma to act. The basic economic reality is that unless we tank innovation, the only sustainable way to reduce drug prices is to reduce drug development costs. There are loads of methods to do this (reduced admin burden, improved patient selection, secondary endpoints, etc.). The FDA and CE have also taken a markedly different stance toward animal testing, which should inject some much-needed energy into burgeoning new areas like in-silico digital twins, organoids, and digital biomarkers

• (Self-driving) Lab automation - I remain bullish on this and have been actively exploring it for some time now, though I will say I’m less confident on what the right play is. Despite the surge in funding for AI Scientist plays, there are few solid answers to how this wave will right what TechBio got wrong. Selling software to pharma is a hard way to make a little money, but developing assets requires the gauntlet of clinical trials (see above). I do think there is a big opportunity to be found within a cloud lab / CRO 2.0, especially for novel molecule synthesis. Chemical/material science is interesting and certainly a more open field. We’ve seen some great success with our investment in Orbital, and that’s just one of hundreds of application areas.

• Agent societies - I don’t yet have a good articulation of my interest here, so I’ve kept it intentionally broad. I’ve written about Agentic Payments already, which will absolutely be a massive opportunity area. Pricing models, governance systems, and identity management are all open questions, although this is generally a more consensus category now. More recently, I find myself thinking about macro orchestration. In many of the cognitive advancements in recent LLM research, you’ll read about chain-of-expert models and collaborative reasoning. If you expand this to its logical endpoint, you would expect a massive volume of specialist agents that self-organize to accomplish a task at hand. This is an orchestration problem at its heart, and it seems that our existing infrastructure is not fit for purpose.

• AI-native discovery - Search and shopping behaviors are rapidly being rebuilt. If attention is the ultimate commodity in the internet age, you want to be the marketplace, not the marketer. There are so many subsegments of this area, and one could plausibly have an entire fund dedicated to investing across the stack—GEO/LLM discovery, agent-native data packaging, wallets and governance rights, agent identities. There is already a lot of investment being plowed into this, so competing requires a particularly aggressive type of founder team from day one.

• Sledgehammer AI - For fun, I’m throwing in a catch-all term for when a founder seeks to absolutely destroy a market. This isn’t just vertical AI or full-stack startups. I’m looking for pirate founders whose stated mission is to conquer. Rare in Europe, but I’d guess there’s one or two.

Fly VC was built to find, back and support the extreme technical talent that can will the future forward so if that sounds like you please reach out.

Greetings from what is likely the last Re(Share) of 2025. It has been a challenging year for a number of reasons, but one of my true highlights has been this newsletter and the small but mighty community I’ve been able to build through it. It has been particularly fun to find readers with like-minded curiosities about robotics, biology, AI, and everything else we cover here. This year I set out to be more consistent with a bi-weekly cadence, and I mostly delivered on that—until Q4, where the wheels fell off. But assuming I survive the chaos of hosting Christmas this year, I’ll be back at it in 2026. Happy holidays to all and enjoy the latest and greatest of what I’ve read.

Stuff Worth Sharing

• Can you be more Pacific? - Naval forces are racing to deploy autonomous underwater drones as a cheaper, scalable way to track submarines and protect critical undersea infrastructure. Driven by lessons from Ukraine and recent attacks on pipelines and data cables, the UK, US, and Australia are committing billions to uncrewed underwater vehicles (UUVs). Australia, for example, is coughing up $1.7 billion on Anduril’s Ghost Shark, and England is building out fleets of contractor-operated drones to monitor Atlantic chokepoints. UUVs offer persistent surveillance at a fraction of the price of crewed submarines. We are structurally anti-weapons as an investment fund, but the potential for massive growth is very clear to me. If these sea dwellers behave anything like their aerial cousins, maritime security will be completely reframed in a few years.

• Chinese fire drill - The bane of Silicon Valley is back! DeepSeek released two new open-source models, DeepSeek-V3.2 and V3.2-Speciale, that claim to rival or exceed GPT-5 and Gemini-3.0-Pro on reasoning, math, and coding benchmarks. The Speciale variant achieved gold-medal performance across elite math competitions, and the standard model delivers near-frontier performance at much lower cost. The impressive release is underpinned by the DeepSeek Sparse Attention (DSA) architecture, which cuts long-context inference costs by ~70%. The article highlights 128k-token contexts for roughly $0.70 per million tokens, which is insane. The V3.2 models also feature “thinking in tool-use” capability, which reasons through problems while simultaneously executing code, searching the web, and manipulating files. A parallelized chain-of-thought of sorts. Both 685 billion-parameter models are free to download, modify, and deploy much to the chagrin of SF.

• That’s that me espresso - General robotics intelligence is the biggest game in venture right now, and Physical Intelligence (Pi) is in the pole position. Last month, Pi released their work on iterative policy generation through human influence and posed the argument that vision-language-action (VLA) robots fail not because imitation learning is weak, but because it breaks down under compounding real-world errors. Their solution, RECAP (RL with Experience & Corrections via Advantage-conditioned Policies), mirrors how humans learn: instruction, coaching, then practice. By providing human corrections when they err on manipulation tasks, Pi’s robots were able to rapidly self-improve via reinforcement learning guided by a learned value function. The learning function was applied to everyday, multi-step tasks like making coffee or folding laundry, with the π* model more than doubling throughput and cutting failure rates by 2x across >90% success rate. The paper’s core claim is that reliable robotics won’t come from bigger models alone, but from systems that can learn continuously from their own physical experience, turning deployment itself into the primary data engine. On this we (Fly) absolutely agree and have aggressively invested against this very thesis, which should be coming out of stealth very soon.

• A helping hand - Pi isn’t the only barista in town. In what may be the cutest robotic demo yet, startup Sunday Robotics unveiled Memo, a fully autonomous home robot designed to handle everyday chores like clearing tables, loading dishwashers, and making coffee in real, messy kitchens. Rather than a walking humanoid, Memo uses a wheeled base with an adjustable column and two dexterous arms, prioritizing stability and reach over bipedal movement. Sunday pays remote workers to perform household tasks while wearing sensorized gloves that closely mimic Memo’s novel hand design, generating higher-fidelity training data than traditional teleoperation. This allows Memo to perform surprisingly nuanced actions, such as grasping multiple glasses in one hand. This really made the rounds in my robotics circles, and the bar for product demos has been seriously raised.

• Forget it - GLP-1s may be the darling of the therapeutics world, but they’re not infallible. Novo Nordisk reported that its EVOKE and EVOKE+ phase 3 trials of oral semaglutide in early-stage Alzheimer’s disease failed to meet the primary endpoint, showing no statistically significant slowing of disease progression versus placebos over a two-year time scale. The double-blind study enrolled 3,808 patients with mild cognitive impairment or mild dementia and was among the largest Alzheimer’s trials run outside the amyloid field. While semaglutide improved several Alzheimer’s-related biomarkers, these changes did not translate into clinical benefit. Novo will discontinue the extension phase and take a pretty massive bath on the whole effort. GLP-1s may still be valuable risk modifiers in neurodegeneration, but this readout sharply reduces the probability that they are stand-alone, disease-modifying Alzheimer’s drugs.

