I had just kicked off a half-dozen parallel Deep Research queries, Windsurf’s coding agent was working on a new frontend feature prototype for me in the background, and while these tasks ran I caught up on Slack and emails — tabbing over occasionally to check on the AIs’ progress.

Hey, I thought, all this clicking around reminds me of when I used to play StarCraft (bear with me here).

If you haven’t played StarCraft, it’s a competitive real-time strategy (RTS) game. Each player controls a mix of workers, combat units, and buildings. The objective of the game is to destroy all your opponent’s buildings.

StarCraft 2 (the latest iteration, released in 2010) looks like this1

A typical game of StarCraft 2 is a 1v1 that lasts about 20 minutes and progresses through a few phases:

• Early game: Gathering resources with workers, constructing an initial set of buildings (a ‘base’), and fighting in small skirmishes

• Midgame: Expanding to new bases, building an army, and pushing to control territory and resources on the map

• Lategame: Large, potentially decisive, clashes with armies of powerful upgraded units

There are important strategic choices to make in a game, such as the buildings and specific army composition you choose to construct (there are rock-paper-scissors-like dynamics in combat where some units ‘counter’ others). But more important than strategy is each player’s execution skill, which players break down into two dimensions: ‘micro’ and ‘macro’.

• Macro is economy management; how effectively each player uses their workers to gather resources, and how efficiently they spend those resources to construct buildings and combat units

• Micro is the fine-grained control of units — focus firing, moving damaged units to safety, targeting abilities. In combat, quick repositioning can be decisive, and small forces controlled by a player with ‘good micro’ can beat much larger ones

A common piece of advice to new players is to practice macro before anything else; simply having more stuff than your opponent can overcome deficiencies in micro or strategy.

Both micro and macro are components of player activity. At the climax of a StarCraft game, a player might need to manage a substantial economy, a few dozen buildings, and hundreds of units spread all over the map.

Since the game space is continuous, the number of potential actions available to a player in any given instant are practically infinite. Think of this pool of potential actions as a resource; a player that can do more things per unit of time can exploit this latent resource and get a decisive advantage.

The standard measure of player activity is ‘Actions per Minute’ (APM). APM is simply the average number of productive commands a player issues in a minute — including moves, attacks, building orders, resource gathering tasks, etc. APM is also one of the most important metrics for evaluating player skill2. A beginner might have an APM of 50 or so, a strong amateur perhaps 200. A professional player might be as high as 400 (~7 actions per second!).

APM by in game rank, sorted from lowest rank on the left to highest. The graph is from a study where the developer of StarCraft (Blizzard) looked for factors that correlated with player skill (i.e. ingame ranks) and found that APM was the best of the metrics they looked at

While APM is an imperfect measure, it does correlate well with skill because to have a high APM you need to:

1. Identify actions you can take

2. Do them (fast)

To see why APM is a good skill metric it helps to view it as ‘load-handling capacity’. A core part of player skill is managing APM budget and focus; there is always something productive to do at any given moment3.

Strategy becomes decisive when players have comparable mechanical ability. Many effective strategies in StarCraft are really techniques to create imbalances in required activity — for instance, attacking multiple bases at once, or using distraction tactics so the opponent’s attention (and APM budget) is stretched thin by having to respond to multiple concurrent threats.

My favourite definition of strategy is from Michael Porter: “The essence of strategy is choosing what not to do.” Strategy is a filter that guides action selection, it is necessary because even the best humans are action-limited.

But if you could do everything, would you need strategy?

Back in 2019 DeepMind unveiled AlphaStar, a reinforcement learning model built to play StarCraft. AlphaStar was strong — it managed to beat professional players in exhibition matches.

While DeepMind capped AlphaStar’s average APM to ~300, the model actually managed to get around this limitation by conserving actions and bursting up to 900 to 1500 APM in short intervals during fights4. If you watch the exhibition games you get the impression that superior micro is a big part of the reason why AlphaStar was able to win: AlphaStar has a tendency to build relatively cheap yet highly maneuverable combat units (‘Stalkers’), take them into fights they shouldn’t win (based on unit composition), and outmaneuver the opponent to get incredible value per unit.

What AlphaStar showed, in my view, is that perfect mechanics can make strategy obsolete. Humans have to use strategy to prioritize actions. Computers, at least from our point of view, don’t need to prioritize — they can simply do everything at once. Even if you have the best strategy in the world it’s basically impossible to beat someone who takes two times, or ten times, as many productive actions as you. Actions dominate strategy.

So where am I going with this?

The reason why I was reminded of StarCraft was because it was the first time I felt like my own APM was an important factor in my own ability to get stuff done at work; the more processes I could kick off, the more useful outputs I could generate.

I suspect that the next few years will be quite strange for knowledge workers; AI labs are starting to roll out more agentic tools (e.g. Codex) that work in the background, and critically can be run in parallel.

These agents don’t even have to be particularly good to generate value, as long as they expand the capacity of the team. For instance, we’re trialing out Devin (“the AI software engineer”) at my company. It’s not a particularly good programmer — it fixes perhaps 10% of problems we throw at it and we have to reject most of its code suggestions. However, the problems it does fix might not have been ones that our team would have gotten to for a while, if at all. It’s relatively cheap and easy to throw a bunch of Devin agents at a problem and see what happens.

Assuming AI agents continue to get better, the knowledge worker’s job may shift away from planning and strategy, and towards triggering and overseeing large parallel jobs and gathering context to ‘feed the agents’. I think we’ll have an intermediate phase where high performers are those with high APM to trigger and manage automations or agents. But these action calling tasks will soon enough get abstracted away into new orchestration and context management platforms. Not long after, the AI will be triggering itself.

Companies today operate much closer to turn-based strategy. Their 'clock speed' is dictated by human decision-making cycles, reporting structures, and meeting cadences. The lower your clock speed, the more important choosing what to do with your action budget is. Skills like navigating bureaucracy, managing stakeholder alignment, and building consensus are valuable because the organizational APM is low. These are skills optimized for a low-frequency environment and are intended to increase the average value of a given action or decision.

There’s perhaps another type of organizational design that is viable in a world of many cheap and ‘good enough’ agents that prioritizes increasing the number of actions. We could see a bifurcation in how companies operate, based on whether they want to optimize for number of actions or average action quality.

But how are ‘AI agents’ different than current software or contractors that allow companies to do many ‘actions per minute’ already?

I see LLMs and their kin as tools to for knowledge workers (and their organizations) to boost ‘abstraction leverage’ (i.e. simple instructions that kick off complex behaviours).

Because actions are expensive today, you need to spend a lot of upfront time in planning mode (e.g. writing software specs, outlining SOPs for contractors to follow, system design). The promise of agents is that planning gets subsumed into action, and the AI can go off and iterate until it achieves some objective (“Action produces information”). This is much messier and error prone than the current way of doing things, but plausibly more likely to find the optimal solution5.

Finding an optimal course is a tree search problem, when action is expensive you need to plan more and use heuristics to prioritize how the tree gets explored. When actions are cheap and fast you just explore the whole tree.

So what remains for humans to do in this new type of organization?

• Gathering rare context for the machine

• Setting the ‘system prompt’ of the organization; taste and preferences

• Managing by exception (out of context problems)

• Maintaining relationships

• Oversight e.g. visual dashboards of what is happening, checking value alignment

These new jobs will require new tools, processes, and decision interfaces. Even if this never comes to pass in practice, executives and capital are placing bets along these lines; I’m already talking to companies who are rethinking how they design their organizations to take advantage of the capabilities afforded by this technology. The risk of inaction is that a competitor will do to them what AlphaStar did to the professional StarCraft players: win by doing more stuff.

Are you a software developer interested in building agent orchestration platforms? Email me at alex@convoke.bio

“The Lackawanna Valley” is an exemplar of a trope in American art, one that the historian Leo Marx refers to as the “Machine in the Garden” in his 1964 book of the same name2. The rapid industrialisation of America in the 19th century bled through to the era’s paintings and literature, where ‘machines’ (mostly trains) begin to appear with some regularity in depictions of the ‘garden’ of the American countryside. Technological progress manifested by the encroachment of these machines was then seen in a broadly positive light. To the citizens of the newly independent United States, the land was a new Eden, or paradise, that they were destined to conquer — with the aid of technology.

One of the more spirited affirmations of the American perspective on progress at the time is given in Timothy Walker’s “Defence of Mechanical Philosophy,” which reads like an 1831 version of the Techno-Optimist Manifesto. In it, he says:

“Where [Nature] denied us rivers, Mechanism has supplied them. Where she left our planet uncomfortably rough, Mechanism has applied the roller. Where her mountains have been found in the way, Mechanism has boldly levelled or cut through them. Even the ocean, by which she thought to have parted her quarrelsome children, Mechanism has encouraged them to step across. As if her earth were not good enough for wheels, Mechanism travels it upon iron pathways. Her ores, which she locked up in her secret vaults, Mechanism has dared to rifle and distribute. Still further encroachments are threatened. The terms uphill and downhill are to become obsolete. The horse is to be unharnessed, because he is too slow; and the ox is to be unyoked, because he is too weak. Machines are to perform all the drudgery of man, while he is to look on in self-complacent ease.

But where is the harm and danger of this?”

