“We don’t know yet. We truly don’t”.

This sentence sounds scientifically correct but may also be the most dangerous sentence you can say. Of course there is nothing shameful about those words - they are practically the national anthem of experimental science. Biology is complicated, experiments fail, assays misbehave, mice possess an almost malicious ability to develop unexpected phenotypes, and supposedly identical batches occasionally behave as though one spent the weekend reading Nietzsche.

The problem begins when scientific uncertainty becomes an excuse for managerial uncertainty. Ask when an important program will reach a decision point and too many scientific organizations answer with an explanation of why dates are difficult. Ask what happens if the experiment fails and the answer is another experiment. Ask when the company will know whether to stop spending money and suddenly everyone becomes extremely interested in discussing the fascinating biology. This is intellectually respectable and commercially lethal.

Sure, biotech is not the same as car manufacturing and needs to be managed differently. Drug development contains genuine uncertainty. A biotech CEO who demands that researchers simply “commit to the date” can’t bully a molecule into having better pharmacokinetics. Excel remains unmoved by this injustice. Payroll arrives on schedule. CRO invoices arrive with almost supernatural punctuality. The lease does not contain a clause saying that rent may be postponed because the experiment result was inconclusive. Your mice are still hungry.

This is where many science-led companies make a subtle conceptual error. They think investors have financed research, while investors think they financed a company that will create a product and generate a nice exit, with research being simple mechanics of the business, not much differentiated from software engineering. Yes, investors understand that there is risk involved. They often manage it by sitting in the ivory tower and looking at your company as a black box, where risk level is just a coefficient they use to calculate their portfolio return. They treat risk just like they treat your development and regulatory milestones. Your company’s value goes up as you pass these milestones, which makes your number look prettier in Excel and increases their chances of cashing out. With 20 to 30 companies in their portfolio, it’s hard for them to be interested in your experiments being delayed because your mouse sneezed.

At the end of the day, investors talk to the CEO and the CEO is a manager first, scientist maybe second, if ever, and managers deliver profits.

So what do you do? Here are some thoughts…

Stop managing experiments. Start managing uncertainties.

Most project plans in research organizations are lists of activities labeled as “the strategic plan”. Develop assay. Screen compounds. Characterize hits. Run PK. Perform tox study. Select candidate. Take over the world. Some might be useful operational mechanics, but they conceal the central managerial question: “Which uncertainty is each activity supposed to eliminate?”

A better development plan might start with an uncertainty register. List the assumptions that must be true for the product to work commercially and rank them by consequence and current confidence: target biology, translatability, efficacy, toxicity, exposure, manufacturability, formulation, regulatory path, patient selection, reimbursement and competitive differentiation. Then attach evidence to each assumption.

This approaches results in a different scientific narrative and managerial frame. Instead of reporting that “the team completed seven experiments this quarter”, the CSO can report that the program entered the quarter with three existential uncertainties, eliminated one, materially reduced another and discovered that the third is worse than expected. This is progress even if the answer is bad. In fact, discovering cheaply and early that a program is doomed may be some of the highest-return research a biotech ever performs. A negative result that prevents a $30M mistake is not a failed experiment, but an unusually profitable exercise, even if the marketing side of it sounds terrible.

Put decision gates before the data exist.

Scientists often decide what a result means after seeing it. Scientifically, interpretation genuinely requires context. Managerially, however, allowing every result to negotiate its own definition of success creates a spectacular survival mechanism for mediocre programs. Before running an expensive or strategically important experiment, write down the possible outcomes and what the company will do in each case. If exposure exceeds X and efficacy exceeds Y, advance. If exposure exceeds X but efficacy falls below Y, stop. If the assay fails predefined quality criteria, repeat. If the result lands in an intermediate zone, conduct one specified resolving experiment.

Once you document your outcomes, question your assumptions, decisions and action items. Did you make the right assumptions? Should they exist? Are your decisions and action items driven by risk and progress or you being scared of failure? Impartial view of results is more likely to deliver good action items.

This is essentially a pre-mortem for data interpretation. It prevents the organization from discovering, after an ugly result, that the ugly result is actually “quite encouraging when viewed in the broader mechanistic context”. Every veteran of R&D has attended a meeting where a program arrived as a corpse just to be resurrected with another six months of experiments. Biotech portfolio reviews can certainly make you believe in reincarnation.

The principle is not that thresholds should never change. New evidence can rationally change them. The discipline is that changing a threshold requires explaining what new information justified the change. Otherwise a development criterion is merely a motivational poster fighting for survival.

Replace single dates with probability-weighted timelines.

Everyone has been there: management asks for a date, science responds that a date can’t be known. Both might be asking the wrong question. What both of them need is understanding of risk and possibilities, which might be as simple as a distribution of dates with simple explanations, and risk factors assigned to them, along with decision points.

For every major milestone, estimate three cases: an aggressive case in which nature gives you a gift and important experiments work largely as expected, a base case incorporating ordinary iteration, and a downside case involving plausible failures. Then assign rough probabilities and, crucially, identify what moves the program from one branch to another. “IND-enabling studies completed in Q3” sounds reassuring but conceals risk. “There is roughly a 60% planning probability of Q3, 30% of Q4 if formulation requires another iteration, and 10% of a longer delay if the current toxicity signal survives the repeat study” tells management something it can actually use.

The percentages will be wrong, but that’s fine. Weather forecasts are also wrong individually, yet nobody concludes that meteorology should therefore communicate exclusively through “I don’t know’s”. Managers want probabilities, not astrology. They want to understand the assumptions that will help them with portfolio planning and reveal when the organization is systematically overconfident. Track forecasts against outcomes for a year and teams can begin calibrating themselves. A group that described twelve consecutive events as “90% likely” and achieved six of them doesn’t have unusually bad luck, but rather a newly-invented numerical system.

Give every experiment an economic job.

Before approving an experiment, ask five questions: What uncertainty does this reduce? What decision will change because of the answer? What is the cheapest experiment capable of changing that decision? How much capital becomes committed after a positive result? What happens after a negative one?

This prevents what might be called scientific feature creep. Researchers understandably want comprehensive understanding, but often not at the level their companies can afford. Sometimes a beautiful mechanistic experiment adds knowledge without changing any decision. That may still be worth doing, but somebody should be able to explain why. Otherwise the company is buying scientific art work.

Value of information is an especially important concept here. Suppose a $150,000 experiment has a reasonable chance of exposing a problem before the company commits $5M to the next development stage. Its value is not merely the scientific information produced. Part of its value is the capital it prevents the company from wasting. Conversely, a $500,000 experiment that produces an elegant publication but would not alter the program regardless of its outcome may have surprisingly little managerial value. This sounds barbaric to researchers until one remembers that companies eventually run out of money and stop paying salaries, whereas unanswered scientific questions have demonstrated impressive financial stamina.

Manage the critical path and the uncertainty path separately.

Traditional project management identifies the critical path, the sequence of dependent activities that determines the earliest completion date. Scientific organizations need another map beside it: the uncertainty path. This identifies the unresolved assumptions most capable of making the entire schedule irrelevant.

Imagine a program in which manufacturing takes six months, toxicology takes four and a decisive translational experiment takes six weeks. A conventional schedule may focus enormous managerial attention on shaving three weeks from manufacturing because it sits on the critical path. But if the six-week translational study has a 40% chance of killing the program, running it three months earlier could be vastly more valuable than accelerating manufacturing. The fastest route through the project is less important than discovering whether the destination exists.

This is a particularly powerful principle for platform companies, where teams can spend years optimizing components downstream of assumptions that remain weakly tested upstream. Don’t merely ask, “What is blocking the timeline?” Ask, “What could make this entire timeline worthless, and how quickly can we test it?”

Budget uncertainty, not just departments.

Most biotech budgets say how much the company is willing to spend on chemistry, biology, clinical operations and headcount. A more revealing budget also asks how much the company is willing to spend before resolving each major uncertainty. This creates capital-at-risk limits.

Management might decide, for example, that no more than $2M will be committed to scale-up before a particular efficacy threshold is demonstrated, or that the company will not initiate a second expensive program until the first has established evidence supporting a shared platform assumption. These rules are imperfect and may force a scientist outcry, but they force work into sequences. Without them, organizations have a tendency to start downstream work “to save time”, which is an excellent strategy when everything succeeds but is also an excellent way to convert scientific failure into expensive scientific failure when it does not.

Parallelization therefore deserves more suspicion than it receives. Running ten things simultaneously looks fast on a Gantt chart. It can also mean spending money on nine things whose necessity depends on the tenth. While parallelization can sometimes look rational if the value of time exceeds the expected cost of wasted work, it can also be a corporate anxiety expressed through purchase orders. Calculate the difference.

Give program a kill budget.

Biotech celebrates program initiation and mourns program termination, which creates exactly the wrong incentives. A program that refuses to die consumes people, capital and managerial attention that could be allocated elsewhere. Companies should therefore create explicit kill criteria and kill budgets. At the beginning of a development stage, define not merely how much money is available to succeed but how much may be spent determining whether the program deserves another stage.

This changes failure from an organizational embarrassment into an expected portfolio outcome. And if failure is structurally normal, then discovering failure efficiently is a managerial capability. A company that kills weak programs rapidly can be better at R&D than one that keeps everything alive. Hospitals have intensive-care units. Biotechs sometimes have pipelines.

Measure decision velocity, not experimental effort.

Executives love dashboards, partly because a collection of colored circles can make almost any situation feel supervised. Scientific dashboards commonly measure experiments completed, compounds screened, samples processed and milestones achieved. Add a more consequential metric: time to important decision.

How long does it take the organization to move from identifying a critical uncertainty to generating enough evidence to act on it? How many experiments are performed between decision points? How many major decisions are delayed because nobody specified the evidence required? How often are programs continued despite crossing previously agreed stopping criteria? These measures reveal organizational quality in a way that raw experimental throughput does not.

A lab can be extraordinarily busy while the company goes nowhere. Hamsters have demonstrated this business model conclusively.

Separate scientific truth from managerial action.

Perhaps the hardest lesson for scientists entering company leadership is that a decision rarely requires complete knowledge. Science asks what is true. Management asks what should be done given what is currently known. The FDA itself makes benefit-risk decisions while acknowledging uncertainties arising from incomplete evidence. A biotech company must do the same long before it reaches a regulator.

This means managers should distinguish several categories that scientists often blur together: known facts, probability estimates, wishful thinking, and decisions. “The molecule caused liver toxicity at this dose” may be an observation. “There is a 50% chance this will prevent a therapeutic window in humans” is an inference. “We will terminate the program” is a decision. Putting all three into one PowerPoint bullet produces arguments in which scientists attack the decision by disputing the probability while executives defend the probability because they have already made the decision. Separating them makes disagreement useful.