• Confestivus for the rest of us - OpenAI developed a technique that trains large language models to confess when they lie, cheat, or violate instructions. Instead of trying to prevent bad behavior outright, the approach asks models to generate a structured self-assessment after completing a task. The secondary process rewards honesty alone, rather than helpfulness or safety, which pushes models to admit shortcuts or deception. GPT-5-Thinking was put to the test and came up with some pretty creative solutions to bending the rules. There’s a certain beauty in teaching LLMs ethics in the same way we might a toddler, although I wouldn’t do either myself.

• Rained supreme - DeepMind has released WeatherNext 2, its advanced AI weather-forecasting model. The model can generate forecasts 8x faster than previous versions and produce hundreds of coherent scenarios from a single starting point. WeatherNext 2 outperforms DeepMind’s prior state-of-the-art model on 99.9% of variables and lead times (0–15 days). Weather is a fascinating computation problem because of the massive amount of interconnected system complexity, referred to as marginals and joints, apparently (fun fact). DeepMind overcame this by injecting noise directly into the model to preserve physical realism across interconnected weather systems through a Functional Generative Network. Brick Tamland, eat your heart out.

• Getting one’s dust deserts - One of my more controversial investment interests is geoengineering. While I’ve yet to deploy against this thesis, a few friends have, so I’ve been learning through osmosis. Stardust Solutions is positioning itself as the first serious for-profit player in solar geoengineering, aiming to cool the planet by dispersing sunlight-reflecting particles into the stratosphere. The company claims its proprietary aerosols could counteract centuries of greenhouse-gas emissions and has a fresh $60 million to prove it. Stardust argues that climate impacts are accelerating faster than mitigation efforts, making geoengineering a necessary emergency option—which is my view as well. Unlike some of the more rebellious players in this space, Stardust says it will only proceed with government approval and proper oversight. It’s a pretty sensible position to take, considering the only viable customer base will be governments facing climate-induced economic or geopolitical fallout. The sustainability space is in the shit right now, at least from a venture perspective, but that won’t last forever.

• It’s a material world - ARIA, the UK’s answer to DARPA, is an organization I hold in extremely high regard and have been lucky enough to work with through their robotic dexterity program. Recently, they’ve unveiled the next wave of opportunity areas, and the Universal Fabricator caught my interest. The thesis attempts to reboot materials innovation by inventing scalable processes to manufacture complex, hierarchical materials. The program proposes building integrated hardware–software platforms that can discover and refine processes for producing hierarchical nanocomposites with performance driven by internal structure across nano-to-centimetre scales. ARIA is fascinating because it perfectly fills the gap between academic research and traditional venture capital. I really couldn’t be more excited about seeing where this funding ends up.

• Cutoff genes - Doctors in Manchester have reported early results from a world-first gene therapy for Hunter syndrome (MPS II), a rare inherited disorder that causes progressive physical and cognitive decline and is often fatal before adulthood. A three-year-old boy became the first of five patients worldwide to receive the experimental treatment, which modifies his own stem cells to produce a missing enzyme in the brain. One year after treatment, the boy is developing normally, which is nothing short of a medical miracle. The trial, nearly abandoned due to funding shortfalls, is now being watched closely as a potential blueprint for treating other rare genetic neurodegenerative diseases.

Portfolio Flex

• Wayve released GAIA-3, their most advnaced world model yet.

• Lir launched a fancy new website.

6 weeks since our last (Re)Share! That’s really stretching the bounds of an “occasionally weekly curation” so I have to start with my apologies. The last month+ has been wild with the bulk of that being on the road in SF (2x), Berlin, London and Buffalo, NY for a funeral. I’m very much enjoying a couple weeks in home base before I inevitably get back to the grind of trying to sell cash to the most technical founders I can find. Fortunately the world is ripe with exciting new research, discoveries and juicy court cases so let’s get to it.

Stuff Worth Sharing

• Right hand man - For those in the know, I have been digging into robotic manipulation for a while now, and that brought me to discover just how insanely complex the human hand is. It’s common to hear that the opposable thumb is what really set us on the genetic path of dominance but, no—it’s the whole five-fingered miracle. With humanoids gaining more and more perception capability, it was only a matter of time before dexterity became the limiting factor. The fine folks at Boston Dynamics released their 2nd generation end effector for their Atlas robot, the GR1, and it is pretty awesome. This three-gripper model is particularly interesting for reasons that I can’t reveal for a few more weeks, but let’s just say we have a very vested interest in this segment of robotics (cough….cough). Besides the insights on mechanistic design and the complexity tradeoffs of multi-digit form factors, this quick watch might teach you a thing or two about your own grippers.

• Bunsen burn the midnight oil - Two of my recent deep dives have been lab automation and AI-generated scientific discovery, both of which I’ve written about in my Thesis Repo. One of the most exciting developments of the year was a July paper from USTC. BioMARS is a robotic platform that uses a multi-agent architecture (LLMs, VLMs, and dual-arm robots) to autonomously design, plan, and execute biological experiments. The system consists of three independent but collaborative agents: a Biologist Agent that synthesizes protocols from literature, a Technician Agent that translates protocols into robotic code, and an Inspector Agent that applies multimodal error detection during experiment execution. When applied to cell culture passaging, BioMARS showed comparable results to manual experimental reproducibility and reduced hands-on time by ~90%. While the paper does the normal academic thing of listing out every possible limitation, the significance of this work—and the broader AI Scientist movement—should not be overlooked. By enabling compute with scientific manipulation, we can drastically reduce the throughput limitations of skilled wet lab technicians. An autonomous biofoundry, at least at the preliminary viability level, is within reach.

• Wait Watchers - Last month, Dwarkesh hosted Andrej Karpathy on his podcast, and it is absolutely worth a listen. Despite being one of the leading minds of the LLM frontier, he is remarkably accessible and surprisingly muted in his AI fervor. Karpathy argues we are roughly a decade away from the true promise of agents because today’s LLMs still lack the ingredients needed to behave like real workers— long-horizon reasoning, memory, continual learning, and robust self-correction. His main premise is less about the technology and more about the public’s unrestrained expectations. AI progress arrives in seismic shifts, but each shift still demands years of engineering to turn demos into durable products. The pod is over 2 hours long, so if you’re short on time, I’d highlight the discussion on the limitations of reinforcement learning (40:05). RL is like sucking supervision through a straw and fails to train or learn in the multi-dimensional refinement that humans do via culture. Improvements via memory, self-play, and multi-agent systems will help, but that’s what creates the decade(+) timeline.