Walker wrote his defense as a response to the pessimistic view of technology espoused by the Scottish essayist Thomas Carlyle. Carlyle feared that the spread of technology would lead to the loss of an essential human character; a spirtual connection to the land and nature replaced with an mechanical, unnatural connection to artifice. Walker responds that freedom from biological shackles and toil allows us time to attend to our minds: the less labour we need to perform, the more “time left free for nobler things.”

“In the absolute perfection of machinery, were that attainable, we might realize the absolute perfection of mind. In other words, if machines could be so improved and multiplied, that all our corporeal necessities could be entirely gratified, without the intervention of human labor, there would be nothing to hinder all mankind from becoming philosophers, poets, and votaries of art. The whole time and thought of the whole human race could be given to inward culture, to spiritual advancement.”

Here Walker speaks of technology in quasi-religious tones; he sees technology as a means of ascension towards a higher state of being. In philosophy, there is the concept of the ‘sublime’ — the reverence, or awe, we feel in the presence of immense power or greatness — that is often linked to such religious feelings. In “The Machine in the Garden”, Marx claims that from the rapid pace of progress in the 19th century emerged the sense of the “technological sublime”: the transfer of these feelings of awe from religion or the natural landscape to technology. Trains, factories, and industry is the sublime power of humanity manifest in mechanism.

It is this feeling of the technological sublime, and mastery over nature, that is common to much of the art that embodies the ‘progress aesthetic’ — an optimistic view of technology as an improving force for humanity. I propose that art in this class has the following attributes:

• Technology as sublime

• Technology as romantic

• A juxtaposition of nature and technology; a harmonious cohabitation

• Orderly nature; nature brought into service of humanity. Mastery and control over the natural world

Below are several additional examples of this aesthetic. First, an engraving of Lombe’s mill (1750), the first water-powered silk throwing mill in England.

“Manchester from Kersal Moor” by William Wyld (1852). A watercolour commissioned by Queen Victoria of Manchester, then the largest producer of cotton textiles in the world.

Robert McCall’s “The Prologue and the Promise” (1983) commissioned by Disney’s EPCOT centre as part of the Horizon’s exhibit that aimed to inspire parkgoers with an optimistic vision of the future (now closed).

A Fortune Magazine cover from 1938.

Miyazaki's works also often have this character of ‘progress art.’ Yet, his films are not purely optimistic; they often deal with the tension between the ideal of technology as an enabling force, and its destructive power. Miyazaki is most romantic when it comes to flying machines: this shot is from the 2013 film “The Wind Rises.”

Progress entails tradeoffs and preferences, and meaningful progress is inherently humanistic. Carlyle was right in arguing that a future wholly separate from nature is undesireable; we are biological after all, and a human-centric concept of progress acknowledges that. Marx, in his book, is also interested in exploring the balancing act between the pastoral ideal and maximum technological progress; the “Machine in the Garden” trope sees control over nature as desirable, but not its elimination.

Along these lines, we generally find art in which technology has fully subsumed nature to be dystopian: for example, in the cities of Blade Runner (1982)

or Akira (1988).

But even in the absence of total assimilation, contemporary art often shows technology and the artefacts of civilization as being discordant with nature. Take Banksy, arguably the most publicly influential artist of the 2000s. In Banksy's world, technology, capitalism, and modern society are oppressive forces:

"We don't live in a world like Constable's Haywain anymore and, if you do, there is probably a travellers' camp on the other side of the hill. The real damage done to our environment is not done by graffiti writers and drunken teenagers, but by big business... exactly the people who put gold-framed pictures of landscapes on their walls and try to tell the rest of us how to behave."

This attitude is perhaps best displayed in his work “Show Me The Monet” (2005), where discarded shopping carts and a traffic cone litter Monet’s water-lilly pond.

Optimism about technological progress seems to be lacking in contemporary art. Perhaps it’s because most of our technological progress in the past generation has come in the form of information and computation — the internet, mobile phones, or AI — which doesn’t translate to visual mediums quite as well as trains or planes do. Yet, we still have modern technological marvels capable of evoking feelings of optimism and the sublime: rocket launches are arguably the modern equivalent of cathedrals.

So perhaps the problem is cultural; we’re lacking new, positive, visions of the future that put humans at the centre without disconnecting us from our natural origins. Perhaps it’s time for a 21st century version of “The Lackawanna Valley.”

Maximum Progress

Surgery Works Well Without The FDA

Here is a conversation from the comments of my last post on the FDA with fellow progress blogger Alex Telford that follows a pattern common to many of my conversations about the FDA: Alex: Most drugs that go into clinical trials (90%) are less effective or safe than existing options. If you release everything onto the market you'll get many times more dr…

Read more

2 years ago · 27 likes · 27 comments · Maxwell Tabarrok

I disagree:

1. I think the FDA is an important — albeit flawed — institution that should be reformed in a careful and incremental manner

2. I don’t think surgery is a good analogue for pharmaceuticals ex-FDA

Since Maxwell quoted me in the opening of his surgery post, I feel somewhat obligated to respond directly and justify my position. While I have some sympathy with the libertarian worldview, I’m here to argue that we’re not yet ready to tear down Chesterton’s fence.

For those who don’t want to read the full linked post, the structure of Maxwell’s argument, as I understand it, is as follows:

1. Both drugs and surgical procedures are complex medical interventions; it is just as hard for patients and regulators to assess the value of surgical procedures as it is for them to assess the value of pharmaceutical drugs

2. Unlike pharmaceuticals, surgery is not regulated by a central (federal government) agency like the US FDA

3. Despite a lack of centralised oversight, surgery functions well; surgery is constantly improving, and is not overrun by scams and ineffective procedures. At the very least, our surgical procedures do not outwardly appear to be of meaningfully lower quality than our drugs

4. Therefore, centralised government drug regulators like the FDA are not required to protect consumers from potential harms associated with complex medical interventions

5. Therefore, we should get rid of the FDA. This would increase pharmaceutical innovation by removing barriers to market without meaningfully increasing the risk to consumers

Before I address the comparison to surgery, and whether or not it’s valid, I want to note that there’s really no need to grasp for analogies like car manufacturing or surgery in the first place. We already know what an unregulated or loosely regulated pharmaceutical market looks like — and, reader, it’s bad.

Maxwell argues that in the absence of government oversight, market forces would prevent companies from pushing ineffective or harmful drugs simply to make a profit. Except that there are precedents for exactly this scenario occurring. Until they were stamped out by regulators in the early 20th century, patent medicine hucksters sold ineffective, and sometimes literally poisonous, nostrums to desperate patients. We still use “snake oil” today as shorthand from a scam product.

If patent medicines are too dated an example, we can instead look to the loosely regulated supplement industry as a proxy. Supplements often don’t contain what they claim to contain1, are frequently found to be ineffective in large randomised controlled trials, or are downright harmful to health. For otherwise healthy people without deficiencies, it seems likely that supplementation is mostly a waste of money2. Homeopathy is another example of a class of products with scant evidence to support their use; clearly people remain susceptible to bogus medical claims when they are willing to pay for treatments that are chemically indistinguishable from water.

The FDA did not spring into existence fully formed, its powers were granted and expanded over time in response to specific safety scandals and societal pressure (I cover these events in some detail in act 3 of my long post on the history of the pharmaceutical industry, thalidomide being the most notable). Anyone proposing we eliminate the FDA should keep in mind that we’ve already tried an unregulated pharmaceutical industry, and we decided that we didn’t much like it, thank you very much.

We can even point to recent examples of drugmakers pushing harmful drugs despite purported regulatory oversight. As Nick Laing notes in a comment on Maxwell’s piece, the opioid crisis is concrete, modern, example of how an overly permissive regulatory apparatus can lead to tragic outcomes:

“I won't get into the weeds of why treating medicine as a regular commodity in the market can lead to disaster, but the obvious example is the opioid crisis in the U.S.A. The market-friendly policies allowed heavy marketing of opioid medication both to patients and doctors (most countries in the world don't allow this) which, along with poor regulation of doctor prescription practises (and a bunch of other reasons) contributed to the U.S.A being one of the first high income countries in the world to have a reduction in life expectancy.”

The generic drug quality scandals documented in Katherine Eban’s book “Bottle of Lies” provide yet another example of how lax regulatory enforcement can incentivise drug companies to cut corners. Quoting a review of Eban’s book in NPR:

“In the U.S., FDA inspectors can show up to a plant at any time, incentivizing companies to always be in compliance lest they be caught off guard by an inspection. But the globalization of the pharmaceutical industry complicated this practice. To minimize diplomatic tensions, the FDA notified plants months in advance of visits. Such lead time gave overseas companies ample time to prepare for inspections. Eban describes accounts of companies hurriedly cleaning workspaces, shredding records and fabricating documents, even destroying visibly contaminated drugs.”

Anyone who wants to convincingly argue for disbanding the FDA should be able to show how we will avoid a repeat of these historical issues.

Now onto Maxwell’s actual argument re. surgery, against which I will advance three main counterpoints:

1. Surgery is regulated to a high degree. It is simply regulated differently to drugs (in a manner which indirectly involves the FDA, anyway)

2. It is generally more difficult to evaluate the risk-benefit tradeoff for drugs versus surgical procedures

3. In many cases, surgery doesn’t actually work all that well

Surgery is regulated

In the US, surgery is heavily regulated, albeit at a more local level than pharmaceuticals. The training process to become a practicing surgeon takes upwards of a decade (4 year undergrad, 4 year medical degree, 5 year surgical residency). State licensing boards can restrict surgeons from performing certain procedures, and take away their ability to practice if they do harm. Malpractice suits, guidelines (whether at the level of the hospital or professional associations), and peer training are other mechanisms that constrain a surgeon’s ability to act with impunity.