It also permits a much healthier phrase than “we need more data”: “What decision-relevant uncertainty would the additional data resolve?” Sometimes the answer will justify another experiment immediately. Sometimes the room will become quiet. Silence is one of the cheapest assays available.

The scientist-manager is a different profession.

None of this requires scientists to become miniature investment bankers who say “value creation” every five minutes and develop an inexplicable affection for quarter-zips. It requires recognizing that scientific leadership inside a company is different from scientific excellence alone. The scientist-manager must translate uncertainty into choices, choices into resource allocation and evidence into decisions. That person still respects complexity, but does not permit complexity to become an exemption from accountability.

Investors do care about science because science determines whether biotechnology works. What they ultimately require, however, is something different from interesting research. They need uncertainty to be converted into value before the cash runs out. Customers are even less sentimental. Patients don’t buy elegant mechanisms. Hospitals don’t reimburse fascinating hypotheses. A commercially successful biotech must eventually produce something that works, can be manufactured, can obtain regulatory approval, can reach patients and can generate sufficient economic value to sustain the enterprise.

So when somebody asks when the program will deliver and the scientific answer begins with, “But you have to understand how uncertain this is,” resist the temptation. Everyone should understand the uncertainty. Then manage it.

Don’t promise certainty where none exists. Do something much harder: define the uncertainties, rank them, price them, design experiments to destroy them, establish decisions before results arrive, allocate capital according to what remains unknown and continuously update the plan as evidence changes.

Science can’t tell you exactly what will happen, but management still requires you to decide what happens next.

I have news! Or maybe a revelation…

The next scientific revolution won’t be powered by AI. Yes, AI is helpful, and it will perhaps find the next 100 drugs and mint another trillionaire, but it won’t be as easy for it to address the biggest bottleneck of progress - humans.

Humans are highly biased information super-compressors, driven by incentives and possessing brains designed for survival rather than processing vast amounts of information. Humans live in social groups with diverse biases and even bigger social groups we call societies and cultures, each one introducing thousands of random inputs with diverse priorities, incentives, and biases of their own.

A scientist working on a project may budget 1000 hours on true scientific work, but he might spend 900 of those hours getting funding and boss’s approval, fighting in-house politics and government regulations, getting upset with the spouse’s to-do list, presenting to committees, updating resume, drinking coffee, going to conferences, picking up kids, complaining about senior management, and then watching the president pull NIH funding or his company changing direction after the investor meeting, who of course care about returns more than lab rats. The remaining hours might look like the art work of a mad scientist, except they are full of biases, assumptions, misunderstanding, and confusion, driven by the incentive to publish or produce positive data so that they can get a bonus or investors can give them enough money to last a few paychecks.

While progress certainly occurs within these walls, environments like this also produce noise, a highly biased and compressed one. Then other scientists peer review it under the same conditions, then others cite it and use it in their own research. Then it turns into a game of a broken phone, culminating in journalists proudly announcing that scientists “solved cancer”, while the original work should have probably been documented as “Assuming target X has anything to do with cancer at all, which we have no proof of, we’ve observed that every once in a while this random molecule docked against the target, while the scientist was doing a voodoo dance he can’t remember, or at least didn’t document well enough to be reproducible.”

This is rarely intentional of course. Humans don’t generally compress information or lie on purpose. They are victims of their brain design nonetheless, simply unable to process all the inputs and contemplations affecting their environments. They are creative at the same time, so their brains try to solve puzzles, and invented solutions or hypothesis quickly turn into perceived reality. But humans don’t generally see the environment around them as that of their own design. They focus on their own micro tasks and immediate incentives, blaming the rest of the world on bad bosses, politicians, spouses, and traffic.

What’s left of this mess is progress, often questionable, sometimes subjective, and randomly not reproducible or requiring a voodoo dance. But it’s progress nonetheless and we are proud of it.

If one was to present this on a figure, it would look like the figure below: a vector of effort, offset by a vector of pushback designed by human flaws, resulting in a vector of progress. Although there is no way to effectively measure this, one might suspect that this vector is progress is fairly small.

Solving humans is difficult but major improvements are possible. While the root cause lies in our biological design, our brains are highly hackable, as we’ve proved through our education system, management, and marketing. It would of course be fun to start by putting humans in futuristic chambers with alien amoeba screening their brains and turning humans into error-prone zombies. But I would start with the education system instead.

The current education system has a great positioning in people’s minds - it shapes humans over the years of brainwashing while humanity believes that it’s mostly good. That’s exactly where shaping should begin. Today the education system focuses more on educating humans how to use someone else’s ideas developed inside the box than on thinking critically, with minimum bias and assumptions, without coffee interruptions, and from first principles. Humans also attend schools to learn how to become members of a diverse society of other humans, whose environment and incentives are aligned more with Netflix in the evening and blaming greedy billionaires for their difficult lives, than with scientific progress of the entire world. It might be a useful skill overall but it still comes with a major flaw - humans behave the same way their society behaves, and society behaves the way humans do, which may be a circular reference and a self fulfilling prophecy. At the end, very few problems are solved but enemies are identified and Netflix announces superb Q3 earnings.

Our education system wasn’t designed for the outcomes we are talking about. It was more focused on enabling millions of people to read and do basic math, so they can enter the workforce which, at the time of the design, was more focused on operating factories and voting. The knowledge workers came in a century later, but the system is still optimized for factory workers and learning effectiveness is now just a statistic, a highly subjective one and not easily translatable to current societal needs. Humans learn to say “I did a great job” and “I’ll do my best”, which is often a general, subjective signal of effort without accountability.

That’s exactly where we should start. An educational revolution, the one that rethinks our needs from first principals and redesigns the entire system without the bias of grades, teachers, and politicians, is the one that will create a true long term impact. A generation better equipped with knowing how to question assumptions, think outside the box, not give in to biases, express emotional intelligence, and understand other humans is more likely to grow into a society with a bigger vector of progress.

It will happen, today or tomorrow, with or without AI. Until then, question your assumptions and biases. Evaluate every word you are saying to see if the concept you are using is true or if it’s a simplified human term that compresses reality into something more comfortable.

Comfort is not a mother of innovation.

Forget AI for a moment. The company pharma needs to study should be the one that didn’t build the model. Not because models are bad, but because the obsession with AI made pharma close its eyes on the gold mine that lays in their existing data, hidden under layers of bureaucracy and protected by incentives to survive rather than innovate.

When I first came into biotech six years ago, I started pulling my hair out and asking a lot of “why” questions. I saw an industry whose “not invented here” thinking outlived the dot-com boom and fall. Some of its information technology and red tape could only compete with those of US federal government and people were equally content with the state of affairs - its employees were often difficult to differentiate from zombies. Evidently, self-driving cars on the streets of San Francisco and the tech revolution didn’t impress enough executives to switch from Windows to web-based ELNs. Benchling “innovated” the industry and finally offered a web-based SAAS product, as Salesforce Tower smiled while celebrating the 20th anniversary of turning on premise software into fossils.

Bureaucracy survived but raised an AI flag to signal that its brand stands for something fresh. The tech industry felt its gravitational pull. They instantly built models. Then benchmarks. Then better benchmarks. More capable reasoning. Generative chemistry. Protein prediction. Autonomous laboratories.

And now, every few months we see another demonstration suggesting that some previously stubborn intellectual task can be performed faster by a machine. This is definitely important and quiet useful. It may also encourage pharma companies to focus on the part of the AI stack where they have the least durable advantage. Frontier models are becoming extraordinarily capable, but they are also becoming widely available and commoditization might be around the corner.

This is great, nonetheless. But the harder problem inside a big pharma is considerably less glamorous: getting the right data, from the right systems, under the right permissions, into the right workflow, at precisely the moment somebody has to make a consequential decision.

That is why pharma executives should spend more time studying Palantir. AI vendors like Anthropic and OpenAI became famous by making intelligence dramatically more accessible. Palantir built its reputation around making complicated organizations computationally legible. Pharma needs plenty of the former. Its chronic disease, however, is often the latter.

Consider what actually happens between a scientific insight and a medicine reaching a patient. Discovery data live in one environment, preclinical results in another, clinical information in several more. Manufacturing has its own systems. Quality has its own systems. Regulatory teams assemble evidence according to exacting requirements. Commercial organizations build forecasts from still another collection of datasets. Supply-chain teams worry about capacity, inventory and geography. Patient information is governed by strict access controls. External data arrive from CROs, laboratories, hospitals and other partners. Each function may be individually sophisticated while the organization connecting them remains surprisingly dependent on spreadsheets, reconciliations, meetings and heroic employees who know which database contains the number everyone else is arguing about. This complexity is sometimes orchestrated by hundreds of people dancing between back-to-back meetings, where confused faces try to assemble puzzle pieces together, hoping that gatekeepers will actually unleash some of these pieces before they retire. Data integration project timelines look suspiciously similar to those of clinical trials, and project managers do voodoo dances similar to those of lab scientists - cute and usually not easily reproducible. Meanwhile, thousands of boxes of paper records from previous decades of clinical trials rot somewhere in the warehouse, waiting for the new generation of zombies to scan them and put behind the red tape.

A more intelligent AI chatbot does not automatically solve this problem. Neither does a better molecular model. What matters is creating a usable representation of the enterprise: what a batch is, what a clinical site is, what a patient cohort is, what inventory exists, which trial depends on which supply, who is authorized to see which information, what actions are permitted and what happens when circumstances change. Palantir calls the organizing layer at the heart of its approach an ontology. The terminology matters less than the idea. Data become substantially more valuable when they are mapped to the objects, relationships, rules and actions through which an organization actually operates.

There are certainly glimpses of this approach and some success in biotech and pharma. You can now get faster access to validated clinical data and shorten the path from trial planning toward submission. Palantir has also described healthcare and biotech customers using its product to harmonize disparate datasets and make clinical and real-world evidence more accessible. Unfortunately, none of this proves that Palantir has discovered the universal operating system for drug development. Vendor case studies deserve the same healthy skepticism as exceptionally attractive before-and-after photographs. But they illustrate a useful principle: the economic value of AI often appears only after information can travel through an organization quickly enough to change what somebody does.