• Partners in crime - California just became the first state in the US to regulate AI companion chatbots. SB 243, taking effect January 1st, holds companies legally accountable when their bots fail to meet safety standards for children and vulnerable users. The law mandates age-verification, clear disclosures, self-harm and suicide-prevention protocols, break reminders for minors, a ban on sexual content with minors, and unambiguous statements that the companion is AI-generated. Penalties can reach $250,000 per offense. I expected far more pushback from the major platforms, but given the string of recent horror stories, mounting an aggressive defense would have been politically toxic.

• It pays to discover - Edison Scientific, the new commercial spinout of FutureHouse, launched Kosmos, a next-generation AI scientist that uses structured world models to handle thousands of papers and tens of thousands of lines of code per run. Kosmos can replace about six months of human scientist effort, and the system has already reproduced three known discoveries and made four novel ones across domains from neurobiology to materials science. For example, the system independently established a human-data-validated mechanism in Alzheimer’s entorhinal neurons. Most exciting is that the Kosmos product is run by my good buddy Ludo, who is featured in this fancy overview. Proud of you, dude!

• Sunlight’s the best disinfectant - A team at UT Austin and the University of Porto may have found a way to turn cancer treatment into something closer to a consumer-electronics problem. They’ve built a photothermal therapy that uses tin oxide nanoflakes, which are cheap, stable, and absurdly thin, to soak into tumors and convert near-infrared LED light into lethal, hyper-localized heat. No lasers, no collateral damage. Their entire illumination rig costs $530, yet in early tests it vaporized 92% of skin cancer cells and half of colorectal cancer cells in 30 minutes while leaving healthy tissue untouched. It’s early-days in vitro work, but the direction of travel is obvious: cancer treatment pushed down the cost curve, out of the hospital, and potentially into at-home light patches that zap residual tumor cells after surgery. If even part of this vision holds, we’re looking at the rare oncology breakthrough that’s not just effective, but universally accessible.

• A picture’s worth 1,000 words - An interesting paper from DeepSeek features a surprisingly simple idea that could reshape how AI systems “remember” by ditching text tokens and storing information as images. Instead of breaking language into thousands of expensive tokens, a new OCR-based model compresses written content into visual snapshots, then retrieves them when needed. This approach packs far more information into far fewer tokens and introduces a tiered memory system: older or less important content gets stored in slightly blurrier form, reducing compute while remaining recoverable. Early benchmarks shows on par performance, but the real breakthrough is this visual-token memory architecture, which could extend context windows, reduce compute bills, and generate massive amounts of synthetic training data. This isn’t a solve for reasoning or dynamic memory, but it hints at a future where LLMs store knowledge more like brains do: richly, efficiently, and visually.

• Grand theft autonomous - One of my more controversial takes on the Fly team is the belief that we need to have better systems for content protection and management. While I will 100% admit this has been a complete vacuum of returns in the past, but capitalism does have this small dependency on property rights. Anyway, Reddit has launched a major copyright lawsuit against Perplexity, accusing the AI search startup of illegally scraping user conversations to train its answer engine. The filing also targets three data-scraping groups, alleging they helped Perplexity launder scraped content by masking bots as real users. Reddit framed the case as evidence of an “industrial-scale data-laundering economy” as AI companies race for high-quality human text. Notably, Reddit already licenses its data to Google and OpenAI for millions, so there’s a quantifiable price of the heist. Reddit recently sued Anthropic for 100,000+ scraping incidents, and this new case adds pressure on AI companies to secure licensed data pipelines—or risk expensive litigation. I’ll bring the popcorn.

• Book report - Not exactly deep tech news, but somewhat considering my newsletter readership - and predominantly considering the book was a 400-page slog - please humor me. Liu Cixin’s sci-fi classic The Three-Body Problem was… a lot. Wang Miao, a scientist drawn into a secret project linked to China’s Cultural Revolution and strange suicides among physicists. Through an addictive VR game about a chaotic alien world with three suns, he uncovers a looming invasion by an alien civilization seeking a new home. On Earth, secret factions either resist or collaborate, hoping to reshape humanity. Now thanks to some GPT-5.1, that summary sounds awesome. In reality, the depth of world-building and scientifically robust plot devices come at the expense of seemingly any character development—save for one loudmouth beat cop. I already know I’ll get trolled for this, but IMO this book was Three Body Problematic.

Portfolio Flex

• CloudNC got some great coverage in Fortune on how they’re revitalizing CNC production.

• The early days origin story of Lir Tx was covered in The Currency.

• Fly portfolio company, Palisade was acquired by Ripple.

• An undisclosed investment may have had some cool research published in Nature.

Gearing up for a three-week travel spell so if you’re in SF, London or Berlin (in that order) hit me up and we can hang. Been spending a lot of time testing my thinking on Lab Automation, so thanks to all those from the (Re)Share community that have reached out to share thoughts. Another healthy mix of deep tech fun is ready for your reading pleasure so let’s get to it.

Stuff Worth Sharing

• Tomb raider - Despite Bryan Johnson’s best efforts, longevity is still a pretty fringe topic within the scientific community (also, save yourself an hour and skip his terrible Netflix doc). But there was some pretty interesting research published in Nature earlier this month. A landmark meta-analysis of 15,000 human tissue samples has produced the most comprehensive epigenetic atlas of aging to date, mapping DNA methylation changes across 17 organ types and 900,000 genomic sites. If you’re unfamiliar with DNA methylation, welcome to the club. This refers to a natural process by which certain enzymes recruit proteins to block transcription factors from binding to the DNA—i.e., silencing that particular gene. This process accelerates over one’s lifetime and is believed to be a major driver of biological degradation as we age. The study shows that some tissues age much faster than others. For example, the retina (63% methylation) and stomach (57%) accumulate changes rapidly, while the cervix (39%) and skin (48%) show slower epigenetic drift. While most tissues show increased methylation with age, skeletal muscle and lung lose methylation, reflecting distinct organ-specific ageing trajectories. This is particularly powerful for life extension advocates who believe the secret to anti-aging therapies lies in moving beyond treating single diseases and instead focusing on aging systemically.

• Where there’s a will there’s a Wayfarer - The money pit of AR/VR is back. At Connect 2025, Meta unveiled the Ray-Ban Display, a new iteration of AI glasses that fuse a full-color, high-res in-lens display with cameras, mics, speakers, and onboard AI. Hidden in these slightly enlarged spectacles is a glance-based display designed for short interactions like checking messages, navigation, or visual prompts. I’m personally more interested in the pseudo-haptic control system via the Neural Band. This is an EMG wristband that interprets subtle muscle signals for hands-free scrolling, clicking, and “hand”writing. The launch videos are pretty neat, and navigation for city dwellers seems like a particularly fun feature. I am surprised to see a $799 (glasses + band) price tag, considering how grand of a flop Vision Pro was. Sure, this is a far cry from those $3,500 but it’s still far above the impulse-buy threshold.