At the state and federal level, regulation is applied to procedures or surgical facilities (e.g. Joint Commission accreditation). Experimental surgeries generally require review and approval of a institutional review board (IRB) before they can be attempted. Much like in drug development, surgeons will test new procedures on surgical animal models before progressing to human experiments — the innovation process is controlled less by central diktat, but instead by the judgement of peers and ethics committees.

Now, one could argue that if the surgical profession can partially self-regulate, why not drugs? It’s plausible we could benefit from a variety of regulatory models adopted across institutions and states. For one, there is the practicality of interstate commerce; drugs can pass state borders more easily than surgeons. The other problem is one of scale. A great benefit of drugs is that they can be manufactured and distributed en-masse. But scale comes with risks, and as we saw with OxyContin, a harmful drug can do great damage when distributed widely. A single renegade “pill mill” can serve huge numbers of customers, whereas a rogue surgeon is limited in how much damage they can do.

And yet, the FDA is involved in regulating surgery (albeit indirectly). The FDA regulates the drugs used in surgery, like local anesthetics, anticoagulants, and antibiotics. The FDA also regulates devices from pacemakers to artificial heart valves to surgical robots (as Maxwell points out). Even stitches and scalpels fall under the jurisdiction of the FDA. Much of the progress in surgical outcomes has come from improved procedures enabled by innovative new FDA regulated devices, like the artificial heart valves used in transcatheter aortic valve replacement (TAVR).

For patients ineligible for open heart surgery, FDA-regulated devices used in the TAVR procedure (in the case, the Edwards Sapien heart-valve system) cut deaths in half versus the previous standard of care (medical management with or without balloon aortic valvuloplasty).

The modern surgeon has many tools at her disposal, most of which have been touched by the FDA on their path from invention to the operating theatre. I happen to think this point is by itself sufficient to invalidate most of Maxwell’s argument. Surgery is not this counterfactual regulation-free analog to pharmaceuticals; surgery is also heavily regulated and anyway so confounded by FDA involvement I don’t think we can conclude much of anything from the comparison. But let’s continue regardless.

Pharmaceuticals are (typically) harder to evaluate than surgery

I don’t intend to minimise the practical difficulty of surgery, but there is often a simple mechanical causal logic to why surgeries work. Reductively, surgical procedures:

1. Remove a bad thing (e.g. a cancer)

2. Replace or repair a damaged or broken thing

3. Disconnect things that shouldn’t be connected

4. Close a hole that shouldn’t be there

Or as the 16th century French surgeon Ambroise Paré put it:

"To eliminate that which is superfluous, restore that which has been dislocated, separate that which has been united, join that which has been divided and repair the defects of nature."

Drugs, by contrast, tend to have subtle effects; it can take years of clinical trials before the effects of a new drug can be statistically separated from noise. I’m optimistic our ability to make these inferences will improve eventually, but we just aren’t good enough yet at predicting when drugs will work to comfortably skip clinical trials. We often don't even know how our drugs work at all. As I said in a comment on one of Maxwell’s articles:

“Most drugs that go into clinical trials (~90%) are less effective or safe than existing options. If you release everything onto the market you'll get many times more drugs that are net toxic (biologically or financially) than the good drugs you'd get faster. You will almost surely do net harm, especially when you consider that companies will be less rigorous about selecting compounds for development than today.”

The low-level substrate that drugs act on in our body is so byzantine that we cannot (yet) disentangle the long and interwoven causal chains that drugs put into motion. Because surgery acts at a higher (macroscopic) hierarchy, much of this underlying complexity can be abstracted away; surgery is to classical mechanics as pharmacology is to quantum mechanics.

Consequently, the effects of surgical interventions tend to be more predictable, and often much more dramatic, than drugs. This relative simplicity is part of the reason why we’ve had effective surgical procedures for hundreds (even thousands) of years and effective drugs for only a century or so.

Even so, drugs as obviously beneficial as Kalydeco (ivacaftor), Spinraza (nusinersen), and Gleevec (imatinib) do come along from time to time — and when they do they are brought to market quickly. It only took two and a half years from Gleevec’s first clinical trial to its FDA approval for chronic myeloid leukemia (the FDA review took 10 weeks).

Surgery often doesn’t work all that well

In his post, Maxwell uses data on surgical inpatient and operative mortality to support his claim that surgery is improving over time. Even though I think data on iatrogenic harm is weak support for Maxwell’s claim (a lead-free sugar pill is hardly meaningful progress), it’s plainly obvious that we’re better at surgery now than in the 1950s. After all, back then we were still lobotomising people.

But have we really improved more at surgery than we have in pharmaceuticals? That’s unclear, and the strong form of Maxwell’s argument would have established that claim using comparative outcomes data between the two classes of interventions.

As far as I know, the dataset to conclusively establish the rate of improvement in the aggregate risk-benefit of surgery as compared to drugs doesn’t exist. We can, however, look to randomised clinical trials on surgical procedures to get a sense of the value of common surgical procedures. The benefits of excision of early-stage cancer, liver transplants, and trauma surgery are fairly unequivocal, but many other procedures fail to outperform placebos in randomised controlled trials. Orthopedic surgery in particular has a shaky evidence base: a review in the BMJ concluded that 7 out of 10 common procedures performed no better than non-operative management. And these are among the most common surgical procedures!

Consider also devices like the surgically implanted transvaginal mesh used to treat pelvic organ prolapse until they were banned by the FDA. These meshes were widely implanted, but frequently led to “mesh erosion, pain and pelvic infection” that necessitated another procedure to have them removed. Had the FDA not been involved in reviewing the evidence and issuing the ban, these harmful devices would probably still be used today.

In the end, all I can really say is that I remain unconvinced of the case for the wholesale destruction of the FDA.

Is the FDA perfect? No. Could it be better? As I’ve written before, I think the barriers to market could be lowered for diseases where clinical trials are infeasible or the unmet need is high. I also could be convinced of the value of parallel paths to market, like Scott Alexander suggests in his FDA post. It’s plausibly good to have diversity in review approaches that increase consumer choice. But, since the real gatekeepers to market access are often payers (and not the FDA) I’m skeptical these pathways would make much of a practical difference anyway.

To close out, I think it’s worth briefly considering the steelman argument for how drug regulators may help rather than hinder progress and innovation in pharmaceuticals. While it’s hard to establish causality here, I find the following arguments plausible:

1. The existence of the FDA is forcing function which incentivises drug companies to generate data that shows that their drugs have a positive risk-benefit profile. Information on what works and doesn’t work is fed back into the development process for future drugs, improving the quality of future drug development decisions

2. (as a corollary to #1) Because drug companies make money if their products are demonstrated to work, they are incentivised to pursue scientific research in a truth-seeking manner. Over time, market forces conspire to eliminate companies that do bad science and reward those who make discoveries that benefit humanity. By creating a mechanism for the market to reward good science, academic science (as a source of talent and ideas for biotech) is also indirectly incentivised to produce higher quality work

3. By creating a regulatory pathway which encodes some value judgements on what type of technologies are worth developing and in what manner, regulators can help funnel talent, money, research, and economic activity (incl. entrepreneurship) down channels that are ultimately beneficial for society

4. A defined regulatory pathway with established value generating milestones (IND, phase I, phase II, phase III…) makes investing in biotech more predictable and attractive3

I’m curious to hear counterarguments to these four speculative claims, or anything else I said in this post. Maxwell, back to you.

At its best, science is a self-correcting march toward greater knowledge for the betterment of humanity. At its worst, it is a self-interested pursuit of greater prestige at the cost of truth and rigor — Ed Yong

Scientific progress appears to be slowing, and scientific results seem to be less disruptive than in the past. A popular explanation for this trend is that young scientists (or newcomers to a field) must master an ever-growing body of knowledge before they are capable of generating their own new ideas. This ‘burden of knowledge’ theory views scientific fields as akin to oceans. Every day, as new discoveries are made, more and more knowledge is poured into the ocean, permanently raising the sea level. A scientific neophyte starts at the seabed and must swim their way upwards to reach the knowledge surface, whereupon they can finally contribute to advancing their field. 

But, this model of science as an ever-growing mass of facts is not how things really work. Scientific knowledge is more like a garden continually tended by a community of practitioners. Old facts and models are weeded out because they outlive their usefulness — or they are simply forgotten, lost in the overgrowth. A scientist doesn’t need to know everything about a field before they can contribute; it’s possible to tend the petunias without knowing about the lilacs on the other side of the garden.

Yet, something about science does seem to be getting harder with time. One possibility is that we’ve picked all the low-hanging fruit, and each new discovery requires a steadily growing investment of time and resources. But, what if the burden is not technological, but social? Perhaps it’s not knowledge that scholars must accumulate in ever increasing quantities, but prestige.