This matters because pharma’s instinct is often to turn AI into a portfolio of use cases. One team builds a medical-writing assistant. Another experiments with protocol generation. Scientists acquire a research copilot. Commercial teams automate content. Manufacturing gets predictive maintenance. Procurement gets another dashboard. Each project can produce a respectable return while leaving the company fundamentally unchanged. The result is a technologically impressive archipelago: dozens of clever islands separated by oceans of organizational friction.

The Palantir lesson is to begin with the decision rather than the model. Which decisions materially affect development speed, probability of success, manufacturing reliability or patient access? What information is required to make them? Where does that information currently live? How fresh is it? Who may access it? What dependencies connect one decision to another? Which actions can safely be automated, and which require accountable human approval? Only after answering those questions does the model become interesting. In this architecture, an LLM is a powerful component rather than the architecture itself.

That distinction becomes especially important as foundation models improve. If model capability continues to diffuse among competing providers, pharmaceutical companies are unlikely to derive a lasting advantage merely from having access to excellent models. Their competitors will have them too. OpenAI’s expansion into life sciences makes the point rather neatly. In April 2026 Novo Nordisk announced a strategic partnership with OpenAI intended to apply advanced AI from drug discovery through commercial operations, including manufacturing, supply chain and distribution. Thermo Fisher has similarly described embedding OpenAI APIs across areas including product development, service delivery and operational efficiency. OpenAI, in other words, is itself moving beyond the chatbot.

The competitive question therefore shifts. If everyone can rent intelligence, what remains scarce? Proprietary scientific data certainly do. So do experimental capabilities, clinical-development expertise and biological judgment. But another scarce asset is the institutional machinery that allows intelligence to encounter proprietary data safely and then affect reality. A model that can predict a supply disruption is useful. A system that knows which products are affected, which patients depend on them, which manufacturing constraints apply, which alternative facilities have capacity, which regulatory restrictions matter and which executive has authority to intervene is something else entirely. Prediction creates information. Operational integration creates leverage. Speed of information availability and automatic representation of interlinked data puts you at a different level against competition.

Pharma has an unusually strong reason to care about this distinction because its decisions occur inside a dense web of regulation, validation, privacy and accountability. The industry’s caution about autonomous AI is therefore not merely conservatism. A system that recommends an action must often provide enough provenance, permissions and auditability for somebody to understand why that action was possible in the first place. Enterprise AI in pharmaceuticals consequently has to answer questions that consumer AI can often ignore: Which data was the system allowed to use? Which version? Who approved the workflow? What happened afterward? Can the decision be reconstructed six months later? Intelligence without governance is a demonstration. Intelligence with governance can become infrastructure.

This also suggests that the industry’s obsession with AI talent may be slightly misallocated. Pharma certainly needs machine-learning scientists, computational biologists and excellent software engineers. But some of the most valuable people in an AI transformation may be translators who understand both the company’s data architecture and the physical or regulatory process represented by that data. Palantir’s much-discussed model of embedding engineers close to customers reflects a recognition that enterprise software rarely succeeds by being dropped elegantly from the cloud. Somebody has to understand why the manufacturing planner distrusts the inventory field, why two clinical systems disagree about site status, and why the apparently irrational approval step exists because a regulator once asked an awkward question in 2017.

There is a also a somewhat awkward implication for senior management. If this argument is right, many supposed AI problems are actually organizational problems exposed by AI. A company discovers that its data are fragmented, its definitions are inconsistent, its permissions are Byzantine, its workflows are poorly documented, and each piece of data has a red tape attached to it. The tempting response is to wait for a sufficiently clever model to reason through the mess. This feels like responding to a chaotic warehouse protected by blind security guards by hiring a more intelligent forklift driver. Intelligence helps. Knowing where anything is helps more. Educating gatekeepers is gold.

None of this diminishes the significance of frontier models. OpenAI and its peers may contribute enormously to drug discovery, scientific reasoning, clinical development and knowledge work. Better models could unlock capabilities that are difficult to anticipate today. The mistake would be to assume that model intelligence automatically translates into enterprise productivity. Between an answer and an outcome lies an organization.

The companies that capture the most value from AI may therefore be those that become unusually good at connecting three layers: models that can reason, proprietary context that makes the reasoning relevant, and operational systems that allow the reasoning to change a decision. The first layer will attract most of the headlines because intelligence makes for spectacular demonstrations. The second and third layers involve data plumbing, permissions, ontology design, validation and workflow engineering, which is why nobody has yet made a blockbuster film about master-data management.

Pharma should certainly watch the AI space. It should experiment with the best models available and exploit every legitimate improvement in scientific reasoning. But it should study Palantir for a different reason. Palantir’s central insight is that enterprise intelligence becomes valuable when it is connected to the machinery of the enterprise itself.

The next pharmaceutical AI breakthrough may indeed be a model that understands biology better than any scientist. But the next billion dollars of AI value may come from something much more boring: a company finally knowing what it knows, knowing who is allowed to use it, and getting that knowledge to the person who has to decide what happens next.

Donald Trump may be less psychologically interesting than the people who think they are nothing like him.

Love him or hate him. Whether you consider him erratic, genius, or both, he represents an interesting lesson in science. No, I am not wondering how chromosomes could have created something the entire world is scratching its head about, and I am not wondering about his policy or cutting NIH funding in the name of pseudoscience. I am wondering whether the rest of the population really looks that much different from him.

Trump’s statements about bombing Oman and Haitians eating dogs in Ohio may sound absurd, but they are not based on clinical insanity. To put it simply, they might just be a representation of a regular human brain and its propensity to simplify, apply bias, and act - a survival mechanism reinforced by years of evolution and an occasional dopamine rush. Whether this specific specimen with his collection of randomness is a good fit for presidency can be a separate topic for discussion, but the fact that millions of peopled voted for him demonstrates a bigger pattern - all humans simplify, apply bias, and act, and the difference between people may lie in the book of biases they managed to collect throughout their lives.

This is where we get to science. Of course everyone going through the PhD program learns about bias and eliminating it in research. But that’s where this bias elimination exercise often ends and people publish research that can’t be reproduced or takes us into the wrong rabbit hole. My recent discussion of a paper with a colleague is a case in point - a well-written paper with lots of data and claims of scientific breakthroughs, except it followed traditional human logic consisting of flaws, biases, and wishful thinking, with publishing being a higher incentive than revealing nature’s secrets. And it’s not flaws that I found curious, but my colleague’s reaction to my findings - yes, this type of logic is a state of scientific writing - it’s basically the norm. In other words, don’t expect papers to publish the truth. Focus on the signal instead - methods, observations, and interpretations. Simplify, apply bias, act. Get a Nobel prize and let someone else deal with reproducibility. Humans know they are biased and flawed, so they invented peer review, controls, statistics, and replication, which remarkably push our flaws towards progress.

This exercise reminded me of a biotech CEO who pitched us a few years ago. After she presented a deck with claims of taking over the world in a year, I asked her for data to demonstrate her claims, to which I got a Trump-like answer: “I have a PhD from Stanford. I know what I am doing.” Evidently, somewhere in the evolution cycle humans created a new type of bias: Stanford PhD = God level knowledge.

While we can wonder about true capabilities of this startup CEO or humans who wrote illogical papers, we should also ask ourselves a less comfortable question: Are we truly different?

Despite what we like to think, the simplification-survival mechanism of the brain likes to turn off critical thinking even in basic evaluation of facts. Thus, we take things for granted and spend very little energy evaluating statements and the environment around us, and even less time evaluating ourselves. This explains of course why we see Trump in others but not in our own bodies.

But then there is something beautiful about these flaws as well. We created common beliefs in god, money, music, and democracy. All are human inventions one can question and criticize, but somehow they created structure, although imperfect, that allows us to not kill each other (for the most part). Pure logic would have never created these things, as it would get stuck contemplating imperfections. Yet, our ability to simplify, apply bias, and act allowed us to walk the progress lane with a smile, so we agreed to live with a set of shared abstractions.

Humans built civilization by simplifying an impossibly complicated world enough to act within it. The same cognitive shortcut that makes action possible also makes error inevitable. Science is our most sophisticated attempt to correct for this defect, yet it is practiced by the same defective machinery. We are excellent at detecting the resulting biases in Trump, CEOs, colleagues and scientific papers. The experiment becomes much more uncomfortable when the subject is ourselves.

Maybe some day evolution will change us. But until then, simplify, apply bias, and act.

Stop! Yes, stop - This might be the most valuable prediction an AI system can make in drug discovery.

Yea, I know. The entire pharma industry story was written around creation and the AI narrative excites shareholders and hungry BD people looking for a golden goose laying new targets and designing molecules and drug candidates at previously unimaginable speed.

Subscribe now

The imagery is invariably generative: more molecules, more targets, more shots on goal. Yet the economics of drug discovery suggest that one of AI’s most consequential applications could lie in precisely the opposite direction. The industry may need artificial intelligence less to tell it what to start than to tell it, earlier and with greater confidence, what to stop.

Drug development is an unusually expensive exercise in discovering that you were wrong. A molecule can look attractive computationally, survive preclinical testing, pass initial safety studies and the scientist’s wishful thinking and still fail when someone finally asks the expensive question that matters: does it meaningfully help patients?

Failure is so common that attrition turned from unfortunate exception to basic economic architecture. Billions of dollars are routinely invested in programs that never become approved medicines. This ocean of mass belief in success of random lottery tickets still attracts investors who are simply bad at math and still prefer to invest in gold over picks and shovels.

What’s curious is that the timing of failure is economically important. A drug that fails in a computer simulation is an inconvenience. One that fails after years of toxicology work, manufacturing investment and clinical development is a financial event. One that reaches a large Phase III program before revealing an efficacy problem can consume substantial capital, management attention and, less visibly, the time and participation of patients. The industry’s objective should therefore not simply be to reduce the number of failures. Biology makes that aspiration heroic. It should also be to move unavoidable failures as far to the left of the development timeline as possible.

This is where AI’s less glamorous opportunity begins. Imagine a system that combines preclinical results, pharmacokinetic data, biomarkers, human genetics, imaging, previous trials, disease biology and accumulating clinical observations to continuously revise the probability that an asset will ultimately demonstrate an acceptable benefit-risk profile. The output need not be a theatrical green or red light. It could instead make uncertainty more explicit: given everything currently known, this program has deteriorated from a 25% probability of technical success to 8%; obtaining another particular dataset could materially resolve the uncertainty; absent that evidence, committing hundreds of millions of dollars to the next phase has poor expected value. That is not automated drug discovery. It is automated skepticism, a commodity of which large organizations rarely suffer an excess.