• Gene shorts - Biotech UniQure had a big few weeks coming off the fantastic results of their AMT-130 gene therapy clinical trial for Huntington’s disease. In a phase 1/2 trial of 29 patients, those receiving the higher dose showed a 75% slowing of disease progression over 36 months versus natural history controls, as measured by the Unified Huntington’s Disease Rating Scale. They also experienced a 60% slowing in functional decline and favorable biomarker trends, including an 8.2% drop in neurofilament light protein (had to look that up!). The therapy, delivered once via neurosurgery, is designed to silence the mutant huntingtin gene. Neurodegenerative diseases are a frequent topic here, and Huntington’s is a particularly brutal condition, so this really does represent a landmark in gene therapies for the field. This new data pave the way for a Biologics License Application in Q1, and a commercial launch is targeted for next year.

• Ticket to free ride - Last month, the European Commission launched a call for evidence on the Carbon Border Adjustment Mechanism (CBAM) pending legislation, seeking input on three key issues before the regime begins in January 2026. Stakeholders are invited to weigh in on the methodology for calculating embedded emissions in goods, how to adjust CBAM certificates to reflect the free allowances still granted under the EU Emissions Trading System (ETS), and how to deduct any carbon price already paid in third countries. The EU gets a lot of shit for over-regulation, generally warranted, but on this, I have to commend them. To be sure, any compliance will incur costs and administrative burden for companies, and on some incremental level that will make it harder to compete. At the same time, there is really no viable path to meaningful decarbonization / emissions reduction without a carbon tax. Historically, this would be a political death sentence for the American right, but in a world of overt protectionism and ever-shifting tariff games, I’d be curious to see how something like this might do on the Hill.

• Brain freeze - Until Labs recently unveiled a whitepaper on their work in cryopreservation of organ donations. The company claims that they’ve recovered electrical activity from cryopreserved and rewarmed acute rodent brain tissue. Using 300 μm cerebellar slices, the team loaded tissue with a cryoprotectant (–196 °C) and later rewarmed with induction heating. Of four slices tested, all showed action potentials post-thaw, with responses that confirmed physiological activity. To be clear, we’re not even remotely close to electrical activity comparable to fresh slices (gross), but this provides early evidence of tissue reanimation. The company recently announced a fat Series A from Founders Fund and Lux Capital to solve organ transplant gaps and one day enable medical hibernation. I’m not going to comment on the latter, but the former is something I can really get behind. Most organs remain viable for only hours after procurement (4–12 for hearts, lungs, and livers; 24–36 for kidneys), which leads to thousands of organs lost every year due to logistics. If you’ve ever had a loved one have to endure the transplant list, you will immediately see the incredible potential this technology could have.

• Red flag-ged little pill - The recently announced link between Tylenol use during pregnancy and autism is, and I believe this is a medical term, bullshit. This article in Scientific American provides a good summary of studies over the past few decades and highlights the strong connection of various factors (parental age, prematurity, maternal diabetes). Tylenol has a handful of historical studies that suggest modest increases in autism risk, but the largest recent study of 2.5 million Swedish births found no causal link, with sibling comparisons completely eliminating any signal. Tylenol is often used to treat fever or infection, both independently linked to higher autism risk. I’m sure most readers will already know that the American College of Obstetricians and Gynecologists strongly rejects claims of causation. Please pop pills responsibly.

• I’ve scene better days - Fei-Fei Li’s World Labs introduced a new pipeline for large-scale synthetic world generation, aimed at accelerating training for embodied AI and robotics. By combining procedural generation with generative AI, the system can create vast, diverse, and customizable 3D environments that simulate real-world physics and accurate layouts. The obvious near-term use case is rich training data for navigation, manipulation, and multi-agent interaction tasks that would otherwise be prohibitively slow or expensive to cultivate. World Labs goes further to argue that better synthetic worlds are essential to advancing autonomous agents through probed reasoning, planning, and adaptability. It’s also a pretty fun to play with all the options on Marble.

• A mind like a steal trap - We all love a good police chase, and we may have a lot more coming our way. Flock Safety, known for its police surveillance tech, is pitching its drone program to private businesses, potentially bringing aerial patrols to shopping centers, warehouses, hospitals, and even retail chains. With FAA waivers for beyond-visual-line-of-sight flights becoming easier to secure, companies could install rooftop docking stations and dispatch drones when alarms trigger. In practice, drones would follow suspected shoplifters, even tracking vehicles, with video feeds streamed to security teams or directly to police. Crime doesn’t pay, but it does entertain.

Summer is behind us and the startup world is returning to full speed. The past few weeks have been a busy one for Fly. We welcomed a new Associate to our ranks and there’s been a lot of big developments on the portfolio side, as you’ll see below. But of course what brings us together, as always, is the range of exciting developments in the world of deep tech. Let’s get to it.

Stuff Worth Sharing

• The doctor won’t see you now - A frequent topic at (Re)Share is the diverging behaviors of AI adoption in the East vs. the West. China’s country’s strained healthcare system has created fertile ground for AI chatbots like DeepSeek, which patients increasingly treat as trusted medical companions. The article follows a kidney transplant patient who now relies on DeepSeek for symptom analysis, lifestyle advice, and emotional support, even reducing prescribed medication based on its suggestions. For her, the bot feels “more humane” than her overworked doctors. This dynamic reflects a wider trend: AI tools are stepping into gaps left by absent caregivers and overstretched health systems, offering 24/7 availability and empathy. This read put me in a weird place within the uncanny valley. On the one hand, digital access to medical care is one of the single biggest impacts that AI could have on the world. It is just not feasible to provide specialist-level coverage to every patient that needs it, and that’s especially true in more rural locales. At the same time, as the article shows, the risks of hallucinations, bias, and patient misdiagnosis are very real. Model overconfidence comes with real-world costs, and there is only so much training that can be done for real patient engagement—at least ethically.

• Kill two birds with one clone - The undisputed leader of synthetic human video has extended its lead once again. Synthesia unveiled its Express-2 avatars offering, pushing their digital clone capability closer to photorealism. Compared to the earlier Express-1, the new system offers smoother body movements, more natural gestures, and voice cloning that preserves accent and intonation. Journalist Rhiannon Williams provides a head-to-head of the two versions using the same base training, and the improvement is quite significant. The gain is attributable to the infallibility of scaling laws, where the Express-1 model had a few hundred million parameters, Express-2’s rendering model’s parameters number in the billions. It still strikes me as odd that a technology this powerful has found the killer app to be corporate training videos, but so speaks the market. Synthesia claims that the next frontier is interactivity, with avatars that can “talk back” in real time and answer questions mid-video. At which point, I would predict live commerce comes to the West in a big way.