At least, that’s the hypothesis I came away with after reading Mark Khurana’s book “The Trajectory of Discovery”1. Khurana, a researcher at the University of Copenhagen, was motivated to write his book after a frustrating encounter with the biomedical research enterprise. As he recounts his inciting incident:

With a completely fresh set of eyes, research initially seemed to be a pure pursuit of truth. Over time, however, this sentiment slowly morphed from admiration to frustration, not toward individual researchers, but rather toward the system more broadly. It seemed to me that there was a vast amount of scientific potential that was going to waste, simply because the incentive structures underlying the academic community seemed to reward the wrong things.

“The Trajectory of Discovery” is an attempt to catalogue the distinct factors that slow down, or speed up, scientific discovery in the biomedical sciences. For example, a chapter on “Citations as Currency” covers how using citations as a measure of scientific value can incentivise scientists to pursue projects on the basis of expected popularity, not necessarily scientific value. All in all, Khurana covers 23 different factors including ‘p-hacking’, the streetlight effect, and outright fraud. The value of “The Trajectory of Discovery” is primarily as a reference work, and as a collection of (interesting) facts — the creation of a unified theory on what is slowing biomedical progress is left as an exercise to the reader.

So, here’s my take: I suspect that many of the individual challenges that Khurana mentions ultimately stem from an over-reliance on prestige as an axis of competition for limited resources.

Science is a social enterprise. In biomedical science, unlike say, maths2, results cannot practically be proven to be true in any definitive sense. What is thought to be true is then a distillation of group consensus among a relevant social collective of scientists. 

Before you can ‘do science’ and be accepted into the community, you must first signal potential to be a valuable contributor. In theory, anyone can experiment in their backyard and publish the results. In practice, the citizen scientist is a relic of a bygone age, replaced by the professional scientist class. The social collective decides who to let into the garden, so to speak.

So how does one get recognised by their peers? Ideally, the quality of a scientist could be evaluated by looking at their track record. But unfortunately, science is not like sprinting — the skill of science is hard, if not impossible, to quickly and objectively measure. The problem for new scientists is that scientific skill is mostly judged retrospectively; it takes decades to judge the quality of a scientific career and a body of work. Nor can we simply give everyone free reign and judge them later. Modern biomedical science is expensive, complicated, and time consuming; limited funding resources must be allocated. 

What is needed is a valid predictor: a way to distinguish good from bad scientists at the outset. This is essentially what prestige is, a signal of quality; prestige is a sort of currency or capital used to compete for limited resources. The artefacts of prestige include degrees from highly ranked universities, Nature/Cell/Science publications, NIH3 grants, and highly-cited papers. Khurana refers to this as "scientific capital": the collection of research experiences, publications, and relationships accumulated over a scientific career. The point of prestige, from the perspective of a funder, is to serve as a predictive signal of valuable future research output4.

If we think of a graph with two axes, scientific results on one and prestige on another, we can plot the main dimensions along which scientists compete. This dual-axis framing isn’t new, as Khurana notes:

Already during the 1970s, the French philosopher Pierre Bourdieu had considered how different actors in the academic world interact with each other. He believed that actors in academia, referred to as Homo Academicus, fought for legitimacy on two fronts. On the first front, referred to as the autonomous pole, academics fight for intellectual and scientific capital. On the second front, referred to as the heteronomous pole, academics fight for economic and political capital.

If prestige was a good predictor of the ability to do important science, we’d see a strong correlation between performance on both poles. Early prestige accumulation should predict future results — and we do see this in the data (although, there’s likely to be a ‘rich get richer’ effect at play). At its best, prestige is a form of evaluation via the wisdom of the crowds. But, as anyone who has been involved in science knows, prestige does not always accumulate where it should.

Invoking Goodheart’s law, there is a big potential problem with prestige as a predictor: if the returns to activities that generate prestige are greater than the returns to activities that generate scientific breakthroughs, scientists are incentivised to optimise for prestige instead. If a system becomes good at conferring and measuring prestige, it will disproportionately attract status-seekers who value prestige over the outcomes we’d prefer them to value (like scientific truth-seeking). Critically, status-seekers are unlikely to think they’re doing anything wrong — they are simply responding to incentives. It is not unreasonable to believe that work is valuable because it gets rewarded. As a result, incentives shape the values of people participating in the system.

Status-seeking behaviour is tolerable if there are enough resources to go around to keep everyone occupied; the status-seekers may take more than their fair share, but at least everyone else can eat. Given enough time, the best scientists would rise to the top anyway.

The problem with this, at least in biomedical fields, is that scientists are increasingly in competition for scarce resources. Khurana notes the severe and growing mismatch between number of scientists trained and the supply of full-time positions. Funding has not kept pace with the number of graduate scientists; the NIH’s budget has stagnated in real terms, following a funding boom and subsequent bust through the late 90’s early 2000’s.

As a consequence of this mismatch in the supply and demand for funding, modern academic biomedical research has become “hypercompetitive”. Khurana mentions several issues downstream of this increased competitiveness: a culture of ‘publish or perish’ leading to a proliferation of incremental or low-quality papers, exploitation of the graduate student workforce, burnout, and high attrition. 

Because prestige is an input into funding decisions, young researchers are being choked out by a system that has little room left to accommodate them. Science is ageing, driven by lengthening training periods and an exodus of young talent. The average age of scientists who receive their first significant NIH grant (an R01) is now in the 40s. In Khurana’s words:

With researchers spending ever more time writing grant applications, resources being concentrated within too narrow a band of individuals, and too few positions available for the supply of trained researchers, there is ‘a growing gulf between the aspirations of biomedical researchers, and the career prospects that are open to them.’

With no track record, early career scientists must start off by accumulating prestige from their degrees, who they work with, and what they publish. They can then ‘cash in’ this prestige for the resources required to do scientific work (grant funding, lab space, support staff, etc.). In the past, the ‘prestige price’ to enter into the scientific community was relatively low. The system breaks when the bar to entry — the amount of prestige that must be accumulated to be allowed to do work — rises to the point that new scientists must spend more time on prestige accumulation than anything else. This results in a pipeline where prestige is used to accumulate more prestige, and the true purpose of science is lost in the search for ever-greater prestige.

A core problem with incentivising scientists to value prestige is that it can be accumulated by validating the status quo, since prestige is awarded by the now-eminent scientists who established that status quo in the first place. At its worst, this dynamic can incentivise scientists to stifle competing ideas and even commit fraud — a prominent example of this being the “cabal” of scientists who stymied research into non-amyloid-beta causes of Alzheimer’s for decades.

A less overtly harmful, but perhaps more insidious effect of overvaluing prestige is missing out on valuable novel ideas. Because genuinely novel ideas are harder to judge in the social validation system of science, they are inherently low status (e.g. Katalin Karikó as an ‘anti-prestige’ role model). Young researchers in particular are a source of novel ideas, and their exodus ossifies science.

Khurana dedicates several chapters to this idea that novel ideas are discounted by the scientific establishment. As evidence he cites the tendency for interdisciplinary work to be published in lower impact journals, and the “pivot penalty” experienced by researchers who shift focus areas. However, it is in the grant-making process where the novelty discount is most impactful. Major funders, including the NIH, are notoriously conservative, and often prefer to fund higher-status work with more predictable outcomes. Funder risk-aversion is plausibly linked to resource availability — by one measure of novelty, the NIH reduced funding for novel ideas when its budget was cut in real terms (around the year 2000).

                            Graph from M. Packalena and J. Bhattacharyab

The novelty discount is just one example of how the funding system is at the heart of the issues caused by an over-reliance on prestige-based evaluations. In Khurana’s words:

The problem with the structure of grant funding processes and the subsequent creation of different classes of researchers is that we are lulled into believing that the system is purely meritocratic when it in fact is not. We have chosen a narrow set of criteria to define previous successes, and by extension to define merit. But these criteria inevitably lead to a stratified research environment where future research ideas are funneled through a sieve of preexisting eminence, meaning that the imbalance in the grant process skews toward prior prestige more so than the quality of the idea.

Of course it’s easy to say there are problems with using prestige to help allocate funding, but what’s the alternative? The answer isn’t trying to stamp out prestige-based evaluation in science — even if we could eliminate prestige, whatever we replaced it with would become the new prestige.

Personally, I’m partial to solutions that increase the diversity of funding sources for scientific research. The biomedical research funding ecosystem is dominated by a small number of risk-averse mega-grantmakers like the NIH. This concentration arguably gives funders too much power, which is part of the reason why researchers need to spend half of their time writing grants instead of getting on with research.

Compare this to the start-up funding ecosystem, which is a much more distributed market; if a founder strikes out with a particular angel or venture capitalist, there are plenty more who might be receptive to the founder’s vision. Because venture capitalists need to compete for access to the best deals, they are also plausibly much more accommodating to entrepreneurs than scientific funding institutions are to the scientists they support.

The benefit of encouraging greater diversity in biomedical research funding sources — whether by founding new focused research organisations (FROs), or breaking up the NIH — is that it engenders a diversity of funding criteria. Because different organisations are likely to value different things when making funding decisions, it’s harder to game the system. The more diverse the decision makers, the more diverse the portfolio of funded research — and hopefully, the lower the burden of prestige accumulation.