The math is attractive because avoiding a bad investment can be valuable even without producing a single new medicine. Consider a deliberately simplified example. Suppose a company is contemplating a $300M late-stage program. If better prediction identifies ten genuinely doomed programs before that commitment, the gross avoided expenditure is $3B. The calculation becomes much less flattering if the model incorrectly kills one medicine that would have generated billions in value, which is precisely why claims that AI can simply become pharma’s automated executioner deserve suspicion. The relevant metric is not accuracy alone. It is the economic cost of false positives and false negatives, calibrated to the decision being made.

This distinction is particularly important in drug development because the data is dangerously unstable. Historical pharma datasets encode previous scientific assumptions, portfolio decisions, trial designs and commercial priorities. A discontinued asset did not necessarily fail because the molecule was bad. Companies terminate programs because competitors arrive, strategies change, patents weaken, financing disappears or another asset looks better. Conversely, a program that advanced was not necessarily scientifically superior. Training an algorithm on historical go/no-go decisions without carefully distinguishing these causes risks teaching a machine to imitate yesterday’s corporate judgment rather than improve tomorrow’s scientific judgment.

At the same time, the algorithm shouldn’t be expected to replace the experiment that actually resolves uncertainty. Clinical trials exist because predictions about human biology have an ugly habit of encountering humans. The more credible role for AI is therefore decision support: finding weak signals across datasets that humans cannot easily integrate, identifying inconsistencies between mechanistic hypotheses and emerging evidence, predicting toxicity or lack of efficacy, improving patient stratification, and showing decision-makers where their confidence is least justified. Regulators are already treating AI in drug development in similarly contextual terms. The FDA’s framework emphasizes a clearly defined context of use and risk-based assessment of model credibility rather than some universal certification that an algorithm is “good at drug development.”

There is also an organizational/management problem that may prove harder than the technical one. Pharmaceutical companies don’t generally lack committees capable of saying no. They lack committees that can say no at the right moment. Once an asset has accumulated a team, a budget, executive sponsorship and years of work, stopping it becomes more than a scientific decision. Careers, forecasts and narratives become attached to molecules. The phenomenon has familiar cousins in every capital-intensive industry: yesterday’s investment becomes today’s justification for tomorrow’s investment (we call it escalation of commitment in psychology). A sufficiently credible predictive system could therefore have value not merely because it sees biology differently, but because it changes the institutional politics of evidence.

This suggests that the most useful pharmaceutical AI systems may look less like molecule factories and more like unusually unforgiving investment committees. They would continuously ask whether the evidence supporting a program has strengthened enough to justify the next dollar. They would compare an asset with historical analogues, expose assumptions embedded in forecasts, quantify what new evidence would change the decision and force teams to explain why a deteriorating probability of success should still command scarce R&D capital. None of this requires surrendering judgment to an algorithm. It requires making judgment more disciplined.

The opportunity extends beyond outright termination. An AI system might conclude that a molecule is unlikely to work in the population originally selected but has a stronger mechanistic case in a biomarker-defined subgroup. It might recommend redesigning a trial, changing an endpoint or demanding another experiment before escalating expenditure. In this sense, “killing drugs earlier” is useful shorthand for a broader capability: allocating experimental capital according to evidence rather than momentum. Sometimes the correct decision will be death. Sometimes it will be a cheaper experiment. Occasionally it may rescue a program by revealing that the original development strategy, rather than the molecule, was the problem.

There is already evidence that AI is moving deeper into clinical development. The FDA says it has seen a substantial increase in drug submissions containing AI components, spanning nonclinical, clinical, post-marketing and manufacturing uses, and its current guidance architecture explicitly contemplates AI-generated information being used in regulatory decision-making. Recent work has also examined AI’s ability to improve the operational efficiency of clinical trials. None of this, however, proves the stronger proposition that algorithms can yet predict clinical failure with enough reliability to transform portfolio economics. That evidence will require prospective validation: models making predictions before outcomes are known, compared against expert decision-making, with both avoided expenditure and the opportunity cost of mistakenly terminated program measured over time.

Such evidence would also provide a better benchmark for the industry’s AI race. Counting AI-designed molecules entering pipelines is easy and photographable. Counting money not spent on trials that never happened is considerably less glamorous. Yet pharmaceutical productivity ultimately depends not on how many molecules an organization can generate, but on how intelligently it allocates scarce capital among them. An AI system that generates 100 additional plausible compounds could even make the industry’s economics worse if it merely supplies more candidates for expensive failure.

The deepest promise of AI in pharma may therefore be selection rather than abundance. Modern computational systems can produce hypotheses at astonishing speed, but experiments remain expensive, patients remain scarce and human biology remains stubborn. As the cost of generating possibilities falls, the value of deciding which possibilities deserve further investment should rise. The bottleneck moves from invention toward judgment.

Silicon Valley has trained us to associate artificial intelligence with making more things. Pharmaceutical companies may discover that its greater contribution is knowing which things never to make at all.

We are screwed, ladies and gentlemen. We are totally screwed. There is a new human-made disease going around, spreading faster than rabbits. No, not the disease you protect yourself from by wearing a mask on your chin, and not the one you ignore under the leadership of Robert F. Kennedy Jr. or by joining the French in proudly ignoring it because accepting it will violate your privacy. It’s a disease that affects everyone and it has no cure.

And technically it’s more than a disease. It’s an allergy.

An allergy to artificial intelligence.

Subscribe now

You’ve seen it. A suspiciously polished email. A photo with immaculate lighting and a perfect body that spells out wishful thinking. Words in LinkedIn posts that provoke an almost dermatological reaction. Text messages from your co-workers using em-dashes. Say “delve” to your friends and you will be accused of being a large language model.

Yes, a lot of it is probably true. Humans quickly adopt productivity tools and AI is a great example. But that’s not where the problem is. The problem is that this allergy to AI is changing human culture. All of a sudden we refuse to be ourselves. We choose our words differently, both written and verbal. We wonder if movie posters are AI generated and treat all written content with 10x the level of suspicion we had for Facebook posts advertising cancer healing with magnetized water or politicians claiming that higher billionaire taxes will solve every problem on earth.

There is nothing wrong with knowledge being transmitted if it’s written by AI. It’s still knowledge. We still learn. And it might arguably be better than human teachers, as AI has less bias than humans and it has no feelings if you throw a fit. Images and videos are equally entertaining and philosophical, whether they are designed by AI or humans. And a lonely old woman dying of cancer in the hospital shouldn’t push back against a cure if it was designed by AI.

Yet we push back. We make our lives harder. We ignore educational materials written by AI and choose human-bias-generated materials instead. A thoughtful scientific LinkedIn post with a call for action will be dominated by comments discussing a far more important subject, criticizing that the article was written by AI. School kids are asked to spend hours on manual research on .gov and .org websites instead of using ChatGPT, because of course the official White House statement that Tylenol causes autism is more authentic than an indexed library of scientific facts.

Yes, we are becoming allergic to AI, and some of my recent Substack posts provide a rather amusing case study. I spent more than two hours writing my last one, only to have Substack inform me, with impressive confidence, that it was 100% AI-generated. This presents writers with a peculiar problem: should I deliberately make my writing worse to establish my humanity? Introduce a few clumsy sentences, perhaps, or misspell something reassuringly difficult? Or should I simply stop sounding like myself because AI has learned to sound vaguely like me? Perhaps I am vaccinated against this disease, but I have decided on the less accommodating option: I am not changing anything. If the machine can’t tell the difference, maybe the machine is the thing that needs improving.

Perhaps that’s the strangest symptom of our new allergy. We are so determined not to be mistaken for machines that we are beginning to edit ourselves around them. We avoid certain words, distrust beautiful images, question polished writing and sprinkle imperfections into our work as little certificates of humanity. AI spent years learning how to sound like us. Now we are learning how not to sound like AI.

And that may be the punchline to this entire technological revolution. We were worried that artificial intelligence would become indistinguishable from humans. We spent considerably less time worrying about what humans would do once it did. Today we are ignoring the fact that AI learns every day, and our quickly-changing communication style will once again be matched by AI within a year, forcing us into this vicious circle of reinventing ourselves for the sake of sounding human.

So use “delve” if you want to. Write a suspiciously good sentence. Send an email with an em-dash. Make something beautiful enough to be accused of using a machine.

Being mistaken for AI may soon be one of the more peculiar compliments a human can receive. And perhaps that is the final irony (em-dash) we built machines to imitate humans, only to discover that humans would start imitating themselves less.

We just published a paper with results that have generated a surprising amount of praise from the scientific community. I’m proud of that. But the most interesting thing about the paper isn’t how well our models performed, but what happened when we made the test harder.

Sure, there are some spectacular numbers coming out from AI drug discovery startups and the noise they generate. But getting a great score and being great at discovering something new are not necessarily the same thing. That distinction could matter enormously for where this industry goes next.

Let me explain…

Your computer can look pretty intelligent when you ask it questions that resemble the ones it studied last night. That’s an inconvenient design flaw in AI models, and an especially important one in drug discovery. Over the past few years, increasingly sophisticated models have arrived promising better predictions of molecular interactions, protein structures, binding poses and affinities. Their progress is usually accompanied by a familiar collection of evidence: accuracy scores, enrichment factors, correlations and olympic style benchmark leaderboards. Higher is better. State of the art is the best. Unfortunately the industry doesn’t award medals, although you wouldn’t know it from press releases.

Yes, I get it. We need benchmarks. Science would be rather difficult without measurement. But a benchmark can answer only the question it was designed to ask, and sometimes this question is considerably easier than the one scientists actually care about.

Current AI drug discovery model makers train their models on what’s already known by humans, so their models… well, tell you what’s already known. That’s just math. Remarkable as a statistical model, but not the most useful one. A more useful model will instead tell you something about molecular systems you know nothing about.

That distinction may sound academic but it can be enormously consequential.

Imagine evaluating a college student with a final exam containing problems almost identical to their homework. The student scores 97%. Excellent. Parents are proud. Student updates his LinkedIn profile. Now give the same student problems based on principles they have learned but presented in unfamiliar forms. Perhaps the score remains 97%. Perhaps it falls to 65%. Those two outcomes tell us very different things about what the student has learned.

Machine-learning models face a similar problem. Models used in molecular science are trained on existing experimental data: proteins, ligands, structures, binding measurements and other observations accumulated through decades of scientific work. We then need a separate set of examples to evaluate whether the model has learned something useful.

The difficulty is that biological and chemical data are deeply related. Two proteins can have similar sequences. Two ligands can have similar chemical structures. Binding sites can resemble one another. So even when we carefully separate the data used to train a model from the data used to test it, the test set can still contain examples that look a lot like things the model has already seen.

