Note this is posted in a personal capacity and this blog does not represent any positions of Relation or Lindus. However, it does include reflections made from conversations with talented colleagues and friends, including Henry Taysom, Patrick Collins, Kevin Foote, Adam Cribbs, and Peter Crane.

Many inefficiencies remain in the decade-long journey to bring a drug to market, and there exist many companies fulfilling various niches within the ecosystem. The story that everyone remembers is that, at the cost of billions of dollars per marketed drug, only one in ~20 drugs programmes is successful, but what about other, less known, endeavours? Who are the other companies that support this path to market? How do the billions get spent?

The “TechBio” portmanteau follows from the concept that a biotech can be built in the style of a software company. Given that, where are the opportunities for TechBio companies to improve the value chain? Having interacted with players both big and small employing a diverse range of business models, it is now time for a (draft) review of how this community fits together. As we progress, we also highlight clear gaps in the ecosystem waiting to be filled.

Broadly speaking, TechBio companies can be imperfectly categorised into: drug discovery (both target discovery and biomolecular design), life sciences tools (including automation), drug development groups, and contract research organisations (including manufacturing and clinical trials), and cutting across this all, some nascent but important SaaS presence. This essay is Part I focused on the discovery stage (above diagram), i.e., before late preclinical and before we enter the clinic. At a later time, we will write Part II on drug development, i.e., late preclinical and once you enter the clinic.

Subscribe now

If there were a simple take home from the following essay, it is that it is all about the data generating process, either:

• If you own a capability or the underlying IP, and there’s a great business model to be had if your technology truly unlocks novel biology.1

• If your primary source of data is public, then the key opportunities relate to building and exploiting networks (and moving towards SaaS-style business models).

Finally, we also consider the incentives for cloud labs and explore why they do not necessarily align as a VC-backable proposition.

Drug discovery

Inevitably, for healthcare to work as an industry, someone has to sell a drug. Invariably, this position is served by pharmaceutical companies, who then sell to pharmacies and hospital networks,2 and increasingly even go direct to consumers (e.g., Eli Lilly launched LillyDirect). Some of this revenue ends up reinvested in R&D, but not all of these activities need to happen internally within pharma companies. Multibillion dollar transactions have minted what we refer to today as the drug discovery industry: the systematic creation and outlicence of potential medicines — or more specifically: their associated IP and regulatory portfolios, see graphic below for the changing dynamics of drug origination. Few drug discovery companies ever decide to go all of the way to market themselves, but partner to both lessen capital requirements and acquire a distributor for their eventual product.

To the layperson, you may imagine that one cooks up new molecules in a lab and then starts testing them in mice, and then perhaps move your favourite one into human trials. At some level this is true, but before we start designing molecules, we must have some specifications for what the molecule needs to do; that is, what biological mechanism the would-be drug manipulates — we refer to this as the drug target, which is often a protein-coding gene. Target-free processes do exist, (referred to as phenotypic drug discovery,3) and were employed to great success historically, but with our ever increasing understanding of patient heterogeneity, the clinical need for most diseases is typically discrete and specific (e.g., non-response to first line therapy, greater efficacy vs standard of care, improved safety profiles, that are challenging to model in vitro), requiring one to incorporate disease biology into the target medicine profile (performance criteria for the eventual drug). Only then do we start thinking about molecular structures.

This split is actually mirrored in how drug discovery organisations position themselves. Startups fall into two (not mutually exclusive) categories: biomolecular design companies (“chemistry” companies) and target discovery companies (“biology” companies). These capabilities do not solely exist within start-ups; pharmaceutical companies have such functions themselves, but in order to move into new disease areas, access the latest technology, or to expand their portfolio, partnering is a natural decision.

With that, let’s discuss Target Discovery.

Target discovery

After various retrospective analysis4 (e.g., AstraZeneca’s 5R’s framework), it can be concluded that many assets were brought into the clinic with data gaps, strategic missteps (or indeed, poor governance), and logical leaps of faith that could have been flagged earlier, resulting in clinical trial failure. Below, historical data show that the primary cause of these losses was due to lack of efficacy and safety signals,5 i.e., our understanding of disease biology was inadequate.6

There’s also been a few shocking case studies demonstrating that for many drugs in clinical trials (mostly limited to small molecules in oncology), our understanding of their mechanism was actually completely wrong and the preclinical data related to a different mode of action, see Science Translational Medicine article below.

These studies (along with many others) have led to the development of heuristics and standards for the target discovery process. Target discovery7 can be approached with various philosophies in mind, from using patient genetics (and other ‘omic technologies), to building disease models that can be perturbed using a range of tool molecules (ideally with well-characterised, known function) and genetic instruments (typically variations of the CRISPR technology). Ultimately, for any line of evidence, historical benchmarking studies can both support and refute the utility of any single technology, meaning that one has to synthesize evidence and “make sense of it all”.

Note, in a future post, Relation will release its “how to” philosophy for target discovery — subscribe to hear it first on the official Relation blog.

What’s been tried?

Target discovery companies broadly fall into 3 buckets:

Novel ‘omic technologies: Imagine you have invented some new way to measure cell behaviour or aspects of human physiology, such assays are what we refer to as ‘omics.8 For most ‘omic technologies, the rights holder does not have the freedom to operate, as their claim really sits under a broader patent family held by another party. With this in mind, instead of selling kits (sets of reagents), some founders decide to use the technology to discover drug targets and outlicence these findings to other parties.

Generally, these approaches have limited success: it’s hard to establish the value of one’s technology without selling it at scale, therefore a frequent solution is to move into drug development as quickly as possible so the company can become valued on their prospective medicines — not their unproved method.9 For examples of this, consider Enhanced Genomics and Nucleome Therapeutics, both own patents relating to the Hi-C protocol (or variations thereof) and operate a target discovery business model.

Therapy area expertise: Another option is to be more agnostic to the employed technology but build expertise in a collection of disease areas joined by some common theme (e.g., cell types, or pathway). Relation achieves this through the systematic (tacit) knowledge in collecting and characterising tissue associated with musculoskeletal disorders, immunology, fibrosis, and dermatology. If you are lucky (with the right clinical protocol), certain economies of scale emerge: Relation so frequently accesses resected tissue from a range of orthopedic surgeries that the effective costs have been drastically driven down (less experimental failure, scaled workflows, faster analyses). However, this comes with the cost of agility — suddenly initiating a new study to access a new disease area requires planning and creates friction. Other TechBio companies with this strategy include: Immunai (immunology), insitro (metabolic, neurology), Cellarity (hematology), Celsius Therapeutics (oncology, IBD; now part of AbbVie).

Amalgamation plays, or “the everything company”: This last bucket is somewhat the “other” category, groups like BenevolentAI who, whilst tied with Recursion as one of the first TechBio companies (and deserve to be credited as such), unfortunately developed technology so broad it is hard to know exactly when it can and can’t be employed. When a company cannot elucidate a narrative for what they’re good at compared to where they can only provide limited value, it is reasonable to conclude a lack of focus.

What could work?

The above states a clear preference for therapy area expertise, but there is value beyond just target discovery in such a strategy: can you reuse the data to find biomarkers to diagnose disease, find patient subpopulations, or even predict who will respond to treatment? In short, the data moat can become multipurpose. Case in point, Relation’s capability in orthopedics can clearly feed discovery efforts in at least 3 complex diseases (so far),10 but has also been designed in a manner to identify potential candidate biomarkers and the accompanying stratification logic.

Two big trends that seem to capture the VC imagination at the moment are the “Virtual Cell” and “AI scientist” concepts. Both areas are still nascent with many claims of victory that more often than not collapse when placed under scrutiny; we seldom have sufficient data (or indeed, the right data) to really bring either of these visions to life in their full generality.

In the case of Virtual Cell companies, there are early signs of life. Consider Noetik who build tissue foundation models for oncology that have clear translational logic. They can virtually perturb cells in silico and appear to recover reasonable gene expression profiles that suggests a certain self-consistency of the architecture, whilst also mapping between different imaging modalities of varying expense (H&E vs multi-omic layers); a technology both for target discovery and clinical practice. The successful companies in the space will be those able to learn the lessons above: choose specific disease areas, learn how to address real clinical shortcomings, and work out how to incorporate this into the underlying foundation model.

(Very recent news story: Relation and GSK have recently announced a partnership in this area!)

The universe of AI scientist companies is less convincing. In all likelihood “we can automate (broadly defined) scientific discovery” will not be addressable in a start-up context but will be partnerships between groups like OpenAI or Anthropic working with automation providers — see automation section.

Biomolecular design

Put simply, biomolecular design is the process by which we determine what specific molecule to progress by balancing a series of complex tradeoffs: from activity (on target vs off-target), immunogenicity (for biologics), to absorption, distribution, metabolism, and excretion (ADME) properties, and even commercial aspects including competition and freedom to operate (FTO) concerns and patent strategies. Various “AlphaFold-for-X” companies have addressed different modalities, including antibodies (BigHat Biosciences, Chai Discovery, Nabla Bio), cell therapy (coding.bio), capsid design (Dyno Therapeutics), of course small molecules (Exscientia, now part of Recursion, Insilico Medicine, Charm Tx, Genesis Molecular AI), and modality agnostic groups (Isomorphic Labs).

Given extended development timelines, as of today in 2026, there are no FDA-approved AI-discovered drugs. Moreover, any proposed molecule is often met with skepticism debating as to whether the design was truly novel or mimicking a known skeleton, or if there is any difference in associated clinical success. The long term view feels the most sensible: these tools appear to be reducing timelines and would we still expect to use the same methodology from the 1980’s decades from now?

What’s been tried?

In the beginning, the first round of biomolecular design companies were fast to raise capital, but perhaps a little slower to realise deals. Often overlooked, one of the biggest challenges in drug discovery is the extended timelines to close big pharma partnerships. As predicted, the world of biomolecular design is gradually becoming commoditised, or more specifically, a trend towards (admittedly expensive) fee-for-service models (see Chai Discovery – Eli Lilly deal) and open source repos (and Boltz associated PBC offering enterprise solutions). There’s a reason why this works: faster to execute, often without stipulations on how the product is used (for example, exclusivities on indications that frequently occur in pharma-biotech dealmaking).

What could work?

The challenge for biomolecular design companies is how to maximally exploit their technology. As more startups get formed, we will continue to see endless technology benchmarking studies as marketing material. For all but the most successful companies, there is a downward price pressure and we can expect the fee-for-service deal trend to continue.

However, these comparisons will never tell the full story, invariably each company may have some advantage in a unique modality or pathway. To avoid the conflict of interest (blanket selling a platform access vs. exploiting its unique advantages), single-asset spin outs will allow specialist executive teams to bring the eventual drug to market. This is the boring expansion opportunity.

In all likelihood, the big business opportunity is upselling and even roll ups. Amazon doesn’t just recommend 3rd party products, it actively looks to offer its own “basic” version, whilst keeping you hooked via a Prime subscription. What does this look like for biomolecular design? An AI company designs you an asset with certain characteristics — what else? Can they propose specific tests to derisk what they’ve created? For example, a specific toxicology study to understand an unusual side chain, a target engagement biomarker for the Phase I study, or a freedom-to-operate (FTO) search that identifies a worrisome patent. Recurring revenue could be secured through programme concierge services, e.g., intellectual property (IP) lifecycle management, and softly advertising acquisition to big pharma. Essentially, if the process of biomolecular design truly has been derisked, the next frontier for an AI-native company is handling the back office operations to bring a product to market.

Life science tools

Because clinical development is so expensive, the preclinical evidence packages are becoming ever more extensive: perhaps some new technology or workflow can further derisk a drug before entering the clinic? From experimental reagents, specialist genomic assays, to giant automated lab infrastructure, drug discovery companies rely heavily on life science tooling — but who to choose?

The core challenge for any tool is to get initial traction. Most of the time a new technology is caught between a rock and a hard place: if the solution has direct competitors, it must be price competitive because the customer will seldom want to change workflows or introduce batch effects into their data; the alternative is that it is a fundamental innovation, and therefore the customer will need to displace budget from elsewhere in their organization — and this displaced budget must be significant enough to keep the startup alive. In short, there is a huge activation energy required for new life science tools companies to become established.

Those companies that have established themselves are dominated by publicly listed multinationals, e.g., ThermoFisher, Becton Dickinson, Danaher. These companies are no longer known for their original products — mostly lab equipment — and sometimes date their founding back to pre-1900’s. In actuality, these conglomerates are multi-decade roll ups of many small companies with the dream that there would be synergies realized within their portfolio, see the graphic below taken from The Republic of Science blog post.