Thanks to Jake Parker for giving feedback on a draft of this piece

There's pleasure in having our expectations subverted, at least that's what some leading theories of humour suggest. By “bribing the brain with pleasure”1 evolution encourages us to root out errors in our models of the world. Funny things are surprising, surprise is synonymous with information, and gathering information about the world helps us make better predictions about it and thereby improve our survival chances. Curiosity and the novelty-seeking drive are similar enticements to seek information; when your base needs are satisfied it makes sense to invest excess energy in learning more about the world. In some ways, it feels good to be wrong.

If (resolving) ignorance is bliss, can we think of depression as a pathological overconfidence that everything sucks? Major depressive disorder is characterized by a loss of pleasure, fatigue, a tendency to pessimistic interpretations, and negative internally-directed thought loops (rumination). One proposed explanation for these traits is that depression is an adaptive form of risk aversion — or learned helplessness. If you believe that the world is fundamentally unpleasant and risky, it makes some sort of sense to disengage with the outside world in favour of ruminating in bed. The association between illness, inflammation, and depression could be interpreted along similar lines as a means to get people to stay home and recover until they’ve fought off an infection.

The loss of explorative joy in depression is in some ways like being at the bottom of a local minima in gradient descent. If you're confident that your model of the world is accurate, then there's no reason to incentivize further exploration. If the local minima you're stuck in happens to make you believe that the world only has net negative outcomes to offer, then you might as well just give up. Yet most people with depression do eventually find a way to get unstuck. One model of how they manage to do this could be analogous to a technique in machine learning that's applied when gradient descent algorithms are trapped in local minima called simulated annealing — injecting randomness and shaking things up so that other, potentially better, minima can be explored in future iterations.

One of the long-standing mysteries about depression is why our treatments work. Psychotherapy, selective serotonin reuptake inhibitors (SSRIs) or serotonin–noradrenaline reuptake inhibitors (SNRIs) are the mainstay of treatment. But there’s a whole panoply of drugs available to treat depression with diverse mechanisms including lithium, ketamine, noradrenaline–dopamine reuptake inhibitors, monoamine oxidase inhibitors, and more besides. The classical psychedelics, in particular psilocybin, have been getting a lot of attention as of late too. Even crudely shocking the brain with electroconvulsive therapy or disrupting it with transcranial magnetic stimulation or deep brain stimulation is remarkably effective.

It's not clear why such a diverse range of treatments should be effective when you apply a neurochemical or receptor-level lens to depression. Our most widely used pharmacological treatments are (were?) thought to work by modulating serotonin receptors and serotonin uptake, but there's no strong evidence that serotonin itself has any effect on depressive mood. Lithium has been used since 1948 and we still don't understand its mechanism. Psychedelics and electroconvulsive therapy are similarly poorly understood. However, if we think of depression as a pathological local minima perhaps what all these treatments have in common is some sort of generally disruptive effect — an injection of novelty — that acts like simulated annealing. Since major depression is associated with impaired neuronal plasticity, temporarily improving plasticity could be one such means to facilitate annealing into a new more positively valenced state2.

Note: I wrote this short essay over a year ago but didn’t publish it because it was written quickly and felt overly speculative. But, as I was going back over old drafts I found it again and figured I’ll finally post it — if it’s misguided someone will correct me and we can all learn something. Let me know if you enjoyed this type of post, or if you didn’t!

All data is the result of some selection process. The first order manifestation of this is bias: a failure to select a representative sample of the real underlying thing the data is describing. But selection operates at multiple hierarchies. For something to exist such that data can be collected about it, that something must have been created instead of some other thing — it must have survived a selection process. Because of this, any dataset includes latent information about the traits that tend to survive these selection processes.

For example, if you counted up all the members of every species on Earth the resulting dataset would give you some information on the types of creatures that tend to do well in Earth's biosphere. If you then counted how many of those creatures had fins, and how many had wings, you could get partway towards an understanding of how to construct an optimal ‘model’ animal. A mix-and-match book made real.

Zooming down to a lower level, Karl Friston (and others) have argued that organisms are embodied statistical models of the states of their environment1. This is a way of thinking about an organism as "a hypothesis of its environment"; in the words of Peter Munz:

"The behavior of a fish and the functioning of a theory of water are exactly identical. The fish represents water by its structure and its functioning. Both features define an initial condition (for example, the degree of viscosity of water) which, when spotted or sensed, trigger off a prognosis or behavioral response which, in case of a fish, fails to be falsified. By contrast, a bird does not represent water"

In other words, the states that organisms are most likely to be in are the ones they are likely to prefer2. You find organisms where they choose to be.

This may just seem like another spin on the familiar process of evolution by natural selection. But I think the selection-eye-view reveals some deeper truth about what is learned by models when they’re trained on real-world datasets. Machine learning models which learn compressed representations of data also naturally end up learning something about the optimal properties of (and selection process acting on) what they are modelling — even if they aren't provided information about those properties from the outset.

Meta's protein folding model ESMfold illustrates this principle. ESMfold achieved state-of-the-art accuracy on protein structure prediction, even though its language model was trained on protein sequence data alone: the training process involved feeding the model protein sequences with gaps in them, and getting the model to predict amino acids to 'fill in the blanks'3.

Nevertheless, the information about the shape of protein folds was latent in the sequence data. In order to learn better and better predictions to fill the missing amino acids, ESMfold's language model had to create an internal representation of the structure of the protein and how it folded. This internal language model representation could then be read out by a folding module (which was trained on sequence data) to generate accurate 3D protein structure predictions.

Protein folding is a non-random process, and there are 'best practices' in how to structure a given protein sequence to achieve a desired structural motif. Evolution has learned these optima and uses them repeatedly, and because they appear frequently, they can be learned. More generally, the set of proteins that exist in the dataset were only able to be sequenced in the first place because they are evolutionary successful. So there is also latent information about the types of sequences and structures that are biologically optimal in the protein sequence dataset.

This principle of latent optimality was exploited by researchers who used Meta’s ESM protein models to improve the properties of monoclonal antibodies that were already highly optimised4. The researchers took antibodies already in clinical use — like Regeneron's COVID antibody — and ran their sequences through the ESM model, tasking the model with predicting where the true amino acid sequence of the antibodies differed most from the model’s predictions. Then, the researchers selectively replaced the amino acids at those divergent sites with the model’s predictions. These substitutions markedly improved the binding affinity, thermostability, and in vitro potency of the antibodies, without a noticeable increase in off-target binding — all without feeding ESM any information but the protein sequence.

This suggests that because samples in datasets have themselves undergone a selection process, as long as we are confident that the selection process is producing outcomes we care about (which we can be reasonably certain of in biology), we can use the statistical properties of those datasets as a sort of optimisation function in themselves. Exploration followed by AI-informed selective regression, or annealing, may be a valuable general approach to optimise new designs generally5.

One of the challenges of applying machine learning in medicine has historically been that it's not clear what optimisation function to use. I've previously argued that a way around that may be to collect all sorts of data on parameters of biological systems — a detailed description of states — and then use that as an objective function. The evidence from the antibody optimisation experiments using ESM are perhaps a small piece of support in favour of that approach. Biological systems have already undergone billions of years of selection and optimisation — we may as well make use of that latent information.

Thanks to Adam Green and David Yang for giving me the idea for this post

Super Mario 64 was one of my favourite games as a child — I still remember racing the penguin down the icy slide in ‘Cool, Cool, Mountain’, or plucking the star from the tail of a giant eel in ‘Jolly Roger Bay’. These whimsical challenges contribute to the game’s enduring appeal: nearly three decades on from its release, a community of ‘speedrunners’ are still actively competing to see who can collect all the game’s 120 stars1 the fastest. 

Speedrunners contest world records by having either the best execution of the fastest known route, or by discovering and exploiting newer, faster, routes. In September, the world record 120-star run stood at 1 hour, 37 minutes, and 35 seconds — compared to the 20 or so hours it takes your average player. In his world record run, the speedrunner Weegee plays with sustained machine-like precision; it’s hard to believe that anyone could shave off the extra 36 seconds needed to break through the 1:37 barrier. By now, the standard 120-star route is so thoroughly optimised that the time advantage from most new routes is measured in fractions of a second. It seemed there was little left to discover in this 27-year-old-game.

That is, until September, when a new ‘holy grail’ trick was discovered that cuts about a minute off the 120 star run time. What I found fascinating about this discovery is not the specifics of the trick per se, but the process by which it was found. Even if you’re not interested in video games, or Super Mario 64 speedrunning, if you bear with me I think you’ll find the story interesting too.

One of the longest sections of the 120-star run occurs in a level called ‘Rainbow Ride’, in which Mario is supposed to ride a slow magic carpet along a rainbow track to get a star atop ‘The big house in the sky’. The big house in question (more like a castle) is shown below, and you can browse a view of the whole level here if you’re so inclined.

This magic carpet ride takes about 2 minutes — painfully slow for a speedrun2. For decades, the speedrunning community thought that it might be possible to skip the carpet ride and jump straight to the top of the house. This would be the so-called ‘carpetless’ strategy. While there had been tantalisingly close attempts at getting up in the past, nobody had been able to find a way up that was actually feasible for a human to reliably execute. At least, not until the programmer Krithalith released a video of a jump technique he discovered and named ‘Orthogonal Jones’

Orthogonal Jones involves a precise sequence of jumps starting from the window on the left side of the house up onto the slanted green roof. Krithalith didn’t find this sequence himself, after all, no human has found a way of getting up in decades of attempts. Orthogonal Jones was found by an algorithm called ‘scattershot’.