• Emergency exit - Annoyingly talented writer/investor Mike Dempsey penned another great piece, this one on the changing dynamics of investing in a world of talent-based arms races and HALO’d (Hire-And-License-Out) pseudo-exits. AI startups like Inflection, Adept, Character, and Windsurf raised vast sums only to be acqui-hired by larger labs with questionable return structures. While these transactions may be justified as talent aggregation, Dempsey argues that HALOs reveal many startups were strategically mispositioned—e.g., betting billions on frontier LLMs without sustainable moats. Inflection, for example, raised over $1.3 billion and amassed a 22,000 H100 arsenal but couldn’t compete with OpenAI, Anthropic, or Meta on compute, talent, or consumer distribution. Character.ai faced similar scale issues despite strong usage. Fast-followers chasing platform shifts with massive early raises often end up boxed into one existential shot, burning capital and talent when incumbents consolidate. In AI, the winners are often first movers or last disruptors—not the rushed middle cohort.

• Come brain or shine - Two Canadian men in their 30s with spinal cord injuries have become the first patients outside the U.S. to receive Neuralink brain implants. Implanted at Toronto Western Hospital, the chip enabled them to control a cursor with thought alone within just minutes of surgery, and both were discharged the next day. The Canadian trial will monitor up to six participants for a year, assessing safety and quality-of-life improvements while watching for risks like seizures or infection. Health advocates urge caution, and a prior U.S. patient experienced implant slippage—the technology remains experimental. Still, the speed of life impact in these two surgeries is materially faster than the previous implantations we’ve covered in past issues. Some of that is due to the nature of the injury, but the pace of technological improvement is remarkable.

• Getting high on your own supply - Long-time readers will know that I am not particularly bullish on the viability of humanoid robotics, and this piece in IEEE Spectrum somewhat explains why. Humanoid robotics startups have promised factories of tens of thousands of robots, but the scaling challenge remains seriously unresolved. Agility Robotics claims its Oregon facility can build 10,000 Digits a year, and Tesla projects 50,000 Optimus units by 2026. But today, only a few robots are deployed and all in tightly controlled pilots. The bottleneck isn’t manufacturing, it’s demand. Companies haven’t found applications requiring thousands of humanoids in a single facility, and multipurpose AI is not yet robust enough to fill the gap. Market viability also hinges on a multitude of critical constraints: battery life, reliability, safety evidence, etc. This is the Perez curve being drawn in real time. I, for one, think that outside of consumer care (child/elderly), there is no environment that can’t be somewhat easily adapted to an embodiment form. Personally, I would much rather bet on basic arm functionality with a highly dexterous hand—but I may be a bit biased ;)

• Embody mass index - This is a great reflection piece from Decoding Bio on the state of biotech capacity in China, from a scientist on the ground. Cam Watson argues that the next biotech leap won’t come from in-silico prediction tools like Atomwise or AlphaFold but from machine intelligence and robotic labs. We’ve covered the potential of autonomous science in a number of past issues - both in coverage on leading labs like FutureHouse as well as my own writing. But I have always had a Western lens, and Watson shows just how shortsighted that was. China, he argues, is uniquely positioned to lead this shift in AI-generated discovery. State-backed investment, national biofoundries, industrial robotics expertise, and reliable grid infrastructure have created fertile ground for embodied AI platforms like BioMARS, which orchestrates multi-agent AI “scientists” in fully automated labs. Integrated biotech campuses such as Suzhou’s BioBAY show the model’s commercial viability with frankly shocking success rates—15 therapeutics to market, 54 unicorns, and 17 IPOs! He goes on to argue that biotech is about to have its internet moment and that China’s scale and speed could set the global agenda.

• Bad Apple - Researchers from Apple pressure-tested the capabilities of Large Reasoning Models (LRMs) vs. predecessor LLMs, and the offspring didn’t do so hot. The models were tested on controllable puzzle environments like Tower of Hanoi and River Crossing to allow for precise control of complexity and direct analysis of reasoning steps (vs. math benchmarking). Findings revealed that on low-complexity tasks, standard LLMs (Claude 3.7 Sonnet) often outperform LRMs (DeepSeek-R1 and OpenAI o-series) by solving puzzles faster and with fewer tokens. At medium complexity, LRMs gain an edge, leveraging longer chain-of-thought to navigate harder problems. But at high complexity, both collapse—with accuracy dropping to zero. Strikingly, as problems grow harder, LRMs sometimes reduce reasoning effort, cutting short their “thinking” despite having unused token budgets.This last bit is a big, big deal as it suggests a fundamental scaling limit in inference-time reasoning. The study underscores that today’s reasoning models don’t possess generalizable problem-solving skills but rely on brittle heuristics that break under compositional depth. While reinforcement learning and self-reflection improve mid-range performance, LRMs remain bounded by structural limits. The authors warn against conflating “thinking tokens” with true reasoning capacity: more text is not equivalent to deeper cognition.

Portfolio Flex

• Lakera announced a $300 million acquisition by security leader, Check Point. Fly has been lucky enough to sit alongside the team since day one and couldn’t be more proud of them!

• Blue Morpho shared a sneak peak of their knowledge-graph powered solution for finance.

• Agemo introduced CodeWords: the fastest way to go from idea to automation, simply by chatting with AI.

Happy Labor Day weekend! Steph and her parents are all traveling so Moose and I have spent the weekend living the bachelor life. And by bachelor I of course mean John Wick movies and several bags of microwave popcorn. I haven’t left the house in days and I apologize for nothing.

Shameless Plug

• My latest addition to the Thesis Repo is out and this time I focus on Geoengineering, which are engineered solutions to alter the climate at will. I’m increasingly cold on anything built on a sustainability bend. That’s not to say these startups aren’t viable, but I no longer have faith in capital markets or broad strokes commercial adoption to warrant (my) venture dollars. Does this all mean that climatetech is an extinct investment space? No but I think the smarter bet is on fail-safe solutions that will almost certainly be needed when climate tipping points are reach in the next 5-15 years. TL;DR - invest in the stuff of Bond villains.

Stuff Worth Sharing

• Just the tip - Speaking of climate collapse. The UK’s Advanced Research and Invention Agency (ARIA) has launched a program to build an early warning system for climate tipping points. The immediate focus is on the Greenland Ice Sheet and North Atlantic Subpolar Gyre. These systems are at risk of crossing self-perpetuating thresholds, with debated but quite serious consequences like sea-level rise, disrupted ocean circulation, and food insecurity. As readers may remember, I was a participant in last year’s Robotic Dexterity program, and I found the whole experience extremely impressive. I am super bullish on innovative public/private models like ARIA, and I wish more hubs would adopt similar approaches. This Tipping Point program will fund a range of low-cost sensing systems and AI-enhanced models to detect early signs of catalyst events. If successful, the effort could transform climate forecasting, offering policymakers trusted, real-time signals to accelerate mitigation, adaptation, and resilience planning against high-impact, low-likelihood events.

• Lobby Hobby - One of the most prominent players in the Geoengineering segment is Rainmaker and this intro video developed by The Abundance Institute is pretty solid. There is very little about the company’s technical approach to cloud seeding (fascinating by the way) but that’s not the point of the piece. Instead there is a heavy emphasis placed on the required work CEO Augustus Doricko has had to do to lobby state legislators to not kill the technology before it gets off the ground. This legal inertia is a big risk to underwrite for an early stage VC, but I’m increasingly of the view that leaning into that is where all the alpha resides. I would have preferred less of the chest-thumping anti-China undertones, but that’s what works these days.