If the business model works as advertised, the concept is simple: identify best-in-class solutions, acquire upstart startups before their price becomes too dear, integrate their products into existing sales channels, and then save on operations via centralized resources.11 What distinguishes the good from the great, is that strong executives go beyond “our portfolio needs a proteomics solution to fill gap X” but ask why a very specific solution will achieve market dominance in years to come, or if there is an option to leverage capabilities within their portfolio. For an example of this, consider ThermoFisher’s acquisition of O-link proteomics: a highly sensitive dual-antibody solution that can leverage Thermo’s biologic production capabilities, but also it has a guaranteed customer base for years to come due to its flagship project with the UK Biobank (meaning it is the reference proteomics dataset for nearly all population genomics going forward). On the other hand, when Standard Biotools merged with SomaLogic (then valued ~$444M; Jan ‘24), it was hard to pluck out the logic from behind the deal; after 2 years SomaLogic was sold to Illumina for up to ~$425M (Jan ‘26).

Whilst one cannot hope to cover the breadth of life sciences tools companies, two areas worth doing a deep dive on are ‘omic technologies and automation.

‘Omic technologies.

If there were one unifying feature of all ‘omic technology companies, it would be the sheer amount of effort they expend building and defending their IP portfolio. As covered in the scTrends market review, much of the history of single-cell and spatial ‘omic technologies is characterised by almost constant litigation. In fact, very few individuals have working knowledge of who holds which patent, what claims are covered, and where there are gaps to exploit.12

A good example is the contested space around single-cell RNA-seq and chromatin-accessibility profiling. 10x Genomics sells Chromium Epi Multiome as an off-the-shelf assay for measuring gene expression and open chromatin from the same nucleus, directly linking RNA and ATAC profiles cell by cell. Parse Biosciences’ split-pool barcoding approach can, in principle, support the same broad modalities — RNA-seq, chromatin accessibility, and potentially paired RNA-plus-accessibility readouts — but its commercial path is shaped by a distinct platform architecture and a different set of patent constraints. This is where the IP picture becomes murky: a company may avoid one product claim while still running into a broader patent family covering a class of chemistries, barcoding strategies, enzymatic steps, or compositions of matter. In practice, the question is rarely “can you measure RNA and ATAC?” but rather by what particular molecular route — and under whose claims?

Strange and wonderful things are happening in the NGS-based spatial transcriptomics space: we are seeing multiple companies converge on very similar technologies but by distinct technical architectures and IP positions. 10x Genomics’ Visium, MGI/STOmics’ Stereo-seq, and Illumina’s emerging spatial technology all aim to produce spatially resolved transcriptome maps, yet they differ in how spatial barcodes are generated, immobilised, read out, and linked back to tissue coordinates. At the research end of the spectrum, methods such as Seq-Scope show that Illumina flow-cell chemistry can even be repurposed to create ultra-dense spatial barcode arrays. So the platforms may look similar at the level of data output (post bioinformatics), but they are not necessarily the same assay, nor necessarily covered by the same patent claims.

What’s been tried?

So, imagine you’re a new ‘omic technologies company, what can you commercialise? Without training as a patent lawyer, probably very little (unless you have an entirely new category of innovation).

In the space of single-cell ‘omics, there’s been remarkably little inroads made by startups vs 10x Genomics. When you run a multi-year study with 100’s of patients, you want consistency with an implicit promise that your supplier will be selling the same thing at the same quality for years. One may switch to a new product line for a new study, but these moments are hard to come by from a sales perspective. One notable break in the 10x sticky moat is Parse Biosciences (now part of QIAGEN), a great David vs Goliath story of some PhD students arguably beating the more “official” spin out from their former lab (ScaleBio, now purchased by 10x Genomics) and fending off a lawsuit(s) from 10x Genomics. A true tale of grit and perseverance to become the 2nd in the market.

Within spatial ‘omics and proteomics, the field is arguably more open: many of the foundational patents have long since expired (e.g., FISH, mass spec), but the newer generations of companies are often cautious to launch products until they’ve lined their ducks in a row with regards to FTO.

What could work?

As platforms take so long to reach maturity with questionable IP portfolios, many groups are staying in stealth for longer periods to engage with a customer base and perfect the product. There is a bit of a race between VCs funding the dream before a bigger player is needed — often to handle the legal battle about to ensue. Two-way licensing deals do occur, but usually this is the result of a dispute, not before it occurs.

The big trend to watch here is the tools vendors are vertically integrating into “selling data and disease insights” to pharma to drive foundation model work and internal target ID. Illumina have been doing this via their Billion Cell Atlas initiative (leveraging their recent acquisition of Fluent Bioscience’s PIP-seq solution) and Parse has been operating a similar model for their larger datasets. It’s likely that in the future many companies in the space such as 10X genomics, PacBio, Illumina, Thermo will have a (pre-)clinical insights division selling these outputs to pharma. The big question that many have raised over these deals is how does this sit with their primary model of all selling razors and razorblades?

A more interesting question long term is: was the 10x/Illumina dual solution a one off? Or are there patent exploits littered across the life science tools ecosystem? The answer is, undoubtedly, yes. From Our World in Data, below is a chart of the rate of pharmaceutical and life science patent filings across the globe.

The number of patent filings is inordinately huge, even though many are ultimately abandoned, especially where universities cannot justify expensive fees for non-revenue-generating assets. Others are later challenged at the Patent Trial and Appeal Board, where in 2024 roughly 78% of claims reaching a final written decision were found unpatentable; even so, there are still far more funded patents than actual new products on the market.

One area ripe for AI disruption is mass-scale patent busting. Historically, this has been the preserve of pharmaceutical companies, generic manufacturers, access-to-medicines groups, and utilising specialist patent lawyers: teams map patent estates, run freedom-to-operate analyses, search for prior art, and challenge blocking claims through litigation or patent-office proceedings. A useful example is Gilead’s sofosbuvir, sold as Sovaldi, a hepatitis C drug launched in the United States at about $84,000 per treatment course. In India, access groups and generic manufacturers challenged one of Gilead’s patent applications, arguing that it covered insufficiently inventive chemistry. The Indian Patent Controller initially refused the application, temporarily denying Gilead that monopoly right and strengthening the case for cheaper Indian-made versions. Gilead later won some Indian patent protection, so this was not a clean “patent busted, drug goes generic everywhere” victory. But it did help open market access: sofosbuvir became available from Indian manufacturers through a mixture of patent challenges, voluntary licences, local regulation, and country-by-country patent rules. The typical lesson is that patent busting rarely collapses an entire IP estate; more often, it removes or weakens one obstacle, creating leverage for cheaper products to enter specific markets. Could this work differently for ‘omics product development?

In pharmaceuticals, patent busting can create room for cheaper manufacture of an existing molecule; in ‘omics, it could enable entirely new assay architectures. Agents could decompose platform patents into their functional primitives — sample preparation, capture chemistry, barcoding strategy, enzyme choice, surface chemistry, sequencing readout, imaging readout, and computational registration — then search across papers, old protocols, conference posters, abandoned patents, instrument manuals, and archived product documentation for prior art or design-around routes. The output would not be an automated legal opinion, but a set of ranked claim charts and “white-space maps” for scientists and patent counsel: here are the claims that look weak, here are the ones that are probably blocking, and here are the molecular routes that appear underexplored. For a new ‘omic technologies company, that changes the question from “can we sell our new product” to “can we reach the same biological output through a sufficiently different technical path without walking through someone else’s claims?” In other words, patent busting as part of a product-discovery engine.

Automation

Automation is quite a different beast. Whilst large-scale automation projects are exceptionally expensive (~10’s of millions of dollars), they are mostly running on expired patents. Attending last year’s SLAS, many vendors are offering subtly different, but mostly very similar solutions — so much so that many people in the field cannot distinguish the relative merits between rigs. The high upfront costs and product similarity motivates “integrator” companies who constantly survey the market, find the best solutions, add them to their portfolio of supported devices, and put your lab together for you offsite, and then ship the whole installation to your facility ‘ready to go’ as a turnkey solution. Even with these integrators, the key operational challenges still relate to interoperability (and associated vendor lock in), coordination, and crucially — adaptability. Large integrated automation systems are typically built for a specific scientific endeavor, for example a high throughput screen. Adapting that setup as scientific and business needs change is challenging, slow, and costly.

For examples of interoperability challenges, consider how each vendor may use the same geometry for a 96-well plate (i.e., the positioning of the wells), but one solution may use a barcode labelling with a specific rim for the plate, another uses a numbering system combined with magnets, but then you want to dispense drugs via acoustic liquid handling and that needs a whole different solution. One liquid handler will use plastic tips sourced in the USA, but another uses a very slightly different plastic composition from Asia, and so on, and so on — a business model that is reminiscent of paper printers and printer cartridges.

Beyond straight automation, this really complicates bringing together multiple technologies. For example, company X offering some ‘omic technology (above) will make very specific claims on their performance when using liquid handler Y. Oh no! You have already bought liquid handler Z, then company X will say “you’re on your own, we’ve never tried this”. Now this is true (it’s not been tested in exactly this way), but it’s likely not an issue, but if you’re a pharmaceutical company and you’re investing in an automation build for some ‘omics technology, how much risk do you want to take on? You would prefer to only run as few QC pipelines internally as strictly necessary — remember, within big pharma, every project has to be documented properly with the correct SOP paperwork (often to maintain institutional knowledge and regulatory requirements).

On the dry-lab side, if you want to connect a device that an integrator company has never seen before, you will have to pay a premium — a great revenue stream for the integrator as they are then able to add said device to their portfolio of supported devices. Attaching a device and communicating with it is just the start. At a lower level, plates move through the system via scheduling software, which often only keeps track of where plates are going. Knowing what is in your plate, at the well-level for reagents, samples and volumes, is up to you! Connecting to your LIMS, parsing logs, updating inventory, and linking the experimental design through to experimental results still requires a load of custom patchwork. And then everything breaks if you want to perform some different science — bringing us back to the adaptability problem. The dream is an automation platform with a software and hardware stack that can adapt to changing business and scientific needs.

What’s been tried?

Taking inspiration from Owl Posting’s automation heuristics piece, various strategies have tried (interpretation ours):

The translation layer: By turning protocols into machine-actionable instructions, we can reduce friction between scientific desires and data generation (e.g., Synthace, Briefly Bio, Tetsuwan Scientific).

The hardware layer: By building intermediary proprietary hardware that connects 3rd party devices, we can achieve greater integration between established solutions (e.g., Automata, Ginkgo Bioworks).

The intelligence layer: By improving a system’s perception of its own state (a world model?), we can increase capacity and reduce failure across automated build outs (e.g., Medra, Zeon Systems).

Two groups who were conspicuously absent from the ecosystem worldmap include the key “integrators” HighRes Biosolutions and Biosero, who do all three layers above. A family-owned business and now operating for over 20 years, HighRes have performed over 500 automation installations and partnered with all of the top 20 pharma companies, meaning that whilst startups may have the “mind share”, the people actually delivering appear as unknown entities. However, they’re family owned and most of their big pharma relationships have been long established — they don’t need to advertise. Having seen early demonstrations for how AI is transforming their business (specifically: upload simple protocol into an LLM, get automated implementation with some inventory tracking), much of the automation world feels like it will soon become a solved problem.

What could work?

Owl Posting’s essay then ends predicting cloud labs as the ultimate destination for “where the field is going”; agreed, but is it a venture backable startup problem with huge upside?

On the large-scale fee-for-service build out, everyone wants “Anduril-for-automation” (i.e., ground up, new hardware) whilst wildly gesturing at Flagship’s Lila healthy fundraise (~$200M) to predict this will be the new model.

This triggered some thinking.

Aside: the challenge of the fee-for-service cell biology cloud lab

The case for cloud labs is a nuanced one, depending on what data you want to generate. Historically, the early automation use cases were all related to chemistry, screening small molecule libraries, a very low margin business that was difficult to scale. But the recent focus is all biology (advanced imaging and ‘omic readouts) to power the training of large foundation models. Whilst some applications have been extraordinarily effective (optimization of growth conditions for biomanufacturing), the use cases that move the needle for drug discovery groups are incredibly artisanal: complex model systems that are highly heterogeneous (e.g., organoids, organ-on-a-chip).

Could this be built, a cloud lab delivering such use cases would be hugely enabling for early stage TechBio companies: log onto a website, upload a protocol for interpretation by an LLM, pay for reagents, and receive an email when your data is ready for download. It would allow companies to rapidly build proprietary models and generate target IP, and thus speedrun to the next round of venture financing.

For the avoidance of doubt, the following relates to the challenges in operating a model that is: 1.) fee-for-service, and 2.) focused on cell biology. Essentially, there is a tension: current installations are designed to meet specific use cases, but with multiple customers, one requires adaptability that is hard to deliver on a fee-for-service basis. Therefore, on one hand, someone needs to fund both a build out and product R&D, potentially requiring decades (organoids take months to grow!), and on the other hand, the platform needs to resist attempts to become captured by special interest groups, aka remaining “apolitical”.

The hurdles:

Short-term price vs. long-term value: The alternative to automation is human labour. Whilst automation-generated data can be assumed to be of higher quality with greater quality control. If a manual CRO will charge reagents plus salary against some multiplier, there is some upper limit for how much one can charge for automated science.