Scattershot is a sort of empirical machine gun; it uses brute force search with random inputs to find the shortest path to collect a star from a given starting point. By simulating the potential paths of thousands, or millions, of Marios, scattershot can discover new optimal routes through levels3. 

This isn’t a video of scattershot or Super Mario 64, but it’s a nice illustration of the principle.

However, random exploration by itself is unlikely to converge on the best solution. The problem with random search methods is that the longer a sequence needs to be, the greater the processing power and time required to explore all the options; random search is inefficient for long input sequences. Krithalith got around this by incorporating learning into the algorithm, and saving shorter optimal move sequences that can be used as staging points for random exploration4. These short sequences can then be chained together to find the best overall long sequence.

As Krithilith developed scattershot, he benchmarked it against established routes and was able to shave off a few fractions of a second here and there. But in September 2023, scattershot was ready to take on carpetless. As the YouTuber Karl Jobst recounts in his video about the discovery:

“In September of 2023 Krithalith would test scattershot on Rainbow Ride and see if he could solve carpetless. He placed Mario on the windowsill and left scattershot running. Within an hour it had already found its way to the roof. An hour doesn't seem like a long time, but [scattershot is] so fast that 1 hour of scattershot playing Super Mario 64 is equivalent to over 100 years of real time.”

It took scattershot under an hour to discover what no human was able to discover in decades.

Even though Super Mario 64 is a constructed system, built by humans, we don’t fully understand what is possible within the system. The ‘rules’ of Mario’s world are known and programmed, but there are emergent interactions between these rules that mean some behaviour is unpredictable from first principles. After 27 years of exploration, we are still discovering new things that are feasible in the system.

The story of carpetless demonstrates that brute empiricism may be the only feasible route to discovering the possibilities present in even simple systems. Many particulars of a system may be compatible with the same underlying principles, and finding the particulars that actually exists requires someone (or something) to go and probe them.

This is a relevant lesson for science in general. In science, as opposed to a video game, we don’t know the underlying rules from the outset — natural systems are governed by ‘hidden rules.’ There are two complementary ways of knowing in science that help uncover these hidden rules. The first is building an intuitive or theoretical model of the shape of a system. Recalling the joke about the spherical cow, these models are useful approximations. They may smooth over certain details, but they provide a general sense of how the systems work, where its boundaries lie, and what may be feasible within the constraints of the system. The other way of knowing is trial-and-error (empirical) testing. This engages with the spiky reality to probe systems experimentally and find where truth diverges from smooth models.

Intuitive or theoretical models are needed to understand the general principles that describe a system, and empirical testing is needed to validate theories and find places where theory diverges from reality. This entails a kind of trade-off between theoretical and empirical understanding; making models of systems is a type of information compression, but empirical testing is a sort of anti-compression that brings details back in.

In the case of carpetless, the speedrunning community had enough of an intuitive understanding of the world of Super Mario 64 built up that they had a hypothesis that it was possible to get up to the roof of the house by jumping from the windowsill, they just weren’t sure how to concretely achieve it. They did know roughly where to point the empirical machine gun, however, so this combination of theory and empiricism played out to discover Orthogonal Jones. 

This is analogous to what happens in science, where ‘scientific taste’ guides experimental directions. The sort of ‘tool-assisted’ route finding that was employed to solve carpetless may also be a useful case study for those looking to automate aspects of scientific research (e.g. Future House). A productive general strategy for scientific automation could be to combine tools for rapid, semi-random exploration with people or models able to direct that exploration to useful areas.

Of course, reality is not as easy to probe as a video game like Super Mario 64. It lacks the ability to save and resume from the save, the ability to emulate the whole system in full and speed it up, and it lacks clear optimisable goals. But what if we found ways to make reality more video-game like?

At first, it seems there’s no obvious analogy to ‘collecting a star’ in the scientific process. However, if we consider that the goal of science is to build ever better theoretical models of systems, then there is a potential way to identify functions that can be optimised. Systems can take on different states, and the set of possible states and state transitions is defined by underlying rules. Science is the process of finding those rules, but only the states and the paths between them are visible to the scientist. We can say that a scientist understands a system when they know all the possible states and how to perturb the system to move between them. With this framing, the analogy to finding an optimal star-collecting route is finding the shortest sequence of perturbations that takes a system from some specified state A to a desired state B. Having a perfect model of the system is equivalent to having a function that spits out the shortest path between any two states.

In practice, this means that a way to approach automating science could be to collect voluminous statistics on system states (deeply characterising both healthy and unhealthy model systems, say) then using scattershot-like guided randomness to find interventions that lead to the shortest path between those states. This approach creates its own function that can then be optimised. If you’re trying to understand a disease, finding the quickest path to reversing that disease and restoring a healthy state tells you a great deal about which factors are most important in the disease process.

Video game-like environments are also potentially valuable to test scientific automation systems. Some natural processes proceed too slowly to use as a test bed for scientific automation processes. Instead, we could create simulated versions of natural phenomena that are well-understood which can be run much faster than real time and loaded with defined states. If scientific automation tools prove capable of uncovering the hidden rules of these simulations, then they could be unleashed on real-world problems. This would greatly speed up the design-test-refine cycle compared to only testing scientific automation techniques ‘in the wild.’

In Super Mario 64, at least, Krithalith’s discovery is already leading to concrete progress. In October, the speedrunner Karin used Orthogonal Jones to set a new world record 120-star time, the first to break the 1 hour 37 minute barrier.

Yet what struck me about the article was not so much the content, but the response. I noticed in the reaction to that article a sort of glee in failure, or schadenfreude, among some of the biotech commentariat. It’s not the first time I’ve noticed this type of reaction. There is often a quickness to take pleasure in the downfall of people or companies associated with bold, even hubristic claims — particularly if they were relative newcomers to the industry.

“Haha, yes! Finally they know our pain.” says the grizzled industry veteran. “They were such fools to think it would be easy; biology is our god — and she hates us!”

When 90% of drugs fail in clinical trials, it’s no wonder the industry has developed an allergic reaction to hype. Bluff and bluster dissolves in the acid empiricism of the randomised clinical trial: “Show me the data,” says the industry vet.

Although AI is a rich vein of hype, it’s far from the only source. Take EQRx, a (now defunct) biotech with the stated ambition of reducing drug costs by undercutting expensive brands with cheap, fast-to-market copycats. If EQRx were successful, they’d have been the Lidl or Kirkland signature of biotech — with ‘own brand’ versions of blockbusters like Keytruda and Ibrance. As the CEO Alexis Borisy said in a launch interview:

“At EQRx we are trying to rethink, reimagine, reengineer the whole way that we make medicines from creating them, proving they work, and getting them to patients… our ambitions are large, we want to have our first drug on the market within five years. We'd like to have 10 on the market within 10 years. By 15 years we want to have literally dozens and dozens of these great innovative new medicines at a radically more affordable price on the market.”

But even copying established drugs is no trivial feat, and EQRx imploded before bringing even one drug to market. The pricing and regulatory hypothesis behind the business was flawed, yes, and in many ways EQRx’s failure was entirely predictable: there were no shortage of post-failure “I told you so’s.”

Personally, I’m glad EQRx tried. Innovation isn’t helpful if it’s unaffordable, and the current drug pricing trajectory is not sustainable long-term. The answer is not to throw up our hands and say “that’s just how it is,” and that EQRx were foolish to think otherwise.

On the one hand, scepticism is often warranted. Biotech is complicated and biological data are often ambiguous, which makes it all too easy to obfuscate in order to take advantage of desperate patients and unsophisticated investors. Those in the industry should be calling out bad actors.

On the other hand, it’s almost tautological to say that biotech is a difficult industry. Every drug development program has warts, and it’s easy to predict failure; if you simply predict any given drug will fail you’ll be right about 90% of the time. Pessimism sounds smart, even if it’s not.

Anyone who has been around in biotech long enough has accumulated plenty of scar tissue. In Derek Lowe’s words, “the default setting for people in the pharmaceutical industry is a sort of enlightened pessimism.” Risk aversion is endemic in the industry, and a career can be built on saying ‘no.’ The fibrotic effect of experience can be tough to resist, but someone needs to be willing to say ‘yes’ if we are to make progress.

It takes champions to drive any drug through development. It takes people who are unreasonably optimistic about their likelihood of success to found industry-defining companies. Often, such people are prone to hyperbole: Bob Swanson, the co-founder of Genentech along with Herb Boyer, painted a suitably grand vision in their 1976 business plan:

“With Genentech’s technology, microorganisms could be engineered to produce needed protein to meet the world food needs or to produce antibodies to fight viral infection. Any product produced by a living organism is eventually within the company’s reach.”

Genentech fell short of this original vision, of course. Yet, along the way it still managed to kick-start the biotech industry, develop the first recombinant insulin, and pioneer therapeutic monoclonal antibodies.

We should welcome people with a vision to do things better in our industry. And while bold visions often seem naïve to veterans, naïveté can be a blessing. Herb Boyer agrees:

“I always maintain that the best attribute we had was our naïveté. . . . I think if we had known about all the problems we were going to encounter, we would have thought twice about starting. . . . Naïveté was the extra added ingredient in biotechnology.”