The model has technically never seen the test example. But it may have seen its molecular cousin. That distinction matters because the world of drug discovery is not conveniently populated with molecules recognizing each other at a family reunion. The most interesting projects often begin precisely where our knowledge becomes sparse: an unusual target, a new chemical scaffold, a binding pocket with little precedent, or a molecule unlike those already characterized.

Discovery, by definition, has an awkward habit of taking us away from familiar territory.

We have been thinking about this problem a great deal at Deep Origin. While developing our molecular docking and scoring technology, we wanted to know how well our models performed when similarity to their training data was deliberately reduced. That meant filtering evaluation systems based on protein sequence similarity and ligand similarity rather than relying solely on conventional benchmark splits.

The result was instructive. Performance changed as familiar examples were removed. Different methods changed by different amounts. Models that looked impressive under one evaluation could look rather different when the test became less familiar.

This doesn’t mean those models are bad. And it doesn’t necessarily mean a particular benchmark is fraudulent. It means something more useful: the way we construct an evaluation can substantially change what we believe a model is capable of.

This is hardly unique to drug discovery. Machine learning has wrestled with data leakage, distribution shift and memorization for years. Molecular science simply gives these problems particularly expensive consequences. If a movie-recommendation algorithm generalizes poorly, somebody watches a disappointing film. If a drug-discovery model generalizes poorly, chemists can spend weeks making the wrong molecules.

Benchmarks became seductive. Leaderboards became attractive. They compress complexity.

A scientific method may have dozens of strengths, weaknesses and assumptions. A benchmark turns all of this into a number. Method A gets 0.71. Method B gets 0.78. Method B wins, except this thinking is wonderfully convenient and occasionally dangerous.

The problem becomes more acute in AI because models are exceptionally good at exploiting statistical regularities. That is, after all, much of what we train them to do. If information in a test set is strongly represented in the training distribution, a sufficiently capable model may perform beautifully without possessing the broader generalization we implicitly attribute to it.

The benchmark is not necessarily wrong. Our interpretation is. For drug discovery, this suggests that we should become considerably more interested in the distance between training and testing data.

How similar are the proteins? How similar are the ligands? How much precedent exists for this binding site? Does performance survive as those similarities decrease? And, eventually, does the prediction survive the most unreasonable benchmark of all: an experiment?

These questions make evaluation messier, and that’s by design.

And then there’s biology: it has no test set. Nature was just never told that it needs to create a retrospective benchmark. There is no hidden Excel file somewhere containing all the molecules humanity has yet to discover. We just simply can’t reserve 20% of nature for final evaluation.

Real drug discovery therefore involves a kind of distribution shift by default. Scientists use what is known to make decisions about what is unknown.

This is why prospective validation matters so much. A retrospective benchmark asks, “Could the model recover answers that humanity already knows?” A prospective experiment asks, “Can the model help us find an answer we don’t?” The second question is much harder. It is also much closer to the reason these tools exist.

This is why, alongside retrospective benchmarks, we have been putting increasing emphasis on prospective tests of our own methods. Predictions are made first. Experiments happen afterward. Sometimes nature agrees. Sometimes it doesn’t (yea, nature is a difficult customer). The failures are particularly useful. They tell us where the map ends.

There is a peculiar incentive in technology to demonstrate products under favorable conditions. This is understandable. Nobody launches a new car by driving it into a swamp. Science should have nearly the opposite instinct.

If we believe a model can generalize, we should seek out cases where generalization is difficult. If we believe it can identify active molecules, we should test unfamiliar chemistry. If we believe it can work across proteins, we should deliberately move away from proteins represented in its training data.

We should make our models uncomfortable. Very uncomfortable. And we shouldn’t give it any emotional support training data.

This philosophy produces less spectacular demonstrations in the short term. Scores tend to go down when tests become harder. Charts become less suitable for victory laps. But a lower score on a difficult test may tell us considerably more than a near-perfect score on an easy one.

This is especially important as AI systems become more powerful. The better models become at finding patterns in enormous datasets, the more carefully we must distinguish between recognizing familiar territory and reasoning successfully in an unfamiliar one.

Drug discovery needs both. But we should know which one we are measuring.

But there is also a larger point here about AI in science.

We have become fascinated by what models can predict. Perhaps we should be more fascinated by what their predictions allow scientists to do differently. A useful virtual-screening model should help a scientist choose better compounds to test. A useful scoring model should reduce the number of unproductive experiments. A useful structural model should reveal hypotheses that would otherwise have been difficult to formulate.

The output of the system is not really a score, a pose or a ranking, but a better scientific decision. Seen this way, benchmarking becomes less about proving that one algorithm is superior to another and more about estimating how much confidence we should place in those decisions.

This requires tougher evaluations, more attention to generalization and, ultimately, experiments that can prove our predictions wrong.

We recently published details of our work on docking, scoring and virtual screening partly because we think these questions deserve to be examined in the open. The methods matter, of course, but so does the way we test them.

There will be larger models. There will be larger datasets. There will certainly be more impressive benchmark scores. The harder task is making sure those numbers mean what we think they mean. Because in drug discovery, the exam begins when the model encounters a question nobody has answered before.

“Congratulations, son!”

He was genuinely proud. Proud of his son’s newly earned PhD, proud of the dissertation whose title he could barely pronounce, and proud of the years of sacrifice that had finally culminated in a black gown and a framed diploma. The only number that impressed him more than the title of the thesis was the balance of the student loan, which he sincerely hoped had been calculated in yen rather than dollars.

“I am really proud of you. I truly am!”

Having no scientific background, he imagined his son joining the ranks of history’s great minds. Perhaps he would discover a cure for Alzheimer’s or maybe he would invent a new class of antibiotics or possibly add years to the average human lifespan. As the family celebrated over dinner, they spoke with the confidence reserved for graduation ceremonies and lottery winners. The world had gained another scientist. Surely humanity would be better now.

This story is beautiful. There is only one problem. It’s the story we like to tell ourselves, not necessarily the one we live.

Most PhD graduates will not become the next Einstein, and we certainly shouldn’t expect them to. Science advances because thousands of researchers make careful, incremental contributions. But there is a bigger question beneath the father’s optimism. If we are producing more scientists than at any point in history, why do transformative breakthroughs still feel so rare?

While the science is certainly hard and full of surprises, I wonder whether we’ve built a system that trains humans to become the wrong type of scientists.

So here’s a question to think about: What if the most expensive bottleneck in modern science is not funding, computing power, or lab equipment, but the way we educate the people who become scientists?

Modern education has been extraordinarily successful at teaching literacy, numeracy, technical competence, and the ability to work within established frameworks. Those achievements should not be dismissed. They enabled industrialization, economic growth, and unprecedented social mobility. But systems were optimized for class attendance, not necessarily outcomes, and when we do focus on outcomes, we optimize more effectively for mastering existing knowledge than for questioning its foundations. Every kid can probably tell you that the moose loses his antlers in the winter but no one knows what to do with this knowledge. High schoolers often stress over biology classes where they learn the words “cell” and “DNA” but they proudly graduate without knowing what these words mean and why they should care.

The same tendency extends into higher education. Doctoral training is vastly more creative than primary or secondary school, yet it still rewards many of the same habits. Students learn to absorb an enormous body of literature, identify gaps that are acceptable to peers, produce work that satisfies reviewers, and build careers through incremental contributions. These are rational incentives. Science depends on rigor, reproducibility, and skepticism. Without them, knowledge quickly descends into speculation. The problem is not that these qualities exist. The problem is that they can crowd out the qualities that produce occasional leaps rather than steady steps.

History offers plenty of examples. Einstein challenged assumptions about space and time. Barbara McClintock’s ideas about mobile genetic elements were dismissed for years before becoming foundational biology. Katalin Kariko struggled for decades before messenger RNA transformed medicine. These scientists didn’t succeed because they ignored evidence. Instead, they were willing to entertain ideas that initially seemed improbable. Scientific revolutions require both disciplined reasoning and intellectual courage. Remove either ingredient and progress slows.

But our current approach is as effective as asking a human to study rubber friction against the road and then expecting him to design a car.

This is not an argument that scientists lack intelligence. Quite the opposite. Many researchers display astonishing creativity within their fields. And it’s not an argument that scientific consensus is inherently suspect. Consensus often reflects the best available evidence. The concern is more subtle than this. Institutions and incentives naturally shape behavior. When promotions, grants, publications, and prestige depend primarily on extending established ideas, researchers have understandable reasons to avoid risks whose rewards may arrive only after their careers have ended.

Education could do more to cultivate productive disagreement. Students can spend more time designing experiments rather than merely reproducing them, defending unpopular hypotheses with evidence, exploring multiple explanations for the same phenomenon, and learning how to recognize their own cognitive biases. They should certainly question every assumption and fact imaginable, even when accepted by the industry as an axiom. Original thinking is not a mysterious talent bestowed upon a fortunate few. Like statistical reasoning or mathematical proof, it can be practiced, criticized, and refined.

Innovation rarely comes from abandoning rigor. It comes from combining rigor with imagination and bravery to question established facts and group think. Our current educational model excels at producing competent experts. The next challenge is producing more people who are equally capable of asking questions that nobody else thinks to ask. This requires an education system that prizes curiosity alongside correctness, rewards thoughtful risk-taking alongside precision, and recognizes that the future belongs not only to those who know the answers, but also to those who discover entirely new questions.

But until then, I wonder if Pink Floyd was right about our education system: “All in all you are just another brick in the wall.”

Forget the AI Scientist. I’ve got a better idea.

Let’s start by imagining the impossible: tomorrow comes and every scientific paper ever written disappeared. Would medicine continue to advance? Slowly, painfully and at enormous cost, but it would most likely would. Now imagine the opposite. Every paper survives, every journal remains online, every citation is perfectly preserved, yet nobody understands how the world actually works. Oddly enough, that is much closer to our present reality.

Scientists love to think of their jobs as a noble search for truth. In practice, their job is also an elaborate exercise in information compression. Millions of researchers spend their lives reducing the complexity of nature into papers, diagrams, equations and models that fit inside a PDF. Other researchers read a tiny fraction of them, remember an even smaller fraction, and make decisions based on the handful of ideas that magically remain in memory after several years of conferences, coffee and random family torture, augmented by social media distraction and political misinformation. Our civilization, one might conclude, is held together by remarkably optimistic assumptions about human recall.