However with regards to long-term value, suppose the artisanal craft of “how to do science” is learnable at some abstract level encodable into a foundation model;13 then who gets to retain that information and how? Imagine you let the customer directly utilise your proprietary automated-science foundation model: it will be expensive to operate and the benefits will be diffuse and not immediately tangible.14 In contrast, if the cloud lab keeps all of the experimental learnings to commercially exploit them later, there is an instant conflict with their customers. How does the customer know you won’t compete with them?15 This cannot be accounted for on a fee-for-service basis — remember, partnerships with FTO restrictions will be slower to negotiate and will limit future customer acquisition. So the cloud lab needs to not “have a horse in the race” so to say – but this is expensive.

The disincentive to externalise: Losing a data generation capability has a number of knock on effects: one loses intuition, which then means missing out on opportunities for more interesting information-rich experimental designs — something the AI scientist narrative has yet to deliver, but may do in years to come.

Moreover, many CROs have preferred customers: pharma companies are more important sources of revenue when compared to startups, and whilst not explicitly verbalised, capacity is often directed towards the customer who is guaranteed to still be doing business a decade from now. In the case of Lila, hypothetically Flagship Pioneering (their multibillion dollar venture creator) could feed all of their portfolio’s automation work through Lila.

Put bluntly, are you a real company or an outsourced ML team ripe for an acqua-hire? Perhaps they will turn the data tap off if you disagree to their terms!

Solution space:

To subvert Ronald Reagan, maybe government is the solution. Accepting that we will need to perform product development and we want to “build in the open” so customers can benefit from its learnings, the next-gen cloud lab should be capitalised via a governmental organisation or an “apolitical group”, e.g., NVIDIA, Microsoft, Amazon etc.16 No one worries that AWS is serving your competition, but startup executives should rightfully be concerned if key platform data generation capability all comes from a specific VC-backed cloud lab — especially if that group is much better capitalised than you (this is actually bizarrely similar to what happened in the early days of Ginkgo Bioworks where they invested in their customers and were subsequently the target of shortsellors).

The other question is on what needs to be built: bluntly, it is not clear where a large unfilled technical gap is beyond specific use cases, e.g., some complex coculture capability. There’s various small hardware and scheduling optimizations that can be made, but there’s providers (like HighRes) who can already integrate lots of diverse systems. Moreover, with so many coding agents, as with biomolecular design companies, the open source universe will eventually fill this niche and the margins will collapse. Anything relating to automating protocols will be the remit of OpenAI, Anthropic, DeepMind etc, and almost all solutions will be a wrapper around their agents. What does need to be built is the open source infrastructure to rapidly redeploy labs for new use cases. With this, one may have a foundation for actual cloud labs that could scale as cloud compute does.

Without being specific on actual schematics, two thoughts worth contemplating on: modularity and franchise. The development within automation that’s most exciting is the transition from rails (where there is a fixed route through the system) to autonomous mobile robots (AMRs) to move between static systems, which can then be upgraded and switched out depending on demand.

You only need to get this working once — and we know this is technically possible. Thereafter, you then have a system that can be deployed across multiple jurisdictions around the globe (Boston, West Coast, London, Singapore etc) via a franchise model. Data robustness is trivial in a sense: does Boston’s data agree with London’s data for the same experiment? This replication then can be used to estimate your confidence intervals. Capacity can also be answered by offshoring experimental work.

Unfortunately, this is hugely capital intensive and for the reasons above, this cloud lab network is not easily a VC investable model with a 10✕ return, but perhaps a 2-5✕ return suitable for private-equity – but with a non-trivial risk profile of a de novo startup building an operational capability. However, from an ecosystem perspective such a proposal is hugely appealing — especially capital constrained TechBio groups outside Boston and San Francisco — we unlock a community-led data generation capability that will allow fledgling startups access the next wave of capital.

That’s it for Part I, but hopefully that gives a feel for what’s going on during the discovery phase.

We realise we’ve not really touched on the impact of China. Approximately a quarter of drug candidates under active development originate from China with 46% of new drug molecules entering clinical trials in 25H1 from Chinese companies. Unfortunately, this is going to get pushed to Part II. Subscribe to hear more!

Subscribe now

Human longevity is substantially heritable — recent estimates put intrinsic heritability near 50% once extrinsic deaths (accidents, infection, violence) are accounted for — yet genome-wide association studies (GWAS) explain very little of it, aka “missing biology”. One of the largest under-exploited clues sits in plain sight: sex, but despite various sociological explanations (violence, risk-taking behaviour) and broad-brush biological explanations (X-linked genetic disorders, immune function, infection risk), we do not yet have a molecular (consensus) understanding for why women live longer than men. This effect is observable at all stages of life, even in utero.

From a gene regulatory perspective, we expect the variation to be driven by development, immune function, endocrine signalling, and genetics — confounded by chromosomal structural differences. However, human longevity GWAS have only produced a small number of sensible signals relative to the size and importance of the phenotype. Several of the few replicated signals even act sex-specifically (APOE’s effect is stronger in women, and equivalently for FOXO3 in men), so pooling the sexes can blur real effects. I would argue that one of the main reasons for this “missing biology” is that we treat sex as a single covariate to be adjusted away, rather than modelling the two biological axes it bundles: developmental effects and sex-chromosome gene dosage effects. And, given that drugs often mimic the effects of genetic variation (with different effect sizes, acting across different timescales), if we truly want longevity drugs, then understanding the genetic basis for differences in life expectancy is a natural starting point.

For an analogy, consider how we use the body mass index (BMI). BMI is defined as a person’s weight divided by the square of their height (kg/m²), which gives a single number that combines two distinct measurements. Because it forces a fixed relationship between height and weight, BMI can obscure associations that height or weight would each reveal independently. In short, two people with very different builds can share an identical BMI. Sex shares some similarities to the BMI example above.

Allosomal aneuploid populations as counterfactuals

Sex is an awkward categorical variable to incorporate into statistical analysis. Neither “male” nor “female” are single molecular exposures, each bundle together (at least) two major biological axes. First, sex determination in mammals is linked to the Y-chromosome gene, SRY, that initiates testis determination and downstream male sexual development, creating lifelong differences in anatomy, hormone profiles, immune function, metabolism, etc. The second is sex-chromosome dosage: XX and XY cells differ even after accounting for developmental differences. Although one X chromosome is largely inactivated in XX cells, a meaningful fraction of X-linked genes escape X inactivation. Reviews estimate that roughly ~15–30% of human X-linked genes may escape inactivation. These escape genes exhibit dosage-sensitive biology that could plausibly contribute to sex differences in ageing.

When we collapse these axes of variation, we then become unable to ask very fundamental counterfactual questions: what is the phenotype of a man with more expression from X-linked escape genes? What is the phenotype of a woman with one fewer X chromosome?1

However, human counterfactuals already exist in the form of Klinefelter and Turner syndrome. Klinefelter syndrome (KS), or “47,XXY”, are males with an extra X chromosome. Turner syndrome (TS), or “45,X”, are females with partial or complete loss of a second sex chromosome. Other rarer sex aneuploidies like 47,XXX and 47,XYY also exist, see below for life expectancy estimates (as known). Whilst these patients often suffer from chronic disease (e.g., type 2 diabetes and metabolic syndrome, cardiovascular disease, osteoporosis, and autoimmune thyroid disease) and infertility, they are mostly healthy but with a reduced life expectancy. They are also not vanishingly rare. Klinefelter syndrome affects about 1 in 600-660 male births and is substantially underdiagnosed; many people are only identified during infertility assessments. Turner syndrome is reported in roughly 1 in 2,000-2,500 live female births, with wide phenotypic variability, especially in mosaic forms. Compared to other “age-associated” rare diseases (progerias, Werner syndrome, Cockayne syndrome etc) which implicate only a handful of genes (mostly relating to DNA repair and genome integrity), these aneuploidies of the sex chromosomes have much broader applicability.2 Their effects are not uniformly harmful either: Klinefelter syndrome is associated with reduced prostate-cancer risk and Turner syndrome with reduced breast-cancer risk, pointing to protective dosage effects for specific age-related diseases.

Complex emergent trans-regulation

Molecular characterisation of Klinefelter and Turner syndromes remains sparse relative to common-disease biobanks. Reanalysing a 2020 study that measured PBMC (white blood cell) bulk RNA expression of a KS and TS patients with age-matched 46,XX and 46,XY euploid controls, we see a clear dose-response relationship for genes on the X and Y chromosomes in relation to copy number. In effect, these aneuploidies act as natural dosage dials; by tuning copy number up and down rather than switching genes off, much as a drug modulates rather than abolishes its target.3

However, much more complex relationships emerge when examining the autosomes, hinting at significant perturbations to human physiology. When examining known genes implicated in longevity, we observe dysregulated gene expression. For example, similar to some autoimmune conditions, Turner syndrome women have higher LPL expression than 46,XX women (suggesting one possible plausible reason for lower life expectancy). Several newer molecular studies now support this broader picture of autosomal gene dysregulation, including transcriptomic, epigenomic, single-cell, metabolomic studies. However, these datasets remain small, tissue- or development-stage-specific, and largely cross-sectional or disease-mechanism-focused, rather than longitudinal, population-scale resources designed to connect sex-chromosome dosage to ageing and survival.

How biobanks help establish causality and reversibility

To discover drug targets likely to treat complex disease, we typically look for characteristics that suggest the drug target is not entirely dissimilar from the targets already manipulated by successful drugs, e.g., expressed in the relevant cell type, druggable from a structural perspective, etc. Most notably, the retrospective study by Nelson et al, showed that when you consider drug targets associated with clinical success, they are ~2.6 times more likely to have some genetic basis for a disease-associated phenotype related to them, often via GWAS. And because germline genetics are broadly assumed to be immutable, then they are causally upstream and are driving disease risk.4 However, causality is not the same thing as reversibility, so we need to build model systems (either in vitro or in vivo) that mimic key parts of pathophysiology, and show that a healthy phenotype can be restored. Biobanks help with both problems because whilst we typically study germline genetics from a blood draw, we may also retain tissue samples that can be used to either study cellular architecture (static) or build dynamic cellular models of disease.

The biobank evidence we already have shows the gap. In the UK Biobank, for KS researchers identified 213 men with 47,XXY and 143 with 47,XYY among 207,067 men of European ancestry, and of the 244,848 women over 40, they found 30 women with 45,X, 186 with mosaic 45,X/46,XX, and 110 with 47,XXX. Because biobanks detect aneuploidies by screening everyone’s genome rather than relying on clinical referral, they capture undiagnosed and mild cases that clinic-based cohorts miss.5 However, these numbers are still shockingly small to be meaningful from the perspective of a population genetics statistical analysis.

So, what do we need to do? We need to build comprehensive data resources that link genetic variation, karyotype status, single-cell or spatial omic profiling of important tissues (e.g., white blood cells to characterise the immune system etc) across thousands of individuals with KS and TS. This can be done by extending the UK Biobank or through entirely new initiatives. Taking a step further than what the UK Biobank already does, we should also collect skin fibroblasts for development of iPSC cell banks to build models of relevant disease processes, and track individuals over time to understand organs and tissues as they accumulate damage.

In closing

Much of modern longevity genetics still leans on finding putative ageing genes across species (not within humans). Human studies are possible (consider centenarian, familial studies beyond standard GWAS) but have yielded few robust loci relative to lifespan’s heritability. We know from drug discovery writ large how challenging preclinical translation can be (i.e., are you actually testing the same biology across experimental systems?). Whilst there are highly conserved biological processes across mammals (e.g., bone formation, cartilage destruction, muscle contraction etc), many of the processes that we are most interested in have diverged over evolutionary timescales, e.g., immune function. For example, humans, wild mice, and standard laboratory mice have very different pathogen ecologies (sterility of environments, exposure to airborne pathogens vs. bacteria/fungal colonies), and laboratory husbandry can strongly shape phenotype. Moreover, the lab animals we experiment on often are bred in artificial conditions skewing their phenotypes in unexpected ways: telomere length varies substantially across mouse species and strains; some established laboratory strains have unusually long telomeres, while wild-derived strains can be shorter or longer depending on background.

Human longevity genetics has spent years looking for variants that explain why some people live exceptionally long lives. That remains important. But perhaps one of the largest natural experiments in human survival has been sitting in front of us the whole time: women live longer, girls survive better even in infancy, and sex chromosomes should not be just binary labels in a covariate table. If human genetics is one of the best maps we have for drug discovery, then allosomal aneuploidy biobanks could help draw a missing section of the longevity map.

This blog post has been adapted from a perspective in Trends in Genetics.

2

One interesting study examined the structure of protein-protein interactions, grouped by the genomic location of the underlying gene by chromosome. Here, they found that chromosomes 4, 13, 21, X, and Y are “less connected” than the rest of the chromosomes, suggesting one possible reason some whole-chromosome aneuploidies are more compatible with survival than others

4

Genetics can also be used to not just understand causal effects of genetic variants or DNA structure, but also intermede dynamic phenotypes via Mendelian randomisation. MR uses inherited genetic variation as a kind of natural experiment to infer whether an exposure causally affects an outcome. Within drug discovery, it can tell us how a biomolecule of interest (e.g., a blood-based protein) drives disease pathology whilst “regressing out” genetic differences.