If we keep believing drug development is near-impossible, it might just stay that way. And while the latest suite of AI tools seems likely to be an incremental advance, stacking incremental advances is how we’ve made extreme cumulative gains over many years.

The industry needs optimists to push it forwards. In the words of Karl Popper:

“Optimism is a duty. The future is open. It is not predetermined. No one can predict it, except by chance. We all contribute to determining it by what we do. We are all equally responsible for its success.”

So yes, let’s be critical of hype and hubris – another Theranos helps nobody. But let’s also be careful not to dissuade optimism. Drug development is unreasonably hard, so it needs unreasonably optimistic people.

At face value, my experience seems at odds with the narrative of the UK as a science superpower. The UK excels at producing students, researchers, and research output: UK scientists publish 6% of scientific papers (#3 in the world), its degree programs are first-class, and the UK biobank and NHS supply deep wells of quality biomedical data. The UK leads Europe in both discovering new drugs and founding biotech start-ups (although it lags far behind the US).

At the time, leaving the UK felt like a personal failure. In retrospect, it worked out well; I was more competitive abroad, where the supply of life science graduates is better matched to demand. The UK directly employs only 70,000 people in its pharmaceuticals sector, versus 900,000 in the US. Contrast this number with the 630,000 students of medicine or biological sciences at UK universities. This oversupply suppresses wages and leads to misallocation of talent into sectors that don’t suit their skills or interests: only 16% of UK biological science graduates are employed in “highly skilled” STEM jobs 6 months after graduating. 

Reductively, the lack of biotech jobs stems from the lack of big domestic employers. Alongside the legacy pharmaceutical companies AstraZeneca and GSK (est. 1913 and 1873, respectively), I know of only one other independent commercial-stage drug developer headquartered in the UK: Immunocore. 

These ‘missing’ employers are ultimately a reflection of the UK’s poor performance when it comes to commercialising its ideas: there’s a “valley of death” between research and commercialisation in the UK that many firms never overcome. 

The valley of death is often attributed to a culture of risk-aversion among UK investors, who are unwilling to provide the capital needed to scale up innovative companies. As a result, start-ups stagnate, sell out, or go seek scale-up capital elsewhere, like Immunocore did with its 2021 Nasdaq IPO. As Immunocore’s chairman put it:

“Immunocore is, I fear, a classic example of what the UK has been losing. We have a massive upfront commitment to the sciences in the UK, then all the commercial benefits go to investors elsewhere”

In theory, it doesn’t matter that companies list in the US. Investors, and individuals, in the UK can simply own the US equity. Capital is mobile, after all. However, in practice they don’t: only 10% of UK household wealth is in stocks versus 38% in the US. Hardly anyone in the UK benefits economically from the eventual commercialisation of drugs discovered in the country.

It’s often proposed that the UK should address this issue by incentivising local biotech clusters and domestic IPOs. Yet, at least when it comes to biotech, this effort may be futile. 

For one, the US pharmaceutical market dwarfs the UK’s: £521bn in 2022, compared to £27bn. The US effectively subsidises the world’s biopharmaceutical industry, while the UK constrains theirs through price controls. Consequently, the talent pool for drug commercialisation is far deeper in the US than in the UK. The powerful gravitational pull of the US market draws in early-stage companies keen to access its management talent and investor networks.

The value of a biotech company is primarily in its intangible intellectual property, so they are only weakly tied to physical locations. Highly mobile assets, capital, and talent exacerbates these winner-take-all accumulation effects. Even if a biotech is founded in the UK, there’s no need for it to stay and continue to support the local fish and chip shop.

Instead of trying to fight gravity, what if the UK played into its strengths and accepted its role as an exporter of talent and ideas?

This requires a shift to thinking about talent and ideas as a valuable economic product and export in and of themselves. Then, instead of seeing the outward flow of talent and ideas as a strict negative for the economy, the UK could set up a ‘sovereign innovation fund’ intended to capture and reinvest some of that value, comparable to sovereign oil funds in Norway, Saudi Arabia, and Alaska.

In practice, the fund could support intellectual property development by making seed investments in start-ups in strategically important areas, like life sciences or artificial intelligence, in exchange for a small equity share. It might also require UK university tech transfer offices give a portion of their ownership share to the fund, as a form of tax.

For its talent mandate, the fund should subsidise education and training in important sectors — and drive individuals towards economically valuable degrees. Emigration to access higher wages should also be encouraged by the fund, perhaps by offering income share agreements in exchange for relocation assistance. According to payscale, the average annual salary for a UK biomedical sciences PhD is £34k. In the US, the average is £87k — 2.5x higher. A UK life science PhD holder effectively gets a £50k raise to move to the US. Why not capture some of that surplus?

Fund profits would feed back into the domestic economy in the form of investments, subsidies, or even direct payments like the $1,600 average annual payout of the Alaskan oil fund. Appreciation in the value of the fund would finance further talent and idea development, like an economic flywheel.

For this model to work, the UK will need to remain an attractive location for talent to train and be trained. Brexit has impeded talent mobility, no doubt, but it also presents an opportunity to reshape the UK’s immigration strategy. Strong ties will need to be forged and permeable borders maintained with strategic destinations for talent, who should be receptive to importing the UK’s highly-skilled labour in a world with increasing dependency ratios.

Ingenuity and human capital has long been a strength of the UK, but without a change this advantage is at risk. The challenges faced by the UK commercial biotech industry mirror the factors that have led to the UK’s broader economic malaise: chronic underinvestment in business R&D, and poor distribution of capital and the spoils of its inventions and economy. Wages are stagnating, labour productivity growth is weak, and the cost of living is rising. Instead of trying to lock in talent and ideas, the UK may benefit from releasing them and embracing its role as the world's workshop for talent and inventions — with a fair cut of the proceeds. 

Personally, leaving the UK was the best decision I’ve made for my career and economic prosperity. I’d like others to have the same opportunity I had; and if the UK gets to benefit, that’d be nice too.

This is my entry for the TxP progress prize, a blogging contest that encourages responses to the question “Britain is stuck. How can we get it moving again?”

We tend to think of health in discrete terms. Either you’re healthy, or you’re sick. Similarly, we put diseases into neat-seeming categories: cancer, arthritis, epilepsy, hypertension, psoriasis, and so on. Yet, our categorical frameworks and language — how we think and talk about disease — are mental shortcuts that may actually be holding us back.

We create categories because we find them useful for decision making. Minds, like metal dies, make sense of reality’s raw extrudate by shaping it into definite forms. Yet this sensemaking process can distort the underlying truth. When it comes to our health, there is an arbitrariness to how we categorise disease that perhaps says more about the evolution of our medical systems and specialties than it does about our biology.

Medical specialties as we know them today emerged in the early-to-mid 1800s1. The impetus to master complex high-volume procedures, improve outcomes, and better compete for patients supplied the initial incentives to specialise. Ophthalmology, for instance, was an early breakaway; the first eye hospital, Moorfields, was founded in London in 18052. By centralising patients, eye hospitals enabled early specialists to gain sufficient volume of practice to reliably perform the intricate, yet effective, surgery required to remove cataracts and foreign objects in the eye3. 

In many cases, operational pragmatism drove the functional specialisation of medical institutions. The spread of ‘lunatic asylums’, intended to sequester the mentally ill away from the general sick population, facilitated the relatively early emergence of psychiatry. Funding for specialised institutions, like Germany’s Bethel epilepsy centre, often came from philanthropic donors with the explicit aim of treating underserved patient populations. It was natural for doctors working at such functionally specialised institutions to adopt a specialty. 

Post-revolutionary France was the first country to adopt a broad system of medical specialisation, starting from the 1840s. As the centre of the Enlightenment, Parisian institutions were predisposed to allow reductionist ideas to first creep into medical practice and take hold. In his history of medical specialisation4, Dr. George Rosen suggested that:

“Specialisation was not a consequence of the accumulation of knowledge, but rather of a new conception of disease: it was specifically the influence of localist pathological thinking, based on pathological anatomy and subsequently on new technologies like the ophthalmoscope and laryngoscope, that created ‘foci of interest’ in organ systems around which specialties could develop.”

Specialisation provided a means to distinguish oneself in the Parisian institutional world, which, at the time, placed great emphasis on the advancement of scientific knowledge. By specialising, physicians increased their chance of remaining at the forefront of knowledge within their chosen domain.

Although Paris may have been the nucleus, specialisation spread rapidly outwards: first to Vienna, then to other European cities, and, by the 1860s, to North America. While growth was slow at first, the proportion of medical specialists exploded at the turn of the 1900s. In 1845, about 10% of Parisian physicians identified as specialists (excluding surgeons); by 1905, that share had hit 35% in Paris, and 30% in Berlin¹. This is not to say there was no resistance to the increasing specialisation of medicine and the rise of ‘special hospitals’. Britain, in particular, was slow to specialise at scale; its leading doctors expressed concerns over wasteful duplication of resources and feared that the loss of a broader perspective would harm patients5. As the surgeon Sir Benjamin Brodie put it in 1860:

“Diseases generally are so connected with each other, and a knowledge of one is so necessary to a right understanding of another, that no one who limits his attention to any given disease, can be so competent to investigate its nature, and to improve its method of treating it, as those are who have a wider field of observation, and who are better acquainted with general pathology.”