I started programming when I was 13 and got my first software engineering job at 17. One lesson has remained surprisingly constant throughout my career: computers are astonishingly literal. They do not misunderstand instructions because they are tired, biased, emotionally invested or secretly convinced they already know the answer. They misunderstand instructions because the programmer told them something stupid, which is certainly a much more comforting form of failure.

Many people casually assume computers were designed to resemble the human brain. The comparison is understandable but mostly backwards. Modern computers were built from mathematics, logic and engineering, not neuroscience. Artificial neural networks borrowed inspiration from biology decades later, but the computer itself is not a mechanical brain. In fact, it excels precisely where the brain struggles.

Human memory is less like a hard drive and more like a novelist with a deadline. We do not store exact copies of events. We reconstruct them. We fill gaps, simplify details and quietly edit inconvenient facts until the resulting story feels internally consistent. Psychologists have spent decades demonstrating that confidence and accuracy are distant cousins rather than close relatives.

What’s even more important is how humans think. Contrary to popular mythology, we don’t reason by searching through everything we know. In fact, our brains don’t even record what we know. We record high level versions of what we perceive, adjusted for our flaws, biases, and sleep schedules. We reason through compressed patterns. The brain constantly throws away detail in exchange for speed. That’s not necessarily a flaw, but an evolutionary optimization that somehow made us survive and thrive. A hunter who stopped to analyze every piece of grass did not remain a hunter for very long.

Suppose you once read an article claiming that a particular manufacturing process used by a factory somewhere in China introduced a harmful chemical into certain metal utensils. You probably believed it. You may even remember discussing it. Yet when a waiter hands you a spoon at a restaurant, your brain performs no forensic investigation into its country of origin. The pattern you retrieve is simple. Spoon. Safe enough. Continue eating. If every meal required reconstructing global supply chains, lunch would become a three-day event.

Scientists are no different. They are brilliant, disciplined and remarkably inventive, but they remain gloriously human. No oncologist simultaneously considers every signaling pathway, every failed clinical trial, every obscure protein interaction, every unpublished experimental result and every contradictory paper buried inside an archive from 1987. The human brain just can’t juggle billions of interconnected facts. It was never designed to.

The obvious human response has been to recruit computers. We built databases. And then we created search engines. Then recommendation systems. Today we have large language models capable of reading more papers in a day than any scientist could digest in several lifetimes. Yea, it’s impressive. But it’s also insufficient.

Giving an AI access to every scientific paper ever written is not the same as giving it an understanding of biology. Papers are high level observations about the world, filtered through experiments, interpretations and human language. They describe fragments of reality. They are not reality itself.

Current language models perform a remarkable feat of statistical compression. They absorb an enormous amount of human writing and learn relationships between concepts. Ask the model about a disease and it often produces a coherent explanation because similar explanations appeared many times in its training data. Yea, it’s useful. But it’s not the same thing as maintaining an explicit computational model of biology.

The distinction matters more than it first appears.

Imagine a system that does not merely know that Protein A is frequently mentioned alongside Disease B. Instead, it explicitly represents that Protein A phosphorylates Protein C under specific cellular conditions, altering Pathway D, which increases the probability of Response E only when concentration thresholds and environmental variables satisfy particular constraints. Every interaction is represented as part of an executable model rather than an anecdote extracted from a paper. New experiments update the model itself, not merely the collection of documents describing it.

Such a system would not search for answers. It would simulate them.

That’s not simple LLM and it’s certainly not a knowledge graph with additional boxes and arrows. Knowledge graphs excel at recording relationships. Language models excel at recognizing statistical patterns. What’s missing is an architecture capable of representing the world’s causal structure with enough precision that reasoning becomes computation rather than recollection.

Biology, unfortunately, didn’t receive the memo requesting simplicity. Living cells resemble a city designed by several committees who never met each other, continuously renovated without planning permission and maintained by contractors paid according to the number of pipes they could reroute before unions tell them to go on a break. Yet, despite the chaos, biology follows rules. The difficulty lies in representing enough of those rules simultaneously without drowning in complexity.

If such a computational model became practical, drug discovery would change in ways that resemble engineering more than lab archaeology. Researchers could ask not merely whether two proteins appear together in published literature, but what happens to an entire biological system if one interaction changes under precise conditions. Failed experiments would become data points that refine the model rather than expensive disappointments filed away in forgotten supplementary material.

The scientific method itself would remain unchanged. Hypotheses would still require experiments. Reality would continue exercising its constitutional right to ignore our expectations. But what would change is the quality of the questions. Instead of relying primarily on compressed human intuition, researchers could interrogate an explicit representation of accumulated biological knowledge that no individual mind could ever hold.

Perhaps this sounds impossibly ambitious. Then again, every major leap in computing has involved replacing an activity humans considered fundamentally intellectual with one machines performed relentlessly and without complaint or going on strike. Math was once a profession. So was mapmaking. Chess was supposed to be safe. We have a poor historical record when predicting which forms of intelligence are uniquely ours.

The next great breakthrough in computer science may not be a machine that creates better social media posts or searches for molecules faster. Instead, it may be a machine that finally stops pretending to think like us. And when that machine arrives, the most surprising discovery may not be what it teaches us about biology, but what it quietly reveals about the peculiar way humans have been thinking all along.

There are few institutions more self-confident than a school. A school possesses the serene composure of a man who has been wrong for two centuries and therefore no longer feels obliged to explain himself.

Walk into almost any classroom in the world and you will encounter an astonishing coincidence. Thirty children, born within roughly the same twelve months, are expected to digest precisely the same material at precisely the same pace while a clock, which knows nothing of mathematics, poetry or childhood, announces every forty-five minutes that wisdom has expired and geography must now begin.

Nobody finds this particularly odd.

If the identical arrangement were proposed for hospitals, the newspapers would call it barbaric. Every patient shall receive the same treatment because they happen to be thirty-eight years old. Recovery shall conclude in June. Those still coughing may return after the summer holidays. Yet when applied to education, the idea acquires the dignified aroma of tradition, which has rescued many improbable customs from the inconvenience of explanation.

The education system did not emerge after humanity assembled its greatest philosophers, psychologists, neuroscientists and teachers in a room and refused to let them leave until they had designed the perfect way to learn. It emerged the way old cities emerge. Somebody built a road. Somebody else added a market. A mayor insisted on a square. A monument appeared because monuments have a habit of appearing. A century later everyone agreed that removing anything would be impossible because everything had become essential to everything else.

Schools accumulated in much the same fashion. Some features reflected industrial society, others political ambition, others administrative necessity, and still others the universal human preference for doing tomorrow what worked well enough yesterday. Institutions are collectors. Once they acquire a habit, they display remarkable reluctance to part with it.

Human beings are especially fond of things that fit into columns.

Understanding refuses to cooperate. Curiosity wanders off. Judgment arrives late, usually after making several regrettable decisions. Creativity has the alarming tendency to disguise itself as distraction until someone becomes famous. Exam scores, on the other hand, stand obediently in neat rows, wearing sensible shoes and waiting to be averaged. Faced with these alternatives, bureaucracy has always shown impeccable taste.

This is not dishonesty. It is simply one of the oldest habits of the human brain. We confuse what is easy to count with what is important. We admire numbers because numbers rarely interrupt meetings to point out that they are measuring the wrong thing.

The incentives perform the rest of the work with admirable efficiency. Governments require statistics. Universities require filters. Employers require signals. Parents require reassurance. Teachers require enough order to survive until Friday afternoon. Every participant behaves rationally according to their own circumstances, yet together they produce a system that occasionally resembles an orchestra in which every musician is following the score perfectly while the audience quietly wonders why the symphony sounds like a filing cabinet.

One should not be unfair. Modern education has accomplished extraordinary things. Literacy, scientific progress and technical expertise did not arrive by carrier pigeon. Schools helped build the modern world. The difficulty is that institutions often continue solving yesterday’s problems long after tomorrow has arrived with different questions.

The most interesting assumption hidden inside the entire enterprise is that learning should proceed according to the calendar. January apparently possesses educational properties unavailable in February. June has been universally recognized as an appropriate month to stop understanding fractions. August contributes nothing to chemistry. Time itself has been recruited as an assistant headmaster.

Perhaps this arrangement survives because it offends nobody in particular. Every generation passes through it, complains about it, improves a corner of it, and then carefully hands the whole magnificent apparatus to the next generation with the solemn instruction that education is far too important to experiment with.

And so the experiment continues.

The strange part is not that the education system contains flaws. The strange part is that the only species capable of redesigning the system is also the species that designed it in the first place.

That leaves us with an awkward possibility: perhaps the hardest lesson our schools have yet to teach is that they, too, should occasionally be asked to raise their hand and explain their reasoning.

Strange things tend to happen when an industry runs out of original ideas. It stops inventing new products and starts inventing new ways to describe the same old ones. In drug discovery, people have reached the point where almost everything aspires to be an AI Scientist. Literature search? AI Scientist. Workflow automation? AI Scientist. A chatbot with access to PubMed? AI Scientist. Apparently, all it takes to build an AI Scientist these days is a marketing team vibe-coding for a day with a good copywriter.

The irony is that building an AI capable of truly contributing to scientific discovery is genuinely hard. We know because, at Deep Origin, we built one of the earliest systems we called an AI Scientist over two years ago. We learned quickly that the problem was never about connecting a large language model to a handful of tools. The hard part was reasoning, experimentation, uncertainty, validation, and all the messy realities that exist in real science. We also learned something else. The first generation was only the beginning. So we moved on. What we called an AI Scientist two years ago is old news.

The market, however, seems reluctant to do the same.

Every week yet another company announces its own AI Scientist. Sure, many are thoughtful products solving real problems. Some are excellent research assistants. Others automate repetitive tasks that scientists would happily avoid. None of that is a criticism. The problem is that almost everything now wears the same label, regardless of what it actually does, and most copy each other and slap an “innovation” label to look cool. And then you add yet another ChatGPT wrapper with an AI scientist logo, and the market drowns in noise.

Here’s the challenge: when every product is an AI Scientist, the phrase stops meaning anything.

This isn’t unique to AI. Jonas Ridderstråle and Kjell Nordström described the phenomenon years ago in Karaoke Capitalism (great book! highly recommend!). Companies imitate success because imitation feels safer than invention. Once a category appears to work, everyone rushes to produce their own version. Eventually, the market becomes a giant karaoke bar where everyone is singing the same song while insisting they’re performing an original.

The drug discovery industry is beginning to sound very familiar.