It’s Christmas and I’m in Okinawa, and to reflect this change of pace I felt like writing a little outside my usual TechBio swim lane. One thing that’s always left me perplexed after visiting Naha for a few years now (and where I am now): why does it have so little name recognition on the international stage?

Purely from a geography perspective, Naha is almost dead centre of a ~2 billion person East Asian market and boasts international flights to 5 major cities within ~3 hours: Hong Kong, Taipei, Shanghai, Seoul, and Tokyo. In 2020, Naha Airport opened a second runway, and offerings now include cities further afield (Singapore, Kuala Lumpur etc) with capacity jumping from 135,000 to 240,000 flights per year.

From a quality of life perspective, there are ~160 Okinawan islands in the prefecture and many of them have (admittedly infrequent) public transport via boat. Every trip I make here is an adventure to see a mini universe contained within a few square miles. Each island will have some rustic cottage industry, from fresh mangos to the most exquisite wagyu beef you’ll ever eat.

In short, amazing location and unique tourism, but I suppose this is ‘necessary but not sufficient’ to become an economic hub. However, I don’t think the barriers for Okinawa are in fact that great in the scheme of things. In fact, it’s now become my 20-year prediction that Naha will become an up and coming regional, or even international, city.

An (under appreciated) ingredient of major economies relates to how easy is it to invest. From a property perspective, when compared to the surrounding regions Japan is comparatively open on to foreign real estate ownership (with certain reporting obligations); whereas many nearby jurisdictions have blanket bans on foreign ownership. Naturally, with the recent change in the premiership, Japan plans to expand these reporting requirements for foreigners buying property, but not prohibitively so — only expanding who needs to report — with implementation expected around April 2026.

The other under-discussed point: Okinawa already has special-zone type levers. Under Okinawa’s promotion framework, there are bonded-area customs provisions and preferential measures, including a reported 40% deduction of corporate taxable income in some cases. Okinawa also promotes an “International Logistics Hub Industrial Cluster Zone” around Naha’s airport/port, where firms can bring goods into bonded facilities to store, sort, relabel/repack (maybe even light processing) without paying import duties up front, then re-export them to Asian markets.

Unfortunately, where Japan as a whole starts falling apart is in its ability to attract international talent. For entrepreneurs, the default pathway (Business/Management visa) has historically been paperwork-heavy and front-loaded with requirements that can be tumultuous for companies moving into the region, or neigh on impossible for those at the early stages of incorporation. Whilst minimum capital requirements have been updated for said businesses (from ¥5 JPY to ¥30 JPY), more onerous is that companies are required to employ at least one full-time worker in Japan at a physical address. Japan has tried to soften this with the Startup Visa (a “Designated Activities” status where local governments vet your plan and issue confirmation before immigration processes it), but that introduces extra layer of coordination — municipal approval first, then immigration with uneven guidance by region. From personal experience, dealing with the Japanese administration does feel like a disproportionately high amount of paperwork, further confounded by the language barrier.

If Okinawa wants international founders and investors, it needs to offer short cuts and exemptions from the mainland’s bureaucracy ideally with multilingual support to connect into the nearby cities. Just before my trip, China’s Hainan free economic zone announcement went live, which should remind policy makers that trade policy is not set in a vacuum and other economies are competing for business and talent.

So what do I think is missing specific to Naha?

First public transport. Okinawa’s monorail is great, but it’s one line and the bus service is limited. Around 80% of movement relies on cars and rarely can the roads handle the level of traffic.

Next, any economic hub needs to train and maintain talent. The Okinawa Institute for Science and Technology (OIST) is a case in point that interactional talent can be attracted to Japan (their first president was Nobel laureate Sydney Brenner), but these were set in place by founding documents that committed to English as the official language with more than 50% non-Japanese faculty and students. Unfortunately, OIST is limited to basic science, and economic hubs need internationally recognised medical and engineering education and research.

Finally, you can’t talk about Okinawa without the US military bases. Public data portals note that ~70% of land for U.S. military installations in Japan is concentrated to Okinawa, despite Okinawa representing 0.6% of Japan’s total land area. However, this can also be an economic catalyst --- redeveloping former military sites may be a route to acquire high value real estate to house new industries. If this is executed well, there is a once-in-a-generation opportunity to have huge swaths of land suddenly put to economic use.

Anyway, that’s a wrap. No idea if this will materialise in how I imagine, but I do think all of the ingredients are there…!

With much interest in how Western economies have fallen behind, the question is how to rebuild capabilities across key industries: chip manufacturing, biotechnology, nuclear power, defence, and aerospace, to name a few. From the 1960s onward, the West built significant scaled operations in these areas, yet, by the late 1990s, we had systematically gutted these capabilities and, crucially, forgotten how we identified and trained people. The AI revolution has been heralded as a mechanism to both increase the quality and reduce the price of higher education, but AI cannot replace tacit knowledge; it does, however, radically reduce the cost of signalling expertise for technical roles. One might posit that we’re heading for disaster when you combine these trends with the capital-intensive nature of these “binary outcome industries” (BOI) where success is measured in a small number of critical events — a clinical trial readout, a factory commissioning, a rocket launch. From specialist mentoring to unwavering academic standards, there may be solutions to rebuild these BOI. However, doing so will require proactive decision-making by individuals, companies, universities and governments with the specific aim to preserve delivery capability as a national resource.

Here we discuss:

1. How do binary outcome industries operate?

2. Why do you hire someone for a BOI job?

3. How do we rebuild human capital?

Subscribe now

How do BOI industries operate?

When you compare the “old industries” to new ones — technology, finance, consulting, and consumer — strange and nonlinear market dynamics reveal themselves. In pharmaceuticals and biotechnology, perfectly safe and efficacious drugs get abandoned for commercial reasons after hundreds of millions of dollars have been invested; nuclear power plants remain in various stages of completeness, spending decades in regulatory purgatory; and launch vehicles are lost in minutes after a decade of development. In contrast to the modern economy, these old industries require huge amounts of upfront investment. Less appreciated is that, akin to highly paid Silicon Valley software engineers, there is an extreme concentration of talent to deliver these mega-projects. A system of training courses and real-world assessments once identified singular leaders entrusted to lead the initiative of the day: from nuclear power plant builds and setting up factories to bringing a drug through clinical trials.

Consider the growth of the pharmaceutical industry. From the late 1800s to the mid-20th century, early pharma was a craft with many “drug hunters” travelling far and wide to harvest plants, fungi, and soil microbes to identify bioactive compounds, often leading to new antibiotics. Simultaneously, glands from various animals, and even humans, were being pulverised to find hormones with therapeutic action (see oxytocin, adrenaline etc). If one could isolate the active ingredient and scale its extraction, then one may have a flourishing business if the unit economics work. Key figures like Henry Wellcome developed the ability to judge what smelled promising, how to purify at multi-kilogram scales, and which impurities would hinder stability a year later. At this time, regulation was light and with clinical trials not mandated until 1962 in the USA (Kefauver-Harris Amendments), one could be commercially successful by merely shipping a product that didn’t obviously harm people.

As chemistry matured, so did patent strategy and regulatory pathways. Firms prioritised composition-of-matter claims and moved away from “Generally Recognized As Safe” (GRAS) routes (formally introduced in 1958) because the most defensible commercial moat was exclusivity provided by patents and regulatory bodies (not just that you could manufacture the drug). With the Hatch–Waxman Act (1984), many “generics” companies selling off-patent drugs built internal groups to attack weak patents and gain an FDA protected 180 day market exclusivity. This meant that big pharma had to walk the line between filing IP early enough to deter “fast followers” (other drugs with the same mechanism) but late enough to cover ancillary aspects, such as formulation, delivery via medical device and manufacturing routes needed before launch, potentially followed by new use cases (“indication expansion”) — if discovered. This led to a professionalisation of regulatory science and executives became experts in reading the tea leaves of “what the FDA will actually accept”.

Eventually, molecular biology and human genetics created target-led discovery and, later, biologics and cell and gene modalities. If you believe the press, the future of health is all around precision medicine; the less appreciated story is around standardisation and integration of the discovery process itself. Various internal reviews became published highlighting factors predictive of regulatory and commercially successful drug launches (see AstraZeneca’s 5R’s framework, GSK’s work on genetic evidence for approved drug indications). This ultimately led to alignment in target validation standards, CMC tractability requirements, and the clinical contexts (patients, endpoints, standard of care) that regulators and payers will bless.

Zooming out, R&D spend per approval has risen over the past few decades even as tools improved, referred to as “Eroom’s Law” (Moore’s Law in reverse) with cost drivers including higher evidentiary bars, safety expectations in larger and longer trials, rare-disease fragmentation, and payer demands for comparators and real-world value. Each layer adds irreversibility: once you commit to tox species, a biologics expression system, or a pivotal endpoint, the project path is locked. That’s what makes pharma quintessentially a BOI: a few irreversible choices compound toward one or two binary events.

Moreover, beyond the cash-intensive R&D process itself with revenue generation pushed a decade into the future, valuations jump discontinuously on data. For biotech, this creates pressure to treat interim analyses and regulatory interactions as pieces of an equity story; and for big pharma, these discontinuities allow for dynamic reallocation of capital across their pipeline as programmes get killed whilst new ones get formed, either organically, or via mergers and acquisitions.

To manage this process, big pharma historically invested in leadership finishing schools: executive development programs, mock FDA meetings, media/comms training, crisis simulations, and rotations across departments and functions. The output was a small cadre trusted to manage $100M+ assets from pre-IND to launch. Their edge was knowing who to call out of hours when your fill-finish yields crash (final manufacturing step); which vendors never hit timelines; and how to negotiate a Special Protocol Assessment without boxing yourself in (a specific FDA meeting).

As we see above, critical events define huge value generation for stockholders. Remarkably, this is not how most modern businesses operate: someone, often young, builds a minimum viable product, generates small amounts of revenue, demonstrates this de-risked concept to investors, and then external capital scales the operation to make millions, if not billions, of dollars (see Meta, Twitter, Instagram, etc). BOIs look nothing like them.

Hiring for BOIs

Why do you hire someone?

At a very high level, a company has work that needs to be done beyond the current available headcount. After posting a job advert, one needs to evaluate applicants based on skills, accomplishments, and some kind of cultural fit; candidates get benchmarked against a role spec before an offer is made. What is less explicit is that experienced hires also get judged on tacit knowledge (alluded to above), and junior hires get judged on potential — basically, can you show that you’re smart?

What is tacit knowledge and why is it so important?

Tacit knowledge is the unwritten ways of working to deliver these outcomes, earned through hands-on experience and overhearing the conversations that one would not wish to speak too loudly. For a small example, in my past blogpost, I stated how AI-generated molecules are often far more expensive than using a traditional CRO, which breaks most of the VC narratives that TechBio companies are cheaper than the alternative — arguably diminishing the value proposition of their portfolio companies. Often, when someone is being interviewed for a senior position, they are subtly (or not so subtly) hinting at the tacit knowledge they have accumulated.

The other challenge with tacit knowledge in BOI is that it’s really hard to measure. Consider that the rate of success for a Phase 2 clinical trial for efficacy is approximately 30%. Suppose we wanted to assess whether an industry leader with presumed expertise (that improves clinical trial outcomes) was better than average at bringing drugs to market.

Let p = 30% denote the market baseline and let the leader claim a success rate y. If we model each trial as an independent Bernoulli outcome with success probability p. We can then ask: how many successful trials, k, out of how many total trials, N, one would have to had run, such that y>p with 95% one-sided confidence and 80% power​, i.e., that they are better than the market.

As you can intuit, you must run 100s of clinical trials to detect small improvements over the market, illustrating why performance is hard to validate statistically in BOI. The larger differences are easier to detect. However, with all of the complexities in bringing a drug through clinical trials and all of the different stakeholders involved — is it even meaningful if a decision maker claims that are reliably be 2x to 3x better than the market? Realistically, if an executive has run over a dozen phase 2 trials, this is already quite a singular person towards the end of their career.

How do you know if someone is smart? Can you prove you’re good at what you do?

When looking at junior hires for a BOI, with or without a PhD, you are looking for evidence of logical and agile thinking under pressure, social grace, and ideally some signs that they have a level of conscientiousness and experience that their compatriots did not: managing a university club, strong academics, personal projects, or even writing a paper relating to original research. While the prestige was diminishing for all the above, AI has now well and truly undercut all meaningful ways a student used to stand out.

At the undergrad level within the Ivy League and liberal arts colleges alike, the prestige of setting up or running a social club is increasingly being treated as a résumé booster or a way to meet influential after-dinner speakers — often at the cost of genuine passion for a subject. From a testing perspective, universities have long been chastised for declining standards; for example, both Harvard and Yale reported that 79% of grades were in the A range in 2020–21 and 2022–23 respectively1.