Yet, specialisation won. Today, two-thirds of doctors in the OECD countries are specialists. The long list of medical specialties is an eclectic collection of domains that are organ-based (hepatology, nephrology, neurology…), disease or symptom-based (oncology, allergy, pain medicine…), and procedure or technique-based (anaesthesiology, pathology, radiology…). Nobody would design such a Byzantine6 system from first principles. Rather, the specialties developed organically from contingencies grounded in the practical realities of medical practice, less so the biological realities.

Molecular biology, pathophysiology, and pharmacology grew up in medicine’s shadow, so naturally they inherited its concepts. This inherited framing has trickled down to how we think about diseases and their treatments more broadly. Consider psoriasis, uveitis, Crohn’s disease, and rheumatoid arthritis. Biologically, these are all complex inflammatory autoimmune diseases which share similar causes and treatments; for instance, all can be effectively treated by a class of drugs that block the pro-inflammatory TNF protein. Despite their similar biology, these conditions are managed by different specialties. In another world, they might all be managed by an inflammation specialist who sees them as manifestations of the same underlying TNF-dysregulation disease.

While current categorisations of disease are not invalid — they are, after all, functional — questioning the inherent assumptions within them may help lay the groundwork for a new way of conceptualising, and treating, illness.

When we create categories we implicitly assume that meaningful boundaries can be drawn between illness and health, and between diseases. Take the debate over prediabetes, which the American Diabetes Association defines as an A1C of 5.7% to 6.4% (where A1C is a measure of blood sugar levels). This threshold is arbitrary — other agencies like the WHO use different thresholds which shift the estimated prevalence by millions7. While pre-diabetics are at heightened risk of progressing to diabetes, the risk is low, on the order of a few percent per year8. Regression to normal blood sugar levels is about as common as progression to diabetes.

This challenge of drawing thresholds exists within diseases too, even among diseases with relatively well-defined genetics. Mutations in the sodium channel protein SCN1A lead to a spectrum of epilepsies that range from relatively mild and self-resolving to the frequent, disabling seizures of Dravet syndrome. Yet, the phenotype of Dravet syndrome is itself highly variable. There is no clean solution to divide up the SCN1A epilepsies; any subdivision contains flaws. Age, response to prior treatment, and seizure type variability further stymie classification. While seizure severity is correlated with the magnitude of structural mutations in SCN1A9, classification on the basis of genetics is no easy solution. Family members with the same genetic mutation can have widely differing expressions — one child may be disabled by daily seizures, while the other is spared10. 

Similar issues are seen in psychiatry too: there is substantial genetic overlap and shared risk factors among attention deficit hyperactivity disorder, bipolar disorder, major depressive disorder, and schizophrenia11. What separates someone who suffers from one psychiatric disease from another, or determines whether someone is considered mentally healthy or unwell, might be a particular quirk of their environment or life trajectory — for example the season in which they were born, or whether they grew up in an urban environment12.

We break diseases into categories to help us understand and treat them, but we must not forget that a category is not real — it’s a label. Although categories, thresholds, and boundaries may be artificial, they have real-world consequences. Categories can be useful abstractions. But it’s worth asking: useful for whom?

When you must allocate limited resources and decide who can and should visit the hospital, it makes sense to draw a distinction between healthy and sick. When you must route people to the appropriate specialty, prescribe drugs with a specific label, and assign an ICD code to bill them, it makes sense to argue over the terminology for their specific affliction.

When motivations beyond the accurate representation of underlying biology creep into the categorisation process, distortions are introduced. To see why, we can think of categorising a disease as akin to the mathematical technique of ‘dimensionality reduction’. This is the process of projecting a multi-dimensional object or dataset onto a lower dimensional mapping — like how light thrown onto a three-dimensional shadow puppet casts a two-dimensional shadow. Once data has been projected onto a lower dimension, you can then draw boundaries around the individual shapes that emerge, and give them separate names.

Throwing away dimensions is a useful trick because it allows us to focus on the variables (dimensions) that we deem most important, while retaining as much information as possible. If you’ve reduced the dimensions cleverly, then the lower-dimensional representation should still look much like the true shape. Imagine twisting a shadow puppet around in a stable beam of light: the puppet’s three-dimensional shape doesn’t change, but its shadow does. The more the puppet is twisted away from its intended orientation, the less faithfully its shadow represents its higher-dimensional form. 

One of the risks that comes with dimensionality reduction is loss of relevant information and distortions of the truth. Depending on how you collapse the dimensions, separate entities can look like the same thing: think of two shadow puppets, one obscured behind the other. Performing a dimensionality reduction means making trade-offs; some information must be discarded. When we created our taxonomy of disease we cast off some information to make room for human constructs and conveniences — like specialties — which necessarily leaves less space to better represent the underlying biology. Biology has been called “the science of exceptions”, but it only seems this way because nature likes to make shapes that don’t fit neatly into low-dimensional boxes.

The modern concept of precision medicine — tailoring therapeutic interventions to an individual’s particular biology or characteristics — is an attempt to address the problem of over-broad disease categories by adding back missing dimensions. Even so, our current approach is trapped within the gravity well of pre-existing constructs. When you talk in terms like ‘a precision medicine based approach to psoriasis’, you are already operating within the existing framework. Enlightenment Paris casts a long shadow. 

What if we were to build precision medicine from scratch? How might we unravel the fan of dimensionality as far as possible, and approximate nature as she really is?

Imagine for a moment you were in some idealised world without computational or instrumental limits. First, you might map all the dimensions along which our biology could vary. Then, having done that, you could plot the values associated with good health on this multidimensional chart. 

This is a way of thinking of diseases not as specific entities, but as mappings to regions in a continuous ‘health landscape’ or ‘phase space’. 

Discrete ways of thinking beget discrete solutions. Like how you might repair a car: diagnose the problem, apply a fix, or replace the part. This is no way to cure a complex, adaptive, systemic condition like cancer — certain types with simple molecular drivers, yes —  but others must be gradually coaxed out of existence. 

People are not a collection of individual symptoms, diseases, or parts. We are dynamic, adaptive systems where genetics, risk factors, comorbidities, and disease processes intermingle into a holistic state unique to each individual. Thinking about health and disease as a continuum engenders a way of thinking that is fluid and systemic — if you lower the mental walls that categories erect, it seems easier to move between them. 

Dynamism is more natural when we think in continuums instead of rigid categories. Within such a framework, we can conceptualise a disease risk factor as a force pushing an organism’s state away from a healthy position — like a boat adrift on a calm sea being pushed out into stormy waters. Certain genetic mutations (like the CFTR mutations that cause cystic fibrosis) would be particularly strong directional currents. This implies the current state of the system is not sufficient to describe its health, we also need the rate and direction at which it changes13. It’s easy to forget about path-dependence when we see the world as static slices.

To treat, or prevent, disease we would have to apply a countervailing force with similar magnitude to push the system back to a desirable position. The further from health, the harder it is to get back. The greater, or longer, the force that would need to be applied. Intervening in the system changes it, so we would have to continually measure and adjust the force and direction of intervention. Drugs are one way of applying this countervailing force, as is exercise, nutrition, and other interventions. 

We could even take a homeostatic perspective and think of health not as a region of the landscape, but as a measure of system responsiveness; the ability to resist forces that might push a system out of equilibrium. Some have made similar suggestions that ageing is akin to a loss of system ‘springiness’. The net effect of living being the slow accumulation of drift.

What I like about this view of disease is that it shows how much more room remains for biotech innovation. All the drugs the biopharmaceutical industry has produced until now account for only a tiny fraction of the ways we could perturb our biology; each a single transformation from point A to point B on an infinite landscape. 

In the future, your annual (or more frequent) medical check-up might track thousands, or millions of parameters. Your doctor could estimate the rate and direction of drift from one check-up to the other, and prescribe a tailored mix of interventions (drugs, exercise, nutrition, etc.) to counteract the net forces pushing away from health — perhaps with the help of a predictive artificial intelligence (AI) model. We wouldn’t necessarily find it useful to assign disease labels in such a world, instead we might rely on exquisitely detailed models that represent our health states as mathematical objects.

To get to this future, we’ll need to overcome the practical limits to this ever more precise description. Every cut we take to define a population more precisely shrinks the size of that population. If we don’t change our approach, there will eventually be an economic limit to precision medicine. Taken to its extreme, we cannot run eight billion n-of-1 trials. To overcome this practical and epistemological barrier we’ll need to train AI models to predict how our health interventions will modulate out-of-sample states, and then validate those predictions. We’ll need to find a balance between high-dimensional representations that capture as much biology as possible, while remaining tractable to measure.

This vision will take decades to achieve in full, of course, but that doesn’t mean we can’t start laying the groundwork. We can engineer better imaging and measurement instruments to more precisely capture and measure biological states, and critically, the rate and direction of how they are changing. We can start testing our expanding library of approved drugs — alone and in combination — to see how they affect biological systems along as many dimensions as we can measure, and characterise their ‘intervention vectors’. These perturbation maps could be built up in ever greater detail: starting with cells, to tissue and organs, then full bodies. We can work with regulators to further explore disease-agnostic indications, and develop and validate new AI-discovered biomarkers for clinical trials.

Perhaps most of all, we’ll need to be open to new ways of reasoning about biology and disease: if we hope to cure all disease, we may eventually have to abandon the concept.