Innovation is not about arriving late with a slightly different user interface. It is not about taking an existing idea, changing the logo, adding “agentic” or “scientist” to the homepage, and declaring it a major breakthrough. Real innovation usually does something much less glamorous. It changes the underlying assumptions. It solves problems people didn’t realize could be solved. Sometimes it even invents an entirely new category because the old words no longer fit.

That is exactly what drug discovery needs right now.

Some of the most interesting companies being built today probably shouldn’t call themselves AI Scientists at all. They deserve different names because they’re solving different problems. Others might benefit from dropping the label altogether and simply explaining what they actually do. Clarity is underrated. Marketing buzzwords have a remarkably short half-life. And pure copycats… Well, let’s just say they have to go back to the drawing board and re-examine their management approach. Simple copying doesn’t generally work well long term.

Science advances because researchers challenge existing ideas instead of endlessly reproducing them. Perhaps AI researchers should borrow that habit as well.

Instead of asking how to build the next AI Scientist, maybe we should be asking a different question: What would you build if you weren’t allowed to call it AI?

Errors are beautiful.

As I was driving to the office this morning, I did something I’ve never done before. I got my mind off of work and started to look around. Buildings, cars, traffic lights and other products of human imagination filled my mind, already overwhelmed by children, work, and political nonsense.

I suddenly began to notice something I never paid attention to before. A tree bending so low, it will probably fall on the house soon. A new businesses opened with a big “We Save You Money” sign, as if putting up a sign will tell me what they sell. A woman driving her car and sitting so close to the steeling wheel that her airbag would probably open into her seat and her face would be featured in a new body horror movie. Trees planted in the middle of the parking lot, as if this subtle sign of random beauty will overshadow the customer’s inability to park and frustration with cleaning the car from random pieces of nature’s presence falling from trees. Public benches divided by metal armrests, carefully engineered to make sure no one accidentally experiences comfort for longer than the city intended. A shopping cart left precisely one foot away from the cart return, suggesting that laziness, much like quantum mechanics, has a finite range. A speed limit sign immediately followed by a billboard advertising a sports car capable of reaching twice that speed before the driver finishes reading the advertisement. Sprinklers watering the sidewalk with remarkable precision while the grass watches from a respectful distance.

I started to think about choices humans made. We created cars that destroy the environment, a school system that teaches compliance while denying critical thinking, and a healthcare system that cures people as long as curing is profitable. These choices weren’t irrational at that time the system was created, but the logic used to design these systems wasn’t always based on first principles or account for all the variables either.

As biologists can proudly admit, life is a complex system of trillions of unknowns. However, human brains aren’t computers and simply can’t process this much information. In fact, the human brain has an embarrassing habit: it compresses impossible complexity into manageable stories. It mistakes patterns for certainty because our ancestors who hesitated too long usually ended up being eaten. People look for high level signals, follow gut feelings, and trust strangers, simply because it worked in the past. And so they design systems that are flawed, proudly announce their achievements and complain about other people’s flaws, turning them into evil billionaires, politicians, and vampires. Millions will die because the healthcare system is not optimized or incentivized to significantly care. Many more millions will die in car accidents because people are not the most optimal drivers and because car safety is constrained by cost, engineering, and consumer demand. And many others will die in wars because of random decisions made by politicians, elected by people using the same logic they used to select what to have for breakfast.

As I was thinking about it, I saw an airplane slowing making its descent towards the airport. I stared at it for a few seconds. Despite its flaws, humanity invented an airplane. And while we know how it works, it celebrates a tremendous achievement in evolution of humanity, as inventing it required the brain to shape into this type of thinking over thousands of years and then learn physics, mechanics, industrial design, and establish a system of logistics, manufacturing, economics, commerce, money, safety, and other things we take for granted.

All of a sudden, this never-ending boring stream of information overload began to make sense. Yes, someone will die in a car accident from improper seat positioning. Yes, millions of people will learn to comply and not invent. And yes, someone out there will spend hours on the weekend cleaning the car from bird poop that comfortably landed on the car because of poor parking lot design. But humanity as a whole will continue.

Our planet has over eight billion people. And while many die, don’t invent, or simply respond to things outside of their control, it still leaves us with millions of people to come up with great ideas, experiment, fail, and try again. And although humanity could progress faster if it was more optimized, we are still making steps in the right direction. And errors we make today? Well, let’s just say that they make humanity beautiful.

And perhaps perfection was never the point. Nature never optimized us for perfect decisions. It optimized us to make just enough good ones that the next generation would have something slightly better to inherit. Civilization isn’t the story of intelligent people building perfect systems. It’s the story of deeply imperfect creatures making enough mistakes that, every once in a while, one of them accidentally invents the airplane.

If aliens visited Earth and watched how we discover drugs, they would scratch their heads wondering why we try to teach computers how to be mediocre scientists.

Imagine hiring Isaac Newton and then spending the next five years teaching him how to fill out grant applications, format figures for Nature, write polite reviewer responses, and remember to cite paper #347 before paper #512 because one of the reviewers might get upset. Congratulations! You have successfully recreated modern science while simultaneously destroying the only reason you hired Newton in the first place.

Subscribe now

But the funny thing is that’s exactly what much of the AI industry appears to be doing.

Every week yet another company announces an “AI Scientist.” The demonstrations are becoming increasingly impressive. The model reads papers, proposes hypotheses, designs experiments, writes reports, and occasionally even generates code to analyze the results. Investors applaud. Pharmaceutical executives nod approvingly. Researchers debate whether their jobs are about to disappear. Somewhere in the background, an LLM patiently summarizes another PDF.

Yet another Claude with a different lipstick was released and raised a billion dollars with a promise of turning biologists into gods.

The assumption hiding beneath all of this activity is rarely questioned. We have decided that the objective is to reproduce a human scientist as faithfully as possible. We teach AI to follow the scientific method exactly as humans do, almost as if the scientific method were a law of physics rather than an adaptation to the peculiar hardware between our ears.

But what if the scientific method looks the way it does because humans are spectacularly limited?

Consider what scientists actually spend their careers doing. They read papers because no one can remember twenty million of them. They specialize in one tiny corner of biology because understanding everything is impossible. They build hypotheses because the brain cannot simultaneously evaluate millions of competing explanations. They look for patterns because processing every variable in a biological system exceeds what evolution prepared us to do. Humans simplify, compress, and approximate. We tell ourselves stories that fit inside our tiny working memory and then spend years testing whether those stories survive contact with reality. And then of course we have bias. We use it to write articles and train LLMs to reproduce it with great precision, while claiming that the AI scientist provides an unbiased view.

There is nothing irrational about the general intent. Human civilization was built by approximation. Newton approximated gravity. Mendel approximated inheritance. Pharmacologists approximate biology every single day. Every model is wrong, as George Box famously observed, but some are useful. Science has always been an extraordinarily successful process of deciding which simplifications deserve to survive another experiment.

But that observation becomes way more interesting once the scientist is no longer human.

An AI scientist does not become tired after reading ten papers. It does not need to choose between immunology and structural biology because it cannot fit both into memory. It has no emotional attachment to the pathway it has studied for fifteen years. More importantly, it does not necessarily need to reduce biology into the same convenient patterns that humans rely upon.

Then here’s the question: If the computational resources exist, why approximate interactions among dozens of variables when hundreds of thousands can be evaluated simultaneously? Why compress the world into an elegant story if the machinery can operate directly on the complexity itself?

Perhaps the greatest advantage the AI scientist can deliver over humans is in how it handles approximation.

Imagine two detectives investigating the same crime. One detective interviews witnesses, notices recurring themes, and gradually develops a theory because there is no other practical option. The second detective watches every security camera in the city, tracks every phone, every vehicle, every bank transaction, every conversation, and reconstructs events without ever needing the comforting shortcut called intuition. Both reach conclusions. Only one depends on pattern recognition because reality had to be squeezed into something a human brain could manage.

Drug discovery today resembles the first detective.

We search for biomarkers because we cannot observe everything. We isolate pathways because whole organisms are too complicated. We classify diseases into categories because continuous biological variation refuses to fit neatly into PowerPoint slides. Every abstraction makes progress possible while simultaneously discarding information. Most of the time, that tradeoff is worthwhile. Occasionally it leads us down a dead end that consumes billions of dollars and a decade of work.

It’s tempting to believe that an AI scientist simply performs these same steps much faster. Sure, it can read more papers, design more molecules, and run more simulations. It can certainly produce more hypotheses before your breakfast than a postdoc manages in a year.

And while speed is certainly valuable, it’s also the least interesting possibility.

The more unsettling possibility is that an AI scientist eventually abandons entire parts of our workflow because they only existed to compensate for human cognitive constraints. Perhaps hypotheses become less central because millions of mechanistic possibilities can be evaluated directly. Perhaps literature reviews fade into the background because scientific knowledge no longer lives inside PDFs but inside continuously updated computational models. Perhaps experiments cease to be fishing expeditions and become carefully selected measurements whose sole purpose is to reduce uncertainty in the model where uncertainty matters most.

Even our language begins to feel strangely human. We speak about “discovering” drugs, as though molecules were hidden somewhere in nature waiting to be found behind a particularly stubborn tree. Engineers rarely describe bridges as discoveries. They build them by working within the constraints imposed by physics. One wonders whether future pharmacology will feel less like exploration and more like engineering once biology itself becomes sufficiently computable.

This raises an interesting possibility. We often celebrate the scientific method as though it were the destination rather than the vehicle. Yet, every generation quietly replaces parts of it. Telescopes changed astronomy. Microscopes changed microbiology. DNA sequencing changed genetics. Machine learning changed protein structure prediction. None of these innovations merely accelerated existing science. They changed what questions could reasonably be asked in the first place.

Perhaps AI belongs in that category. Or perhaps it belongs in an entirely different one.

There is an old habit in technology of assuming that progress means building machines that imitate humans. Cars replaced horses by refusing to gallop. Airplanes conquered the sky without flapping their wings. Calculators became useful precisely because they ignored how mathematicians perform arithmetic in their heads. We accepted those differences without much philosophical discomfort.

For reasons that are less obvious, we seem determined to teach AI scientists to think exactly like us.

Maybe that’s a good decision. Biology is humbling, and humility has an excellent safety record. Or perhaps, many years from now, historians of science will smile at our first generation of AI scientists in much the same way we smile at early airplanes covered in decorative wings and feathers, carefully engineered to resemble the birds they had already surpassed.

If that happens, the biggest breakthrough in drug discovery may not be a molecule, a model, or an algorithm. It may be the day we stopped asking artificial intelligence to become a better scientist and instead asked a far more unsettling question.

What if science itself has been optimized for human limitations all along?