At the graduate level, the recent step change is that AI can even write semi-convincing, and perhaps even convincing and correct, research papers — too much for any journal or conference to accommodate and gatekeeping has scaled accordingly. Historically, the main recipient of high-volume research articles were the pure mathematics journals: there was a certain romanticism that appealed to the amateur mathematician of being the next Ramanujan and finding a proof to a Millennium Prize problem. The journals generally took a pragmatic, if exclusionary, approach: blanket reject any submission without institutional backing or appropriate recommendation.

With the rise of newly formed non-research-intensive Asian and African universities, and the general growth of paper mills and low-quality venues, low-quality scientific work started seeping into the mainstream, purporting results that could not possibly be true. Retractions have surged: more than 10,000 research papers were retracted in 2023 across publishers; in 2024, Springer Nature alone retracted 2,923 papers. The established journals had to follow suit by more aggressive triage — pre-screening investigators through informal “Meet the Editor” webinars and assistant-editor checks before peer review.

Finally, the ML and AI conferences known for their rapid publication culture may now be at breaking point. NeurIPS 2025 received over 21,000 submissions in total and accepted ~25% in the main track (in previous years ~5.5% of submissions were flagged for ethics review). By submitting a manuscript, you also agree to review a handful of other manuscripts (the only way one could organise such an event). At that scale, even well-intentioned processes strain: heterogeneous reviewer quality paired with limited venue capacity has led to frequent overruling of decisions weakening the link between reviewer mark and acceptance of the work.

The power of AI is not imagined and is clear that for those without track records, the number of job postings is drying up.

The system for identifying talent worked when elites could find other ways to distinguish themselves. Unfortunately, now there’s a smaller and smaller aperture to definitively say “this person is smart because they did A, B, C”. AI reduces the dynamic range for signalling talent — which is an even bigger problem for BOIs, where performance is revealed only at a few binary events.

How do we rebuild human capital?

For individuals: Seek mentorship (especially across disciplines and departments) and avoid full-time WFH, you want to essentially build an understanding of the value chain that your business operates in so to understand how and why decisions are made. Most mentorship that people seek out defaults to prestige matching (titles, brands), but may not actually provide any meaningful insight into how business gets done.

For companies: Startups should hire senior operators early for tacit knowledge before irreversibility locks in — hugely important as your burn rate ramps up as you will only get a few shots on goal. In many BOI settings, this is a well trodden path, see Space X (Shotwell, Koenigsmann), TSMC’s founding team included Morris Chang, John C. Martin was seen as pivotal for Gilead --- and Relation are trying something similar2. We need to institutionalise apprenticeship. Finally, we could have a golden age for recruiters and hiring teams as they will be needed more than ever to do much more in person assessments when compared to online tests.

For conferences: Introduce submission fees, require code/data disclosure or registered protocols where applicable to increase the cost of false signalling.

For universities: Stop computational assessments (obviously) and return to unwavering academic standards by re-emphasising closed-book and oral exams3, so the cognitive burden is not solely on businesses as to work out when to hire someone.

For governments: Offer tax benefits and create programmes for retired executives to help and advise start-ups and growth stage companies. We need mechanisms of transferring tacit knowledge.

But don’t forget. Eventually, the old BOI industries will reinvent themselves to be industries of the future. Artificial intelligence, genomics, and nuclear fusion are going to be a young person’s game. There are always distinct differences between people who grew up living a topic and people who learnt later. You can’t fake intuition, so let’s build systems that build it.

Share

There is enormous excitement for autonomous science: letting algorithms (or agents) control key aspects of experimental design to optimize the capability and capacity for inferring scientific truth. Naturally, this is most exciting when experiments are expensive and/or technically challenging to implement (e.g. single-cell sequencing) — which is basically all of the time if you’re doing anything in genomics.

Generally, I’ve seen two (not necessarily mutually exclusive) competing visions for autonomous science:

1. An LLM-centric view of the world trained via next-token prediction on scientific texts; entities (protein X, cell type Y, etc) are emergent phenomena. The model would then suggest future experiments to perform to resolve contradictions in the training data or build greater resolution over its world model.

2. An entity-based view of the world where variables defined a priori interact with each other — typically with error bars associated to these interactions. However, in most of biology, we do not know the underlying rules or parameters governing these interactions.

   1. In reinforcement learning (RL), we learn how act (a “policy”) by trial and error to maximise a reward in an environment with sequential decisions.

   2. If there are parameters to infer, then active learning (AL) is your tool of choice to learn a predictive model.

As I explained in an earlier post, the current “AI scientist” research arc appears to have a way to go before vision (1) is close to reality.

However, vision (2) is very much a reality… or is it? Whilst reinforcement learning is hugely popular, it is exceedingly data hungry, so nearly all of the applications relate to scenarios whereby one can computationally generate new data, e.g., AlphaGo.

Active learning on the other hand never made it out of the starting gate (see Google Trends of RL in red vs AL in blue).

Active learning (AL) is concerned with how we improve the performance of predictive models by acquiring more training data: we require intelligent decisions about which data to acquire next, often in an iterative setting.

Virtually every conversation I’ve had about applying active learning (AL) in the real world with industry insiders ends with the premonition: it’s not worth your time, it will take years to get working, and acquiring data randomly is probably a safer option. My take on the reasons behind this is that:

1. There seems to be minimal mathematical justification behind some of the most simple and very popular algorithms.

2. Without such foundations, it is hard to model how the “real world” will undoubtedly break your base assumptions and the subsequent consequences thereof.

3. Therefore, no one has a strong understanding of how the marginal acquired data point relates to model error, and one cannot estimate how many experiments are actually needed to reach some acceptable tolerance.

In short, we are playing in the dark and we cannot expect AI to have a real impact on scientific discovery until we bottom this out.

Below, I give:

• Historical experience applying sequential model optimization

• My take on the field and how various ideas have been conflated

• The beginnings of a mathematical foundation for AL regression problems

Subscribe now

Practice before Theory: The RECOVER Coalition

Go back to 2020: Relation had recently formed, COVID-19 had gone mainstream, and the Gates Foundation had announced their therapeutics accelerator. Seeing the announcement, we were lucky enough to work with the team at MILA and Turing prize winner Yoshua Bengio to put together a proposal (“RECOVER”) that received funding. The idea was simple: due to the challenges in understanding the mechanism of small molecule drugs in preclinical and clinical development and how drug-drug interactions occur, can we use AI to select pairs of drugs1 that may work together to achieve an antiviral effect? This is a combinatorially difficult problem, if you have 10,000 drugs of interest, then you have ~50M pairwise drug combinations — completely infeasible to screen experimentally.

To achieve this, we created a deep learning model to predict how pairs of drugs impact an in vitro phenotype, set up wet-lab experimental systems to evaluate drug combinations, and then let the model choose pairs of drugs for evaluation. By performing this process iteratively, the idea was the model would identify increasingly efficacious drug combinations that human reasoning would not naturally put together.

Building the deep learning models was challenging but achievable, the real difficulty was agreeing a strategy for how future drug combinations were selected. A common approach from AL was proposed, called “Least Confidence”, whereby we pose a collection (or “ensemble”) of deep learning models that each have slightly different versions of the training data, and subsequently see where they agree and disagree. If the models disagree, then conceptually the models are unsure about how some pair of drugs are going to work together, and in which case evaluating them experimentally and collapsing this uncertainty (typically referred to as “epistemic uncertainty”) should be highly informative to improve predictive accuracy.

At this stage, I started noticing a few key problems with the field. Notably, it was not clear how to deal with experimental replicates. Due to the instability of various cancer cell models we used during our proof-of-concept, we could run the same experiment 3 times and get 3 different (but similar) answers! For example…

It occurred to me that if you’re trying to minimize model uncertainty (epistemic uncertainty), does this even make sense conceptually? The underlying process is noisy, so if a model is uncertain in some way, this may in fact be the correct inference.

To get around this (and noting that time was not on our side), there was a simple solution: skip trying to optimize for predictive accuracy, but to identify a different objective to optimize for. In our case, the objective became to find highly “synergistic” drug combinations (i.e., drug pairs that would be efficacious at low doses).2 This way, we could move away from “vanilla” AL methods and use related approaches like Sequential Model Optimization (SMO) and/or Bayesian Optimization (BO), which are themselves built upon active learning principles. For the full write up, the associated publication is here.

Unfortunately, if you want to understand very complex biological or medical problems, there are rarely simple metrics that optimize for. In fact, no one has identified a quantifiable objective to suggest when an AI model actually understands biology — other than predictive accuracy! If you can make a highly accurate predictive model, then one can study the underlying structure of the model and try to assign meaning to it. In other words, if you don’t know the generating process for biology, then you have to infer it, and a necessary condition for the correct process is that it is predictive.

This does not however mean that an incorrect process cannot also predict the correct outcome, or that a model with substandard predictive power cannot capture the broad structure of the underlying biological process — but we don’t yet know many “core rules” of biological regulation beyond a few key motifs, e.g., signal transduction (ligand → receptor → downstream change, incl. TFs etc) and the central dogma of molecular biology (DNA → RNA → protein). So, if the field of active learning is about improving predictive accuracy, we really want it to work in practice!

Historical context

What is the core problem?

Fundamentally, our goal for active learning is to reliably outperform randomly designing future experiments, or “random selection”.

For a brief demonstration of how good random selection is, consider the following: we wish to estimate the area of a circle (because we have temporarily forgotten the formula due to selective amnesia, but in fact any shape would work). We can draw a circle inside a box, define a set of points that can be selected (e.g., consider the pixels on your computer screen), and then if you choose points at random, an estimate for the area of the circle is the proportion of points inside the circle multiplied by the area of the box; we call this Monte Carlo simulation. When compared to using a regular grid, random selection provides a much more accurate estimation of the circle’s area with the same number of points (on average — each realisation is of course different).

However, random selection can hypothetically be beaten: if you stop the randomly selected points from bunching too close together (e.g., via Halton sequence selection, which is deterministic), we build even more accurate estimates.

With Active Learning, a key weapon in our arsenal is that experiments can occur sequentially: we can select a point (or multiple points in a batch), assess its usefulness (however we define this), and then adjust our strategy accordingly.

A brief taxonomy of AL terms and state-of-the-nation

Most of active learning’s greatest hits were written for classification: pick the next image whose label (“cat”, “dog”, etc) you are most unsure about. The math and metrics lean on likelihoods and mutual information, but when people port these ideas to regression problems (the scientific setting where we care about quantities, e.g., concentrations, growth rates, binding affinities etc), the field gets confusing.

Here’s a classic example of a mismatch: Least Confidence.

• In classification it means that if there’s a low max-probability, then we assume that there’s a high “least confidence”:

• In regression it quietly morphs into “pick the largest predictive variance.” Same name, totally different quantity.

Before talking methods, a few working definitions:

• Epistemic uncertainty: model uncertainty that can shrink with more (or better-placed) data.

• Aleatoric uncertainty: measurement/process noise that won’t go away, no matter how much data you get. This quantity can be heteroskedastic and vary over the input space.

• Bias: the systematic gap between your model’s expected prediction and the truth.

All three are local quantities (they vary across the input space). You’ll often see high epistemic uncertainty where you’re far from training data; aleatoric spikes where the physics/assay is noisy; bias where your model class is just wrong.

To reason about real experiments, it helps to sort problems into three buckets:

Type I problems: Noiseless scenarios resulting from deterministic mappings (the simplest case).

For example, suppose you want to learn the relationship between Celsius and Fahrenheit: there is a direct relationship between the two defined by:

These are typically very boring problems, and sadly a lot of AL benchmarking stops here.

Type II problems: Real-world scenarios with uncorrelated aleatoric noise, i.e., experiments are performed and we can essentially view all of them as independent and identically distributed (The “i.i.d” assumption).

Imagine we are trying to learn the dynamics governing various weather patterns from sensors located around the globe; each sensor may corrupt the true signal but we may wish to make the assumption that the functioning of one sensor is unrelated to the functioning of another, i.e., there are no correlations in the underlying noise process. However, this noise can still be heteroskedastic, e.g., we are better at quantifying slow wind speeds vs. fast wind speeds.

Type III problems: Real-world scenarios where aleatoric noise is correlated, i.e., we see systematic noise corrupting the whole system as we observe it.

Consider the scenario where we’re running hundreds of experiments in parallel (e.g., cell growth assays) and the incubator temporarily shuts down and restarts without anyone noticing, then all of the cellular models in this batch all experienced the same random event, correlating the resulting experimental measurements. This is especially relevant to hot experimental systems biology techniques, e.g., pooled CRISPR screens with single-cell readouts (“perturb-seq”), whereby we see imperfections in the manufacturing process for microfluidic chips leading to systematic changes in gene expression measurements.

If you want to see versions of the same problem with different noise structures, see the two-dimensional function below: left is Type I (no noise), middle is Type II with independent noise for each (x,y)-coordinate, and right is Type III where (in this toy problem) the noise is similar in nearby regions of the state space.