• The good subsidie young - 200+ companies and associations signed an open letter to the European Commission urging immediate action to secure sovereign digital infrastructure. The EuroStack initiative argues Europe is at risk of near-total dependence on non-EU technologies within three years without a decisive industrial strategy. I have mixed feelings about this letter. At a spirit level, the message is clear: Europe must shift from regulation to action or risk permanent digital dependency. On this, I wildly agree. Digital intelligence is as much a strategic priority as steel or energy or defense. In a world of Allied free trade, interdependency was viable, but the political winds have clearly shifted, and if America pursues an isolationist strategy, Europe obviously needs to adjust. At the same time, national industrial strategy never works—especially in Europe. The stated goals are noble, but falsified demand and cross-country asset pooling is such a central planning mindset. There are certainly capital challenges, but the real driver of growth is going to come from the levers that EU nations have always struggled with: deregulation and tax incentives.

• Don’t shoot the messenger - A recent paper in Nature described the COMET project, a transformer-based AI model for designing lipid nanoparticles (LNPs), which are one of the key delivery vehicles for RNA medicines. Unlike previous models that focus on single lipid molecules, COMET encodes full formulations: lipid identities, ratios, and synthesis parameters. The model was trained on LANCE, an LNP dataset of over 3,000 formulations and (I think) the largest in the world. COMET achieved strong predictive accuracy and screened 50 million LNPs, all in silico. Experimental validation confirmed that COMET-designed LNPs showed higher protein expression, with some candidates surpassing benchmark formulations by >40-fold. I’m biased towards a preference for AAV delivery (Lir Tx), but RNA is such an exciting modality that all news is good news.

• A mind like a steal trap - The battle of AI vs. IP rights had an interesting turn this week, with one AI giant dodging a major bullet. Anthropic reached a preliminary settlement with a group of authors, averting what could have been a financial catastrophe. The class action alleged Anthropic pirated works from “shadow libraries” like LibGen to train its models and therby exposing it to statutory damages starting at $750 per book across an archive of ~7 million titles. That put potential fines near $1 trillion, enough to make even Altman blush. The presiding judge had earlier ruled Anthropic’s model training was largely fair use, but allowed the piracy claims to proceed to trial. The settlement, expected to finalize September 3, keeps terms under wraps for now but represents a major development for the dozens of ongoing AI copyright battles that hinge on how training data is sourced and justified.

• Head in the cloud - I’ve been exploring the BCI space for the past two weeks as part of some investment due diligence I was running. This brought me to a recent article in Stanford Medicine that describes some new research on detecting inner speech. To level set, a lot of BCI technology, both invasive and non, involves attempting to decode the neural patterns associated with speech so that the signal can be converted into text and then later audio speech. But what about our only inner monologue? Though clearly related, inner speech does not trigger many of the follow-on brain activities for vocalizing (e.g., breath and tongue control), so the signal-to-noise ratio is significantly worse. These researchers have developed a pea-sized micro-electrode array that is implanted in the motor cortex and records brain activity, translating it into phonemes (building blocks of speech) via machine learning. The array was tested in four patients with severe speech and motor impairments, enabling a proof of principle for silent decoding.

• Speechy keen - Microsoft AI threw its hat in the GenAI ring by unveiling two new in-house models. The more interesting model is MAI-Voice-1, a high-fidelity, expressive speech generator capable of producing a full minute of audio in under a second on a single GPU. The quality of inflection and tonal variability is impressive, though it’s only available in a handful of voice types. I was disappointed to see that I could not use it to generate a Christopher Walken study buddy, my standard text-to-speech eval. The second model MAI-1-preview, is Microsoft’s first foundation model trained end-to-end with a mixture-of-experts framework.

• Skinny genes - Researchers at Stanford, Princeton, and Berkeley introduced CRISPR-GPT, a multi-agent LLM system that acts as a co-pilot for gene-editing experiments. In addition to the expected functionality of structural design automation, CRISPR-GPT can also engage in Q&A troubleshooting for bio teams. The tailored model integrates chain-of-thought reasoning, retrieval-augmented generation, and specialized fine-tuning on 11 years of CRISPR discussion forums to provide domain expertise. Compared against GPT-4o, it achieved >0.99 accuracy in task planning and significantly outperformed baselines in gRNA design, delivery method selection, and troubleshooting. In wet-lab trials, junior researchers with no CRISPR experience successfully executed multi-gene knockouts and epigenetic activation experiments on the first try, which is widely impressive and demonstrates just how much of a macro lift fine-tuned AI can be in the bio domain.

• Adobye bye - Google made a pretty big update to the image editing capabilities within the Gemini app, powered by a new Flash 2.5 model, codenamed Nano Banana. This new functionality is specifically designed to preserve a consistent likeness of people and pets throughout the editing process, ensuring that the final images feel authentic and natural. I will say, it is pretty fun to play around with. The post also highlights how users can now blend multiple photos to craft entirely new scenes. As expected, the provided examples are quite a bit better than my first-hand experience, but it’s still worth some experimentation. Also noted in the post: all images created or edited using these AI features will be marked with both a visible and an invisible digital watermark, which gets ahead of the forthcoming deepfake doomer rants.

Portfolio Flex

• Orbital released OrbMol, their newest model optimized for modeling metal complexes, biomolecules, and electrolytes.

Greetings from Seattle where I’m visiting family and hopefully surviving our Mt. Reiner hike that’s scheduled in a few days. I have a concerningly busy travel schedule over the next few weeks (Seattle → LA → London → KeyStone, CO → Phoenix) but dealflow is light this time of year so I might just have time for a coffee if you DM me. What’s not light is the amount of news I’ve consumed over the last 2 weeks so let’s get to it.

Stuff Worth Sharing

• Tired of the rack race - A visually stunning deep dive into the single biggest issue of the day: compute infrastructure. For a tech nerd, this is photojournalism at its finest. The Financial Times piece tours the monumental data centers being built across the country and explores just how rapidly the supporting technology has had to evolve. Some of the CPU vs GPU info might be a bit basic, but about halfway through the author gets into the intricacies of GW-scale HVAC and the insane levels of capital now flooding into the sector. The scale of growth is hard to comprehend, but the photo evidence of the physical sprawl does a solid job. Like nearly everyone in tech, I find the surge of compute investment extremely exciting. With billions pouring in month after month, it’s reasonable to wonder if we’re in the middle of an inevitable bubble—something I question somewhat daily now—but at present, there are no signs of slowing.

• Embody mass index - As my readers will already be very familiar, I am very bullish on robotics and their rapidly expanding potential due to emerging semantic understanding. This wonderful piece by Joanna Lichter outlines three major technical bottlenecks that remain today (data scarcity, policy learning, and hardware constraints), with some theories on how we might overcome them. The piece covers the full gamut of startups that are experimenting with solutions like low-cost data collection, world models, and vertically integrated platforms, but the sim-to-real gap is a wily beast and has been the cause of many a company wind-down. Lichter argues the future lies in fusing VLMs’ semantic strengths with RL’s autonomous learning to build general-purpose robotic intelligence, which is a view I strongly resonate with. This broadly summarizes the Fly robotics thesis, and you could easily map our Fund III portfolio to most of the main segments outlined. (Hardware dexterity to be announced…)

• Bind the gap - UK startup Latent Labs has introduced Latent‑X, a frontier AI model enabling push‑button design of de novo protein binders with atom-level precision. The model was able to deliver lab‑validated binders with macrocycles showing a 91 %+ hit rate, and mini‑binders had 10–64 %. The major unlock comes from Latent’s ability to simultaneously generate both sequence and 3D structure, which allows for 10x faster throughput than existing pipelines. Latent is playing the mindshare game by making this model early-accessible to upload targets, design binders, and rank candidates. Engineered binders remain one of the most fascinating areas of techbio today. We work closely with the team at Lir Tx, which is focused on AAV engineering for gene therapies, but obviously every modality has its relative place. Very excited to see stuff like this rapidly coming out of London. Paper here.

• Syndey Sweenie’s got great Genies - Google dropped Genie 3, and it is awesome. This model can generate diverse, interactive environments from a simple prompt and allows for real-time navigation at 24 frames per second. The work builds upon previous models, Genie 1 and Genie 2, and represents a breakthrough in allowing for real-time interaction while maintaining consistency over several minutes. Genie 3 can simulate physical properties and create natural or fantastical worlds on demand. Users can now prompt 'world events' that alter the environment, such as changing the weather or adding a herd of deer. Genie has always been rooted in gameplay, but this latest release has garnered a lot of buzz in the robotics community as a means to seriously scale training data for long-tail deployments. Open-Endedness Team Lead (and friend of Fly) Tim Rocktäschel has been retweeting some great demos, so I highly recommend a visit.

• In God we rust - Dutch startup Ore Energy has launched the world’s first grid-connected iron–air battery at TU Delft. This is a big milestone for long-duration energy storage (LDES) and comes at a time when energy independence is being seriously rethought across Europe. Ore’s approach uses a rust–derust cycle, where a reversible chemical reaction between metallic iron and oxygen stores and releases energy. I never really thought of rust as a two-way street, but apparently it’s exceptionally good at releasing electrons and enables a battery with 100+ hours of storage made entirely from abundant materials. Ore claims the system is 7–10× cheaper than lithium-ion alternatives and can cut system-wide energy costs by up to 63%. The startup aims to scale production to 50 GWh/year by 2030.

• Strategic Open Sourcing - Last week was a big one for OpenAI. The first of two major drops was the introduction of a new family of open-weight language models called gpt-oss. Zuckerberg eat your heart out. These models are available for free on Hugging Face under the Apache 2.0 license and mark a notable shift for the ironically named but to date closed-source company. The gpt-oss-120b model is quite impressive, with capabilities close to OpenAI's o4-mini model, while the smaller gpt-oss-20b is similar in performance to o3-mini and can run on edge devices. Both models utilize a mixture-of-experts architecture and some of their latest training methods. OpenAI is also running a Red Teaming Challenge with a $500,000 prize to encourage the community to find new safety vulnerabilities. I’ve not played with this but the X vibes suggest this is a welcome addition to the OSS ecosystem.

• GPUseless - Not to be outdone by themselves, OpenAI also released GPT-5, a launch whose fanfare was notable only for the middling reaction it received. The most advanced LLM is described as, “the smartest, fastest, and most useful model yet, excelling in domains like math, science, and law”. The two-hour-plus unveiling highlighted a number of new capabilities, with a particular emphasis on safety and accuracy to reduce hallucinations and provide more reliable answers—including for health-related questions. As mentioned, reactions have been mixed. While things looked like a near-immediate win, especially with developer capabilities, the mass market reaction seemed to turn, or at least flatline. To some extent, this is inevitable; humans are fickle after all. With that said, this did feel like a letdown. From my personal vantage point, I was unable to access GPT-5 for four days despite using a VPN to every major city in the country. Once I got access, I did notice an upgrade in speed and clarity, but I’d say it was incremental at best. Core user experience issues were frankly more of a problem for me, with complete memory drops and total freezing being consistent all week. Even now, as I try to proofread this, I’m having to move to Gemini.

• Implanting evidence - Recent deal work has pulled me down a BCI rabbit hole, which led me to this recent coverage on Synchron. We’ve discussed this company in past issues, and it's quickly emerging as a major player alongside Neuralink. Synchron is making (brain?)waves with its minimally invasive implant, the Stentrode, which is inserted via the jugular vein rather than open-skull surgery. The device is currently in trials with patients, decoding neural signals to control devices, send messages, and interact with consumer tech. While less precise than Neuralink’s direct brain implants, Synchron’s approach is safer and much more scalable. The first half of the article focuses on the technology and the relative benefits of semi-invasive methods, which I’m infinitely more bullish and am actively pursuing. There is also quite a bit of ink on the rich history of the company, which startup nerds will enjoy. If you’re more of an audio learner, this is a good interview with the company’s CEO.

• The Talented My. Gripley - A fun DIY tutorial on building and operating your own robotic gripper. This is a notable departure for me, but I've been thinking about getting a 3D printer for my home office. That thought led me to invent reasons to own one, which led to this project. This Universal Manipulation Interface (UMI) guide is a fantastic resource for anyone who wants to dip their toe into at-home robotics. It includes everything you need for a complete, full-stack solution.

Portfolio Flex

Orbital launched a shiny new website, which is just in time to take in all the traffic from being named a Future 50 by the Generalist.

Back at home base coming off a fantastic week in NYC. The weather was as good as the pizza and I can’t wait to go back. With the summer dealflow doldrums now in full effect I’ve been given a more than normal time to process through reading so I hope to get issue #62 out in short succession, but this one is full of the deeptech news, papers and puns that you’ve come to expect. Let’s get to it.

Shameless Plug

• Last week I posted a short piece on the research I’ve been conducting on European university spin-out terms and the response was…strong. At the time of writing this newsletter the post has had over 350,000 views, 1700+ likes and a lot of DMs. Turns out that people really, really respond to ranking lists. I have definitely ruffled some feathers at a few TTOs but the community is overwhelmingly aligned with me that the EU spinout system is broken. If we want Europe to fulfill its innovation potential it starts with treating founders like founders and I intend to keep fighting the good fight so watch this space. Full post here.