To close, most AL theory is rock-solid only in narrow regimes (linear/GP models, i.i.d. homoskedastic noise, one-at-a-time querying). Add deep nets, batching, heteroskedasticity, correlated noise, or model misspecification, and the proofs thin out. Information-theoretic approaches (e.g., BALD) are principled in spirit, but with deep learning models they rely on approximations and can end up chasing irreducible noise. Net: foundations are solid in special cases; elsewhere we’re mostly dealing in heuristics plus empirical evidence. Scientific regression needs a cleaner story.

Towards a principled approach to AL

In our recent preprint, we start dealing with some of the above problems.

First, we note that for regression problems the bias-variance tradeoff can be applied and the expected mean squared error (EMSE) can be interpreted as the sum of an epistemic uncertainty term, the bias squared, and an aleatoric uncertainty term.3

One would ideally like to select points in the state space such that the bias is close to zero, but also that the predictive distribution approximately matches the underlying noisy oracle — we do not want our predictions being more or less certain than the underlying truth. Essentially, we cannot allow our predictive distribution (in purple) to be more confident than the underlying noisy process (in green):

Naturally, when considering regions in the state space to select, some will correspond to areas whereby the bias or epistemic uncertainty rapidly collapses — and these should be prioritised for labelling when compared to the aleatoric uncertainty (that never changes)! We achieve this through calculating an approximation to the derivative of the EMSE between experimental rounds — leading to “difference” acquisition and our preprint title: two rounds of experiments can be used to estimate the gradient of the EMSE, which is then exploited in a third round of experiments (or indeed, any future experiments).

Finally, through noticing a mathematical trick, the bias-variance tradeoff can be adapted into a cobias-covariance tradeoff that allows for a neat way of proposing batches via eigendecomposition.

Recapping this in math-heavy language:

1. Target the right quantity. Decompose prediction error into

   We call the sum of the first two the reducible error.

2. Estimate reducible error – not just uncertainty. Build a covariance-like operator over the pool:

   where ΣF​ is epistemic covariance (from an ensemble/GP), Δ=δδ⊤ captures bias via the co-bias matrix, and ΣY​ is aleatoric covariance (including batch correlations). Two practical tricks:

   • Difference acquisition: compare consecutive rounds so stationary ΣY​ cancels, revealing where reducible error is actually collapsing.

   • Quadratic estimation: predict entries of Δ directly (pairs), which is numerically more stable for batching than first predicting δ and squaring.

3. Batch smartly with eigenmodes. Work on the reducible operator

   Take its top eigenvectors and, for each, pick the input with the largest loading. You end up spreading the batch across orthogonal directions of reducible error, not just piling into one noisy hotspot.

Why this is different to much of the previous literature:

• Optimizes the metric that matters. We explicitly target expected MSE, not a proxy for “uncertainty”. Such methods are referred to as “Expected Error Reduction” approaches.

• Separates signal from noise. The difference trick cancels stationary aleatoric terms; the cobias-covariance view makes correlated noise an explicit object you can model or subtract.

• Smarter batching by design. Eigenmodes give diversity: cover distinct, high-impact directions that reduce error after retraining (given various caveats about how well the underlying neural network can capture the underlying data).

• Model-agnostic but practical. Plays nicely with deep ensembles, MC-dropout, or GPs; no need for idealised assumptions to get started.

Ultimately, we are interested in designing algorithms for the messy realities of drug discovery and systems biology — replicates, batch effects, shared lab equipment, etc — and this is a small step on this journey.

In 2021, a who’s who of academics at Stanford popularised the term “foundation model”, exclaiming how by sinking millions of dollars into training runs, one could achieve vast out of distribution generalization capability. Across a range of industries, these models have boasted a range of use cases, from audio and vision generation to robotics and automation. Counter claims have however haunted the term: first, does it make sense to spend so much money? Second, providing such expenditure makes sense, can one achieve similar results at a fraction of the cost?

One field I’m most familiar with has been biomedical research. Below, I give my take on what does (and doesn’t) make sense at the biomolecular level.

Protein foundation models

In 2021 DeepMind dropped AlphaFold2, and triggered by this, a slew of “AlphaFold-for-X” companies were formed and funded, each with their own vaguely interesting twist. From antibodies to peptides, gene or cell therapies, every therapeutic modality under the sun had a Techbio startup1 waiting to sell you a story: the reason said modality never achieved commercial success was that an expensive, crack AI team wasn't available to steer the ship.

Unfortunately, the one thing many VCs forgot to check was the unit economics: essentially when you hire a team of ML researchers and engineers, you need to benchmark the strategy to the traditional way of doing things. By and large, when you compare the hundreds of millions of dollars required to bring a drug to market, spending low single-digit millions with a CRO to make a molecule for your target is an acceptable price to pay. However, the cost-benefit analysis of adding an ML research team and associated compute into the mix has yet to be elucidated… very long term, perhaps you can drive down the timeline (or even cost) of biomolecular design, but short term you are bankrolling platform creation of unclear value.

To be clear, I am specifically referring to performing ML research with large compute requirements in a company setting. As Charlotte Deane laid out in her talk at AIxBIO, published models are already an incredibly useful support tool, but (IMO) the economic argument in performing research in the area outside of academia is limited. Essentially, the models are competing over very small performance gains — gains that do not necessarily translate into real world value. Whilst I could dig out results from a survey paper and try to translate test set performance into a real world consequence, a more pertinent anecdote is that when we have been provided costs by a number of providers to a initiate a small molecule discovery campaign: the AI-enhanced companies came in at a ~3x multiple. Unfortunately, the multiple is not justified: there is no compelling data pack to date that demonstrates superiority over a skilled medical chemist.

With this in mind, it is why it’s slightly disappointing to see that 4 years later, the UK government didn’t get the memo and is funding a new protein-ligand database (~£8M for 500k structures) explicitly as part of an AI-driven drug discovery strategy2. On one hand, I’m really pleased to see the government lean into the sector. On the other hand, £8M is trivially fundable a pharmaceutical company if they felt this dataset would move the needle. Moreover, every retrospective analysis and commentary of why clinical trials fail (the largest cost in clinical development) tells us that lack of efficacy (and also safety) is the main reason for clinical trial failure, i.e., we didn’t understand the underlying biology3. One of the few reproducible signals that are predictive of clinical success is as to whether the drug target in question is linked to the genetics of the underlying disorder (Nelson 2015, Nelson 2024).

To summarise, I do not think the numbers add up — at least for small molecules. Moreover, I don’t think that designing biologics will change the paradigm a huge amount without clear forethought on what the use case is. For example, we can generate millions of candidate antibodies through a range of display methods, so merely optimising for binding isn’t differentiating; other properties pertaining to practicality and scale up may make more sense. Finally, companies like Dyno Therapeutics seem to be bucking the trend by bringing together structural models with an experimental platform to screen capsids for gene delivery — an area that has historically been challenging to optimize — everything ends up in the liver!

DNA foundation models

Contrast this to DNA foundation models: typically we are trying to learn the relationship between novel genetic variants and the downstream regulatory consequences within a certain window (say, up to ~1M base pairs). In fact, this is really two problems in one: when a genetic variant falls within a coding region of a protein (called an exon) and the mutation causes a change in the protein’s amino acid sequence, then we obtain an altered structure (perhaps providing greater justification for protein foundation models). However, approximately ~90% of variants found in genome-wide association studies (GWAS) are non-coding.

When a genetic variant is in a noncoding region, we see changes in how DNA is transcribed into RNA. This is a tricky problem as DNA is wound around histones and densely packed within a cell’s nucleus, generating complex nonlocal relationships that we wish to learn. We achieve this by feeding a transformer-style deep learning model genetics with other genomic modalities, say RNA-seq, ATAC-seq, or ChIP-seq etc. Ultimately, if you can learn the relationship between arbitrary DNA sequences and downstream gene expression, then you have the means to identify new drug targets4. For example, if overexpression of a gene causes a deleterious phenotype, one can design a drug to inhibit the function of the corresponding protein.

One aspect people seldom factor into their reasoning of why DNA foundation models are so interesting is how they can hypothetically reduce cost, labour, and practicality constraints of performing tissue profiling at scale. Imagine you decide to run a study to collect paired genetics with single-cell profiling of some tissue of interest. You need to pay for sample acquisition (e.g., patient fees, hospital fees, couriers), sample preparation (e.g., tissue dissociation, an appropriate single-cell technology), and sequencing. Fraught with logistical challenges, you will also need a capable and experienced lab that is frequently on call to receive samples! It goes without saying, this is a very expensive endeavour.

However, imagine you have an AI team that built a DNA foundation model trained on the data you have generated (say ~100 patients or so), you can now make inferences on much larger groups of patients, courtesy of large national biobanks. For example, the UK Biobank has ~500,000 volunteers, All of Us are aiming for ~1,000,000, and Our Future Health is aiming for ~5,000,000! Within such biobanks, we have patients presenting different disease pathologies donating their DNA, but seldom with the associated disease tissue of interest collected. This is a huge opportunity to detect subtle differences in how genetic variation manifests itself in large patient cohorts that would be undetectable in the original small patient group.

Regardless of how much VC capital has been raised, one cannot perform single-cell profiling on hundreds of thousands of patients!5 Remarkably, one unique property of using DNA foundation models in such a way is: if you increase model performance, one may be able to make more nuanced inferences with smaller patient numbers. So theoretically, improving benchmarked performance can translate to valuable real world gains (justifying investment).

“Virtual cell” foundation models

A new term has gained popularity, the “virtual cell”, encompassing a range of foundation models largely trained on single-cell perturbation data. Ironically, these models are not at all virtual cells (in the sense that we are simulating molecular dynamics), but for ease of reading, I will reluctantly continue using the term. At a high level their goal appears to be simple: to predict the effect of unseen perturbation on a gene expression profile. However, there are several much deeper (more valuable) problems related to this:

1. If you can link predicted gene expression (or some other dynamic ‘omic readout) to functional phenotypes you actually care about (e.g., T cell exhaustion, cancer growth etc), then one has a mechanism to propose new drug targets off the back of this.

2. If the interventional data generated can be faithfully mapped to observational human biology in situ, one may be able to find drug targets that transform a diseased state into a healthy one.

3. If one has an exquisite time-resolved understanding of gene expression dynamics, then one can definitively say what the causal chain of events was leading to the emergence of phenotype of interest. This is very important if you want to understand the potential for on-target toxicity effects.

4. Finally, depending on the out of distribution performance of the model, one may be able to perform transfer learning between cell types, or even say something about multicellularity, advanced model systems, and emergent phenotypes.

If you can do (1-4), then you have an exceptionally valuable platform — it is akin to dropping both target discovery and potentially other elements of preclinical development in the diagram below:

However, there is one key problem: there is nowhere near sufficient data in the public domain to do this properly. If analysis of recent press releases are anything to go by, there is an excessive focus on having a very large headline number of cells sequenced — not paired phenotypes or a clear focus on specific disease areas6. Naturally, I’m very interested to fix these problems, but I will save the end of this analysis for future post.

The AI scientist is coming and it’s so much better than your silly human brains — it finds hidden patterns and crazy ideas that are beyond the basic reasoning of us mere mortals. Well, that’s the narrative, but encouraged by the glitzy press release from FutureHouse, I decided to look a little deeper and give my take on what good looks like.

Below, I give:

• A mini-dissection of the FutureHouse claims

• A how-to guide to game “scientific discoveries” via a drug repositioning narrative

• What is actually useful.

ASIDE 1: I should say that I’m really excited by the concept of AI scientists, but premature declarations of victory only hold back the field. For example, there are at least 5 papers showing simple statistical models outperforms deep learning for transcriptomic profile prediction. This being said, I’ve been playing with LLMs to do some quite challenging maths and I am insanely impressed.

The FutureHouse claims

Jumping right to it, three bits jumped out at me:

1. "To be clear, no one has proposed using ROCK inhibitors to treat dry AMD in the literature before, as far as we can find..."

As soon as I read this, I felt I’d heard this story before (thank-you fellow Relationeer Cristian Regep and ALSA’s Katie Sunnucks), but perhaps related to diabetic retinopathy. A split second google search revealed a review article on ROCK inhibitors in ophthalmology more generally, so it felt like well trodden ground… typically experimentalists like to stick molecules into assays of vaguely related diseases and see what happens, so this is hardly a surprise.

I posted this on twitter, there was a bit of back-and-forth, but soon people pasted repeated links showing ROCKi in dry AMD.1

There’s a lot of talk about what constitutes “novelty”, but it seems that if there’s this much debate, it’s safe to assume that this isn’t novel.

However, there’s actually a more pernicious point: LLMs tokenize words not concepts. Therefore, it is not clear that was not data leakage; essentially, if ROCK inhibitors were in the training data for wet AMD, does the model differentiate between “wet AMD” and “dry AMD” as separate entities? As far as I am aware, the answer is no.

There’s also a softer claim, for example:

2. “This is the first time that we are aware of that hypothesis generation, experimentation, and data analysis have been joined up in closed loop”

Depending on how strict your definition is (particularly on ‘data analysis’), this claim is also questionable. For example, people have been for quite some time using active learning (AL), reinforcement learning, or sequential model optimization (SMO) for experimental design. For example, if you want to search through the combinatorial space of drug pairs, we used SMO to find synergistic drug pairs in 5 wet-lab/dry-lab cycles. We’re certainly not the only ones.

At the end, we have the key caveat:

3. “Also, this discovery is cool, but it is not yet a "move 37"-style discovery. At the current rate of progress, I'm sure we will get to that level soon.”

This part I’m really not sure of. The reason is that we will never see nonlinear results if we constrain ourselves to highly gamable drug repositioning narratives. Essentially, if this is the benchmark, it’s so easy to cheat, how can we expect to make genuine progress? Let me explain.

Gaming “scientific discoveries” via drug repositioning

A few observations about drugs and disease:

• Many chemotherapies and immunotherapies work across many types of cancer.

• Many antibiotics work across many infectious diseases.

• Many anti-inflammatories therapies work across many autoimmune diseases.

Basically, core biological mechanisms are repeated over and over, and new cutting-edge drug discovery revolves around finding new mechanisms specific to some organ system or disease, not rehashing old mechanisms.

However, imagine you want to come up with a (fake) compelling story, here’s what you do: consider the graph of drug-target-disease triplets with edges like this

• Edge 1: Drug A is interacts with target B

• Edge 2: Target B modulation can be used to treat disease C

and then we can state the common knowledge that “Drug A treats disease C”. With me?

Suppose now you compliment this data that:

• Edge 3: Drug X is linked to target B

then you can have a “eureka” moment from your “AI scientist” that “Drug X treats disease C 🚀🚀🚀” — Not much of a revolution.

ASIDE 2: As an alternative to Edge 3, perhaps we can also use Drug A to treat (new) disease Y if target B is also appropriate for disease Y. Again, not groundbreaking.

But why haven’t we caught onto this trick? In my mind, it’s due to a lack of knowledge for how the pharmaceutical industry actually works. Essentially, pharma companies seldom want to publish literature specifying how drugs work outside of their primary FDA approval without extensive scrutiny. This is for a bunch of reasons, namely:

Pricing: Imagine you have a drug on the market in an expensive disease area (say, some rare and aggressive cancer), but then someone tells you that you can also use that drug to treat an inexpensive disease area (say, mild headache), what do you do? The savvy businessman would tell you: do nothing! This is because as soon as you get an approval for headaches, then you will need to offer the drug at bargain basement prices so it is competitive with ibuprofen, but then suddenly you will see that lucrative oncology market disappear as cancer patients buy copious amounts of your new headache drug!

Whilst this may contravene the intended label for the drug and you have a delightful intellectual property portfolio and litigation strategy saying that prescribers can’t do this, it’s not hugely enforceable.

Crossover safety warnings: Here’s another issue, imagine your drug is being used in both acute and chronic settings. Perhaps after long term use, an epidemiologist notices an increase in the rate of heart attacks. Now you have an obligation to bring this to the FDA, and perhaps they may decide that due to the availability of other treatment choices in the acute setting that maybe your drug should be withdrawn!

Basically, there’s a lot of very interesting non-scientific reasons why interesting scientific ideas do not go very far — or get published.

What would a useful AI scientist do?

This part is quite a struggle. If not recovering plausible hypotheses, how do you evaluate success and what are good project ideas? A few principles jump out:

1. Making quantitative estimates. For example, what target when knocked out will inhibit my phenotype by >90% in a 3 week physiologically relevant assay? Here, we need formal representations of variables that describe disease, e.g. how is the phenotype measured? How to incorporate time? What is a knock out? This will all need to be extensively experimentally validated.

2. AI for lab automation interoperability. Imagine if you run an experiment in Lab A; is this the same as Lab B — even if they have different equipment? If they’re different, how different is this? Are there rules where experiments replicate, and rules as to when you’re comparing apples with oranges? Before you even build an AI scientist, it would be great to know what training data you should use. It’s worth listening to folk like Vincent Alessi who have smart ideas in this space.

3. Chasing value. Something we wrestled with at Relation: how to know if a project makes sense to begin with? At a high level, one should research what the “old fashioned” way of doing something is; build an understanding of cost/time/practicality of doing it in this way at scale; understand the incremental benefit of doing it using new sexy “AI method”; and then calculate if it moves the needle. A great idea is to look for new use cases.

ASIDE 3: The other option is finding some new way of analysing data that no one has done before, but I suspect this is substantially harder for AI if there is no precedent – now wouldn’t that be exciting?!

Finally, it’s just worth remembering that we’ve been here before in molecular design, and any progress we make will take years before its in the clinic.

Jake originally started this work in 2024 and as such it is written in the first person; Michael got involved to provide some much needed historical perspective and runs a specialist consultancy in the area.

Motivation

Here’s a pretty harmless example: have you ever wondered whether you can claim VAT back on entertainment expenses? Thankfully, a consultant has provided a rather definitive flowchat (below, from 2017), but it begs the question: why on earth is the system for something so trivial so complex? If you were to start from scratch, no one sensible would design such a policy.

When everyday interactions induce such friction, what are the economic costs to such complexifying triviality at a societal level? Could this be part of the explanation for why the UK economy is flatlining?

Observations

Consider the following:

• From trains to nuclear power plants, all major infrastructure projects appear to be hopelessly slow, overbudget, and uncompetitive when compared to European peers.

• Despite having a single national healthcare provider, the UK was unable to build a cost-effective (or even, effective) COVID-19 tracking system.

• The UK has the 2nd largest HR sector globally (normalised against population size).

• Less than <5% of projects by the Ministry of Defence were given a “green” rating by the National Audit Office, with all remaining projects labelled with serious shortcomings or as “unachievable”.

• In spite of the UK’s global reputation in STEM prowess, our premier AI research institution appears to have been largely irrelevant in the face of modern deep learning advances, especially Large Language Models.

• The UK appears unable to jail violent repeat offenders, with various loopholes often attributed to disappointing jail terms and conviction rates.

Below, we give thoughts on:

• The current situation: why we are here and what we should be aiming for.

• Challenges with radical reform and what we can learn from them.

• How should one design a system from scratch with policy examples.

• What is unique about the UK, when compared to the world at large.

Theory

The way we, British, live our lives is governed by case law, leading to rules, regulations and customs slowly accumulating over generations. As the empire expanded, the United Kingdom thrived from subjects building significant fortunes overseas, before eventually returning home and encouraging the adoption of laws to protect their new-found wealth. The world economy has changed, soon to be transformed further by AI, and Britain is struggling to navigate this new globalist age; she is without an empire but with the legal and regulatory baggage of a bygone era.

In recent decades, many former colonial states have overhauled historical laws and constitutions with impressive results (e.g., Hong Kong). Whilst many analyses have highlighted cultural factors, I argue that simplification is key — not only the bureaucratic process in and of itself, but crucially the underlying principle, policy position, and mechanism should be simple. In short, simplicity is transparent and if set up correctly — hard to exploit whilst motivating productive behaviour. A frictionless experience of everyday life should be the goal of a globally competitive Britain and one I advocate for below.

Subscribe now

The British Policy Landscape: A Maze of Complexity

After reading Foundations, an idea solidified: UK government policies and administrative processes are bogged down by two intertwined problems. First, we have myriad specific “if-then” code blocks that are vaguely understandable (and you find glimpses of them on government websites). But second, we also allow for discretion and extenuating circumstances to be considered.1 This affects all political persuasions and inconveniences every socio-economic group.

Here are two everyday examples. First, consider how your personal tax bill is calculated: income tax is banded based on earnings; capital gains have a fixed rate; if you buy a house, you pay stamp duty etc. However, all these forms of tax have various exemptions and rebates (marriage, children etc). The logical conclusion to this is that past a certain point, people (and companies) inevitably engage with accountants and lawyers to reclassify one form of income as another to minimise their tax bill.

Second, imagine a homeless person attempting to access housing. This task is apparently so complex that the leading UK homelessness charity, Shelter, primarily provides advice on how to get housing from the government — not providing it themselves. Our desire to algorithmically classify the relative severity and neediness of our most disadvantaged also leads to perverse outcomes: incentives now exist to misrepresent one's personal circumstances to game the system.

This all becomes hugely wasteful: too many of our most gifted and able members of society build expensive esoteric expertise for hire so that others may effectively navigate daily life.2 3

The Power of Simplicity

When pitching to potential investors or collaborators, I’ve always been (strongly) advised that any core idea should be encapsulated in a single, clear sentence. As a mathematician and scientist, I found this extremely challenging; after all, every idea demands context, scale, and nuance. Yet, the ability to distill complex ideas into a single, resonant statement demonstrates clarity and vision.

This principle of simplicity has proven to be a major force behind some of the world’s most successful companies. Consider Apple; Steve Jobs’ relentless pursuit of simplicity created products so intuitive that even toddlers and seniors can operate them after minimal instruction. This isn’t just about consumer electronics; it’s a principle that applies to the broader economy. By stripping away unnecessary complexity, we can create systems that are both accessible and robust.

The British state is in drastic need of simplification, but why is it complex in the first place?

Elites lobby for complexity — hindering economic growth

In “The Rise and Decline of Nations”, Mancur Olson warns that when small, well-organized interest groups gain power, they steer government policy to serve their narrow benefits — even if that means layering on complexity that ultimately stifles progress. For example, in pre-revolutionary France, a patchwork of feudal dues placed heavy burdens on peasants while enriching the nobility: tenants had to surrender a portion of their harvest (champart) and were forced to use the lord’s facilities (banalités), such as his mill or oven, creating an opaque and exploitative system that hamstrung agricultural innovation and deepened rural hardship. The French Revolution dramatically transformed this legal landscape with the National Assembly abolishing the feudal system via the August Decrees (1789), canceling personal servitude and most feudal dues (with complete cancellation by 1793). The natural conclusion drawn is that revolutions reset such adverse policies.

Within “The Shock Doctrine”, Naomi Klein expands this point and details how deadly events like wars, revolutions and natural disasters can quickly lead to major shifts in corporate ownership, organizational structures, and political philosophies — in particular, highlighting free market economics. Regardless of whether you believe such revolutionary moments shift the political compass left or right, they do allow for such nonlinear shifts to occur. Peaceful, painless and well-functioning examples of transformation involving zero fatalities are regrettably hard to come by. It should go without saying that revolutions are not desirable,4 however finding ways to “reset” key aspects of British policy is becoming hugely necessary. It also somewhat goes without saying that Britain hasn’t had a revolution in quite some time (1651), so now we discuss lessons from history.

Undoing complexity: qualities of successful change management

If we are conceptually happy that complexity ferments over time, then what impact does this have on state capacity? Can one execute on major infrastructure projects? Finally, what do attempts of reforming runaway complexity look like?

Change management literature argues that project leaders focus too much on project design, and less on project implementation. As Kennie and Middlehurst called it, the most common approach is the ‘”Tsunami”: everybody involved, massive communication, then hope’. There is clearly a gap between expectations and reality, even when the stakes are high.

More nuanced, quantitative studies have aimed to build solid evidence to help delivery, particularly on large, expensive projects. As studies from Oxford Global Projects (OGP) have shown, drawing on thousands of case studies and data, most large projects overrun on cost and time, see below. Not well visualised are “black swan” projects, like the Olympics, which have historically overrun sometimes over 400% of budget.5 Naturally, overruns are very clearly linked to more regulated industries.

So why do project managers not simply adjust their project budgets, usually upwards, to be more realistic?

The OGP researchers identify ‘uniqueness bias’: people believe their projects are more unique than they really are. Instead, they encourage their readers to consider their project as less ‘unique’, embedding it within the evidence available. Indeed, the evidence presented is encouraging enough that we listen to it: simplifying the state, particularly in a time of relative peace and security, will be very hard.

Political interest in transformation

Interestingly, debates about wholesale work reorganisation have historically taken place amidst a curious mixture of sociologists, management consultants, industrial activists of both the radical left (e.g. the Communist Party of Great Britain; Socialist Workers’ Party) and right (e.g. British Union of Fascists), and science fiction authors, for example Francis Spufford’s Red Plenty and Kim Stanley Robinson’s Ministry for the Future.

By comparison, and crucially, the parliamentarian centre-right and centre-left have paid much less attention to such organisational or technological topics over the past century and more. This means these themes have rarely featured in political manifestos or indeed government policies.

Interesting experiments or ideas in work organisation tend to manifest in outliers such as Stafford Beer’s experiments with Cybersyn in Chile from 1971-73 or the 1976 Lucas Aerospace Workers’ plan.

An exemplar of a top-down project, Cybersyn was a project implemented by Chile’s socialist Allende government from 1971, augmented by British management consultant Stafford Beer, to use operations research and ‘management cybernetics’ to control Chile’s economy and resources centrally. While initial experiments seemed promising, Pinochet’s violent coup in 1973 effectively shuttered the project.