Stuff Worth Sharing

• ‘Tis the reason to be jolly - Last week I attended the AI + Bio Summit in NYC, and one of the most prominent topics was the need to rethink clinical trials. It’s widely accepted that clinical development is broken. Perhaps the single most cited datapoint I see across my dealflow is that it costs over $1 billion to bring a drug successfully to market and that number is steadily increasing thanks to Eroom’s Law. Most performance improvement solutions focus on better patient selection, but this paper takes an innovative approach: predicting trial outcomes at the operations level through a highly explainable model structure. AUTOCT merges the reasoning capabilities of LLMs with the interpretability of classical ML to predict trial outcomes by autonomously retrieving public information and refining tabular features for Monte Carlo Tree Search. The paper claims SoTA performance with ROC-AUC scores of 0.753, 0.639, and 0.702 on a test set for Phase I, II, and III respectively. The black-box nature of deep learning has always been a major impediment to clinical adoption, so this offers real promise in a super high-stakes space.

• Sound judgement- Deepfakes and GenAI-powered fraud is a topic we’ve covered a lot in (Re)Share. Audio models specifically have reached uncanny levels of realism, but that capability poses real privacy and security risks. A new technique called teacher guided unlearning (TGU) enables text to speech systems to “forget” how to mimic specific voices while preserving general performance. When prompted to clone a forbidden speaker, TGU produces a random, untraceable voice instead. It reduces speaker similarity for protected voices to the level of unrelated speakers, yet only degrades performance for retained voices by about 2.6%. The jury is still out, but unlearning tools appear to be a far more resilient defense mechanism than audio-based watermarking or off-limits category segmentation. Paper here.

• Overcompensation station - Born-again fratboy Mark Zuckerberg unveiled Meta’s bold new vision for AI superintelligence: a fleet of multi-gigawatt data centers to rival the scale of an entire city. While several sites are in development, the centerpiece is Hyperion, a Louisiana facility that Zuckerberg says will eventually draw 5 GW of power—or “enough to cover a significant part of Manhattan.” The AI gold rush is not a new topic here, but the scope of this commitment is astounding. Meta plans to invest $68 billion in capital expenditures this year alone, dwarfing previous tech infrastructure spending and marking a major escalation in the AI arms race.

• Cookie monster - Google’s AI Overview summaries, which were integrated into Search in late 2024, have radically altered user behavior. A Pew Research Center study tracking 900 users in March claims that once an AI Overview appears, the click-through rate to websites drops nearly 50% and sessions end much earlier. Good for users, catastrophic for the digital marketing ecosystem. Desktop click-through rates for publishers are down about 48% over last year, and we have the lawsuits to prove it. Despite the drop in overall traffic, Google is killing it, with advertising revenue rising 12% in Q2 ($54 billion). Search and publishers have always had a tenuous relationship, but their unholy alliance is falling apart and a business school case on platform risk is being written in real time. It’s for this reason that I’ve been actively exploring creative solutions around Generative Engine Optimization (GEO) and will happily speak to any founder in the space.

• Copyright and wrong - Trump rolled out an expansive America’s AI Action Plan, built around three executive orders aimed at asserting U.S. AI dominance. Unsurprisingly, the policy takes aim at Biden-era AI guardrails and continues the blitz against DEI-focused frameworks or “woke AI.” I can get behind a number of these efforts, but I take real issue with the comments made days later at the Hill & Valley AI Summit, where the President dismissed legislative efforts mandating payment for copyrighted content used in AI training because it stifles innovation. Now, to be clear, I’m biased in this debate because I’m married to a wonderfully talented author - available wherever fine books are sold - but I strongly disagree. The whole reason we have innovation, and capitalism itself, is because of property rights. Take away the incentive and I’ll show you the outcome.

• Mask and you shall receive - Google unveiled the LSM2 model, an innovative Adaptive and Inherited Masking (AIM) approach designed to overcome the inherent data limitations of wearables. “Real-world missingness” refers to the natural occurrence of gaps in bio-signal data, typically due to battery drain, charging, showering, etc. Most wearables rely on data imputation where data gaps are filled after the fact, which can introduce bias or break entirely under real-world failure modes. LSM2 embraces missingness and uses the AIM technique, where learnable mask tokens are injected at both real and synthetic missing points. “Adaptive” masking randomly removes inputs during training to simulate dropouts, similar to dropout in vision models. “Inherited” masking reflects real-world collection patterns. The model learns to interpret these absences in context, not ignore or guess them. It was trained on 40 million hours of multimodal data (Fitbit and Pixel devices) across 60,000+ users and outperformed its predecessor, LSM1, across classification, regression, and generative tasks.

• Good griddance - A great interview with Zach Dell, CEO and co-founder of Base, one of the most interesting startups in energy today. Like most Invest Like the Best episodes, this one is wide-ranging and a bit of a time investment (84 minutes), but it’s an excellent listen for anyone interested in energy, climate, or how to avoid a complete grid collapse under the weight of AI and EVs. Dell gives a clear and insightful overview of the market structure behind energy generation, transmission, and distribution. He then dives into the specifics of Base, which installs distributed home battery systems that serve as both backup power and flexible grid resources. I found the conversation refreshingly specific and economical. So many startups are working on exciting problems but fail to recognize that they’re operating in commodity markets. Engineering excellence is table stakes but so is financial discipline. Cost is what matters and there are no premium electrons.

• Mine the gap - My newest deep dive (pun obviously intended) is the world of mining. MP Materials operates the last major US rare-earth mine at Mountain Pass, CA in an effort to reclaim domestic control over critical minerals. A lot of ink gets spilled on China’s dominance in rare-earth metals, but the truth is North America has plenty of them. What we really lack is the supply chain and processing infrastructure to turn those raw materials into functional components like magnets. The article traces MP’s checkered history of boom and bust, but the company is now on a tear. It has aggressively expanded beyond extraction by building the first US-based facility to produce neodymium‑iron‑boron (NdFeB) magnets at scale. Capacity has jumped from 1,000 to 10,000 metric tons per year, supporting production needs for Apple, the Department of Defense, and probably some other secret things.

• A chip off the Soviet bloc - I recently finished the audiobook version of The Cold War by John Lewis Gaddis and it was awesome. Terrifying to realize just how close we came to complete nuclear annihilation multiple times over the Cold War’s ~40-year history—but as far as reads go, yeah, it was awesome. I chose to read this because of the accelerating similarities to the West’s current-day relationship with China. I’m not deep enough in this very complex time period to offer any insightful takes, but the motivating forces that drove that conflict - hubris, fear, opportunity, unfamiliarity - are all alive in today’s news cycle. If you’re a fan of history, or just someone interested in current geopolitical dynamics, I highly recommend this book. The Cold War was an unquestionable win for the West and capitalism, of which I am extremely thankful and proud, but that path to victory was enormously contingent on dozens of specific decisions and idiosyncratic personas. It would not take a lot for us to be in a very different version of the multiverse, so to all my VC contemporaries who orate on their soapbox about abundance and “winning AI”, take it down a notch.

Portfolio Flex

• Dash Bio was featured in the Wall Street Journal for their rapid fire funding success.

• Transcelestial is now space-qualified and ready to launch with Europe’s 6G StarLab.