The 1976 Lucas Aerospace Workers’ plan was drawn up in response to job cuts at the Lucas Aerospace Corporation. In contrast to Cybersyn, the Lucas plan was developed from below — by shop stewards and union leaders. The shop stewards researched and prepared a detailed plan of new product design, workplace reorganisation, and training to simultaneously achieve two goals: i) to stop job cuts by generating new, desirable products and therefore work, and ii) to shift Lucas’ outputs from military equipment to sustainable and socially useful products. While the Lucas plan did not proceed, its bottom-up nature, foresight and attention to detail has prompted a number of documentaries (for example in 1978 and in 2018) and political debates since.

Noted above, there is a reason why these outliers tend to be utopian: execution and implementation are much harder than conception or design.

British reform is tepid compared to American reform

Within the UK, the language of cutting red tape has been popular for some years. For example, under the Conservative-Liberal Democrat coalition government, in 2011 the Prime Minister David Cameron launched a new regulatory system to reduce ‘costly, pointless, and illiberal government red tape’. This involved ‘introducing a new one-in, one-out rule, meaning Ministers have to identify an existing piece of regulation to be scrapped for every new one proposed’. This became known as the ‘Red Tape Challenge’. The Red Tape Challenge involved consultation, and essentially crowdsourcing, challenges to existing regulations from stakeholders. The idea was that ministers would then be summoned to defend their regulations against these challenges. Within several years the government claimed to have removed or abolished thousands of regulations. The Red Tape Challenge appears to have dissipated with the centrality of the Brexit referendum from 2016 onwards.

Contrast this to the US Department of Government Efficiency (DOGE), conceptualized in 2024 and implemented in 2025, as a highly apt case study. Having been underway for around three months at the time of writing, DOGE has gained a reputation for chaos, duplicity or even outright failure. For example, consider the reported firing and rapidly rehiring of 300 employees of the US National Nuclear Security Administration (NNSA), which controls the US nuclear weapons stockpile.

However, this interpretation underrates Musk’s risk appetite for organizational experimentation; he clearly believes such experiments reveal essential components versus faulty or inessential ones — no matter how public these apparent ‘failures’ are.

Consider:

• The relaunch of social media platform X featuring an interview between Musk and then-presidential candidate Donald Trump, beset with technical difficulties (August 2024)

• Experimental SpaceX Starships which disintegrated after launch in January and March 2025.

Moreover, while it has a reputation as simply an aggressive cost-cutting exercise, part of DOGE’s remit also relates to government transparency. A glance at the DOGE.gov website reveals novel new ways of conceptualizing the state at the top level:

• Restructuring and headcount reduction (the most reported aspect)

• Cancelling projects deemed wasteful

• Interrogating state reliance on private consultancies and cancelling contracts

• Calculating savings (aggregate and per taxpayer)

• Developing an ‘efficiency leaderboard’ of all federal agencies

• Transparency of employment by agency including i) years of tenure ii) annual salary and iii) average employee age. (DOGE’s source: US Office of Personnel Management)

• An Unconstitutionality Index: the ratio of agency rules to laws passed by congress, per year. (2024 – 18.5:1)

While in aggregate these changes may seem unfamiliar, eccentric, even, most are in fact the culmination of many organizational transparency proposals, or the logical conclusion of existing ones.

At the present moment, when you compare us to our trans-Atlantic partners, you see a distinctive vision of what 2020s efficiency looks like: with a remit across all federal agencies driven by a business elite with a reputation and, to an extent, track record of cost-cutting success. In contrast, the British approach feels very consultative, very admission-based, barely radical, allowing for significant omission when someone gets too close to an interest that someone wishes to protect.

After 100 days of the Trump presidency, observers are now attempting to calculate DOGE’s actual savings. Time will tell if the DOGE’s interventions are successful in terms of achieving long-term cost savings, increased state efficiency, and better value for taxpayers.

A recent study has shown that historically US and UK efforts over the past century were much closer in tone and style, and quite different from DOGE’s current approach. Indeed, whilst most overseas governments and policymakers will reject the tone and pace of DOGE, in fact the reception of what DOGE wants to do has generally been positive, including from centre-left governments such as the UK’s. Labour established the Office for Value for Money in late 2024, and, at the time of writing, Labour Together’s ‘Project Chainsaw’ is gaining traction, with the Prime Minister Keir Starmer and other Labour figures suggesting that while the language may be overheated, necessary change is very real.

Simple mechanisms decrease cognitive load and economic friction

Recently, I read “Designing Simple Mechanisms” by Shengwu Li, which revealed something obvious but profound: you can have completely different outcomes depending on how two mathematically equivalent marketplaces are presented to participants.

A classic illustration of this is found in auction design. Consider the “second-price sealed-bid auction” (SBSPA) conceptualised by economist William Vickrey, each bidder submits a bid privately and simultaneously (to acquire some item of value). The highest bidder wins the item, but the price paid is equal to the second-highest bid.

Imagine you’re participating in such an auction. Depending on how competitive you are, you may immediately start playing mind games against the other bidders. For example, trying to work out their spending power in relation to yours, perhaps bluffing how much you would bid, or trying to form coalitions to underprice the item. Although theory shows that bidding one’s true internal valuation is technically the optimal strategy, this rarely happens in practice when real humans participate in such auctions.

In contrast, consider the dynamic ascending auction. The auctioneer starts by announcing a low initial price for an item, and bidders signal their interest by raising their bids in small increments. At each step, every bidder simply compares the current price to their own private valuation—if the price is below what the item is worth to them, they continue bidding; if it exceeds their valuation, they drop out. The auction continues until only one bidder remains, who then wins the item at the final price. However, this final price is only a fraction more expensive than what the second highest bidder would pay — therefore the two auctions are mathematically equivalent! Except in one version, everyone acts together at a single point in time and in the other, we have a sequential process. Empirical studies consistently demonstrate that in ascending auctions, participants converge on truthful bidding far more reliably than in sealed-bid formats.

The key advantage of the ascending auction is that it makes the optimal strategy transparent. At each increment, bidders need only focus on a single, clear comparison: Is the current price lower than my private value? If yes, keep bidding; if no, withdraw. This simplicity minimizes the room for error and levels the playing field for bidders with varying levels of strategic sophistication. It also enhances trust and transparency because the decision process is visible and intuitive — participants see in real time that deviating from truthful bidding offers no benefit. Economists often refer to such processes as “strategy proof”.6

In a sense, participating in the UK economy is like participating in an SBSPA, we spend huge cognitive effort trying to understand a complex system. Even though the government would prefer people to play fairly and there to be trust between citizens, from a game theory perspective, this is not being encouraged. In today’s complex global economy, adopting simple, transparent rules can help align incentives, reduce errors, and foster a fairer, more robust marketplace.

What does good look like?

Imagine a world in which every key policy can be summarized in a sentence that captures its essential meaning and applies in almost every situation. Such clarity would serve as a common reference point for both citizens and policymakers.

For inspiration, consider Hong Kong wherein corporate profits, personal income and property tax (applied to rental income) are all approximately taxed at a flat ~15% rate above a modest threshold. This system does not recognise citizenship and residence is irrelevant. There is no sales tax (or VAT) and no capital gains tax. It's simple to import and export with the coastal city operating as a free port. Whether you agree with such a policy or not is missing the point: such an approach is incredibly simple. Whilst the form is getting more complex as years go by, it is still only a few pages in length.

Consider yourself a random member of Hong Kong society. Whether rich or poor, there is very little to be gained from trying to game such a system. Many of the micro-optimizations we try in the UK (see types of businesses in the UK), would have no benefit — the tax rate would remain the same. There are also second order benefits: the need for a service economy of tax accountants and lawyers reduces, and other professions can take prominence.

Whilst such a simple worldview can draw criticism — Hong Kong does have a sizable population in poverty (see works in progress piece explaining this) — its taxation system is considered one of the most straightforward systems globally. With such a system in place, there is the aforementioned tendency to make things more complicated over time, for example, the recent move to introduce a top tax bracket (compared to changing the base rate). Whilst this may encourage an ultra-high income individual to game the system, this is still a far cry from the UK.

Some very simple policies

There are some rules and systems that cannot be gamed and when market participants have fewer options available to them, the more likely it becomes to design fair systems. Below, I muse over two areas whereby I believe one can achieve strategy-proof outcomes.7

The astute reader will note that when I refer to numbers X and Y below, one can superimpose political positions onto these. For example, perhaps a left wing position may state that X should be very large, and perhaps a right wing position may state that X should be very small. Here, I’m actually less interested in what the actual numbers should be, but more about making a point about how a mechanism operates.

Universal basic income (UBI) & tax: Imagine every citizen received £X per year from the government and paid a flat Y% of their income (regardless of source). Black markets aside and assuming that UBI covered some minimal quality of life, if one wanted a higher standard of living they would have to earn more (assuming that all other benefits were wrapped into the UBI). Bizarrely, such a motivation is not necessarily true in the UK anymore: there are numerous reports on how increasing the number of hours can lead to a disproportionate reduction in benefits for part time workers.8

Migration: Advanced economies around the world universally agree that some level of migration from both developed and developing countries is desirable. Imagine each year we randomly accept X individuals on humanitarian grounds and Y individuals using some agreed upon cost-benefit analysis.9 By creating two transparent routes for entry, one can move away from political point scoring. Moreover, economic migrants know that they have been selected on some meritocratic basis.

The above was quickly conceived and needs a real economist to sweat the details. However, such a data-driven approach makes it easier to monitor, adjust targets over time, and behaviour becomes more predictable. Regular reviews based on objective metrics help ensure that the policy remains fair and resilient against attempts to exploit any loopholes, making the system more robust and accountable overall.

The difference between British problems and global problems

Many aspects of the above are arguably problems applicable to the whole of the West (or even the whole world), however there are some points whereby the UK suffers from very specific problems.

At a cultural level, we overly reward and ennoble strategic executive roles over operational roles. Historically, this may originate from how we structure the army: we have the officer class who went to university and the “squaddy” who may have dropped out of high school. This manifests as an oversupply of various nontechnical consulting roles with the few technical strategy roles focused almost exclusively on quantitative trading. Outside of finance, this dichotomy between strategy and operations is in direct conflict with most strategies that emerge from an on the ground understanding of specific technical tactical tissues at hand.10

If you follow the Crush Crime initiative, it seems there is a huge backlog in bringing dangerous criminals to trial. Connected to this, the UK’s legal system is widely regarded as one of the world’s most complex due to a unique combination of historical and structural factors. It has evolved over centuries through judge-made common law (precedents) rather than a single written code, leading to a densely layered body of rules. Uniquely, the UK encompasses three different legal jurisdictions within one country – England & Wales, Scotland, and Northern Ireland – each with its own laws and court system (plus 14 overseas territories, most a hangover from the British Empire). Compounding these factors is an unusually deep court hierarchy: cases can progress through up to four tiers of courts on appeal (from local magistrates’ courts all the way to the Supreme Court). By contrast, most countries lack this degree of jurisdictional complexity and have far fewer court tiers, which is why the UK’s legal framework stands out as exceptionally intricate in global terms.

Lessons from Europe: Finland is committed to simplicity by default

Finland is often studied by other countries for its social welfare, education, gender and, increasingly, defence aspects. However it is also valuable to note its direct and explicit commitment to government simplicity and efficiency. Indeed, the Finns realise definitions are important, as efficiency can be conceived as going beyond cost-effectiveness. To quote the National Audit Office of Finland (Valtiontalouden tarkastusvirasto; VTV):

“Efficiency means on one hand economic efficiency of an authority or the achievement of objectives by using as little resources as possible. On the other hand, it also means flexible and error-free operations of a public or private party. According to the requirement of efficiency, a public party should consider not only its own expenses but also the expenses of other parties.”

The Finnish logic, including why it experimented with UBI for two years, is that policy complexity leads to additional costs for the state and its citizens. Essentially, that simplicity is cheaper for citizens’ taxes and time, and therefore should be a clear goal of state entities. This has helped to ensure the continuance of these principles across successive Finnish governments of many political complexions. A similar rationale exists in Finnish customer protections, in that businesses are expected to provide services and information focusing on the least-informed customer. This in turn reduces the incentive for businesses to provide services requiring their customers employ advisors to help them navigate the customer experience.

In closing

Historically, Britain has prided itself on being the banker for the world — a role that ultimately became untenable as global capital markets evolved. Today, the true sources of wealth lie in land and intellectual property, not in the labyrinthine financial services that once defined our economy. The arguments presented here advocate for a one-size-fits-all, principle-based approaches that might serve as the foundation for a modern British constitution. By embracing radical simplification, the UK can untangle its regulatory web, restore fairness, and re-establish a foundation for genuine, sustainable growth.

If we manage to fix this, looking forward Olson’s insight remains vital: without constant vigilance against the influence of narrow special interests, even sweeping revolutionary reforms risk evolving into new, complex bureaucracies that can stifle long-term economic dynamism.

Subscribe now