Evolution had been running a respectable business for several billion years. Its rules were admirably straightforward. Whoever adapted stayed. Whoever did not adapt became sedimentary rock. The system required very little paperwork, employed no committees and had never once commissioned a report entitled “Towards a More Inclusive Extinction”.

Then Homo Sapiens arrived. At first, evolution must have been delighted. The newcomers were inventive, energetic and perfectly willing to eliminate one another over such pressing questions as whose mammoth was larger or whose god preferred Thursdays. Everything appeared to be proceeding according to established practice.

Then humans began organizing themselves. Instead of allowing the unlucky fellow with the broken leg to become an educational example for future generations, the tribe carried him home. Instead of abandoning the old woman who could no longer hunt, they asked her to teach the children. Before long, they had invented villages, then cities, then governments, all devoted to the thoroughly suspicious activity of preventing people from suffering the full consequences of reality. One imagines evolution staring at this spectacle with growing irritation. “Excuse me,” it might have said. “The weak are supposed to disappear.” “Thank you for your opinion,” replied humanity. “We’ve opened a hospital.”

For hundreds of thousands of years, the principal challenge was to avoid becoming lunch. A tiger possessed the admirable quality of announcing its intentions clearly. One either outran it or became an anecdote. The modern world is much more sophisticated. It sends emails. Nobody’s ancestors evolved to cope with twelve messaging apps, six streaming services, financial markets that never sleep, artificial intelligence, social media, breaking news every seven minutes and a refrigerator that insists on downloading a firmware update before dispensing ice. Our brains, which performed brilliantly against wolves, now spend their afternoons wondering whether a stranger’s opinion on the internet requires an emotional response. The predators have disappeared, but the adaptation problem has merely changed costumes.

Some people appear unusually comfortable in this magnificent confusion. They absorb impossible quantities of information, recognize opportunities invisible to everyone else, persuade thousands of employees to pursue the same improbable dream and somehow emerge with companies worth more than the economies of respectable countries. Modern capitalism rewards this ability spectacularly because information has become the most valuable raw material on Earth. One might conclude that evolution has simply found a new arena in which to reward adaptation, replacing sharp teeth with sharp minds and forests with financial markets.

Except the story takes a turn that no evolutionary biologist could have predicted. The wealth created by these entrepreneurs does not remain entirely under their control because democratic societies claim part of it through taxation and redirect it toward healthcare, pensions, unemployment insurance, education and countless other institutions designed to soften the consequences of economic failure. Whether the entrepreneur considers this enlightened civilization, a regrettable necessity or daylight robbery is largely beside the point. Evolution never asked anyone’s permission, and neither does the tax authority.

The irony becomes sharper when the wealth exists largely on paper. An entrepreneur may spend thirty years building a company whose estimated value reaches tens of billions of dollars without possessing anything remotely resembling that amount in cash. The fortune consists of factories, software, patents, private shares, research laboratories, long-term contracts and optimistic expectations about the future. Yet the state examines this intricate construction and calmly announces, “Excellent. We shall require a percentage of it.” “I don’t have a percentage,” replies the entrepreneur. “I have a company.” “We’ll accept the estimate,” says the bureaucracy, with the serene confidence of someone who has never had to sell a private business by next Tuesday.

At first glance, this appears to be humanity’s greatest trick. We have defeated natural selection by replacing survival of the fittest with survival of the most heavily subsidized. The productive support the less productive, the adaptable cushion those who struggle to adapt, and civilization congratulates itself on having finally improved upon four billion years of biology.

Only that is not what happened.

The joke, as usual, is on nature because cooperation was never evolution’s failure. It was one of its greatest inventions. Long before there were ministries, there were tribes. Long before there were welfare offices, there were hunters bringing food back for injured companions because next season the roles might be reversed. The tribes that cared for one another often outlasted those that treated every setback as an opportunity to reduce headcount. Compassion, inconvenient though it sometimes appears, turned out to be a remarkably successful survival strategy. Governments did not abolish evolution. They industrialized one of its oldest ideas.

Perhaps, then, billionaires and welfare states are not opposites at all. They are competitors locked inside the same evolutionary machine. One side adapts to an increasingly complex world by creating extraordinary concentrations of wealth, knowledge and productive capacity. The other adapts to the consequences of that concentration by creating institutions capable of redistributing part of it, whether willingly or not, in the name of social stability. Capitalism relentlessly separates. Democracy relentlessly reconnects. Each exists partly because of the excesses of the other, and each justifies itself as the necessary correction.

This may be the defining evolutionary innovation of civilization. Nature evolved predators and prey, parasites and hosts, cooperation and competition. Humans added something entirely new: institutions capable of compelling cooperation among millions of strangers. We often describe this as defeating evolution, when it may simply be evolution expressing itself through a different mechanism. The battlefield changed. The contestants changed. Even the weapons changed. What never changed was adaptation.

Evolution taught us to compete, but it also taught us that the competitors who occasionally stopped competing long enough to help one another had a curious habit of inheriting the Earth. Today the contest looks rather different. Entrepreneurs adapt to markets. Governments adapt to entrepreneurs. Citizens adapt to both. Every group insists it is merely responding to the behavior of the others, and all of them are probably right. Evolution, one suspects, is watching the entire affair with professional curiosity. It has seen predators. It has seen parasites. It has seen symbiosis. It has never seen an audit.

“You know, 9/11 never happened”

I was sitting in an Uber, going down the mountain road, admiring the Mars-like landscape of Sedona, as the driver, probably in his early twenties, confidently explained that 9/11 never happened the way history recorded it. There was no airplane at the Pentagon, he said. The government invented the story to justify more control over society. He spoke with complete conviction, not as someone entertaining a possibility but as someone describing an established fact.

As he spoke, I found myself transported back twenty-five years.

On the morning of September 11, 2001, I lived less than half a mile from the Pentagon. I saw the smoke rising into the sky. I saw the damaged building. I saw the aftermath with my own eyes. I watched thousands of frightened people walking home along highways because they were too scared to drive or take metro. I remember picking up the telephone only to hear nothing but a busy signal because the telecom infrastructure was never designed for an entire metropolitan area to call their loved ones at the same moment. The internet was still young. News websites struggled under the sudden surge of traffic. Much of the information arrived through television, radio, and whatever fragments people could gather from those around them. And the Russian news site (the only one that would open for me at that time), taught by 70 years of propaganda reporting, gave the same type of report you get from Russian news today - powered by wishful thinking and decorated with smoke and mirrors.

I did not learn about those events through a documentary, a social media thread, or a YouTube video assembled decades later. I witnessed them.

That conversation stayed with me long after the ride ended, not because I was surprised that conspiracy theories exist, but because it forced me to confront a much larger question. If someone can reject one of the most documented events in modern history, complete with eyewitnesses, physical evidence, thousands of victims, independent investigations, and endless photographic records, how should we think about our understanding of more distant history and, more importantly, evaluate our civilization and decisions we made?

How does someone comprehend the Second World War? How do they judge the fall of the Roman Empire? What confidence should anyone have about events that occurred two thousand years ago when the available evidence is inevitably thinner? The question is not whether history is perfect. Historians constantly revise interpretations as new evidence emerges. The question is how the human mind decides what deserves belief in the first place.

The answer, unfortunately, is that our brains were never designed to be objective truth-seeking machines. They were designed to help small groups survive. That means using shortcuts, trusting your tribe and copying successful people. It also means noticing patterns even when they are not there and reaching conclusions quickly because hesitation on the African savannah could be fatal. Although these instincts served us remarkably well for hundreds of thousands of years, they leave us vulnerable in today’s world flooded with information.

This explains many of the strange behaviors that surround us every day.

A shop with a French sounding name instantly feels more luxurious. Add the words “Depuis 1898” beneath the logo and many people unconsciously assume heritage, craftsmanship, and quality. Nothing about those words guarantees an exceptional product, yet they reliably change perception.

Put the words “Matcha Ritual” on any storefront and it suddenly feels elevated, mindful, and exclusive. If you think about it, the word “ritual” refers to repeated practices whose significance comes largely from shared belief and belonging rather than objective necessity. Yet in modern marketing it has become shorthand for sophistication, wellness, and intentional living. Our brains happily make that leap with very little evidence.

None of this makes humans irrational in every circumstance. It simply makes us human. We constantly use mental shortcuts because careful analysis of every decision would be impossible. Most of the time, these shortcuts work well enough, but sometimes they produce spectacularly wrong conclusions.

The difficulty is that innovation demands the opposite way of thinking.

Progress rarely begins by accepting the assumptions everyone else considers obvious. It begins when someone asks an uncomfortable question. Why do we do it this way? What if everyone is solving the wrong problem? What would this look like if we started from physics instead of tradition, from economics instead of habit, from evidence instead of social expectation?

This is what people often describe as “first-principles thinking”. No, it’s certainly not mystical. It simply means stripping away convention until only fundamental constraints remain, and then building upward from there. It is extraordinarily difficult because it requires overriding the shortcuts our minds naturally prefer.

Worse still, societies often make this harder. The person who questions accepted assumptions is frequently labelled as arrogant, unrealistic, naïve, or dangerous long before they are celebrated as visionary. Every generation praises yesterday’s innovators while instinctively resisting today’s.

That is hardly surprising. Conformity is comfortable. It reduces conflict. It strengthens social bonds. Civilizations need a great deal of conformity simply to function. But they also need a small number of people willing to look absurd in pursuit of a better answer.

As my Uber ride ended, I told the driver that I had lived near the Pentagon on September 11 and that I had witnessed the attack myself.

He became quiet. Perhaps he wondered whether I was mistaken. Perhaps he thought I was somehow part of the conspiracy he believed in. Perhaps, for a brief moment, he considered that one of his assumptions might not survive contact with someone who had actually been there. I will never know. The conversation ended with silence.

What I do know is that humanity advances because of people willing to question assumptions. The irony is that the assumptions most worth questioning are often our own. That requires humility as much as courage.

The future may not necessarily be built by people who automatically reject conventional wisdom, and certainly not by people who automatically accept it. It will probably be built by those who are willing to ask a more difficult question:

“What is actually true?”

This may sound deceptively simple. It may also be the hardest question our species has ever tried to answer.

I’ve got an idea. Want a scientific breakthrough? Visit a modern art museum.

I just visited the Istanbul Modern Art museum expecting to either admire something I don’t understand or maybe to leave slightly confused, as one often does after walking through a modern art gallery. Instead, I found myself thinking about scientific research. Not because of the pieces where someone appeared to have thrown paint onto a canvas and declared the result profound, or because a pair of glasses left on the floor somehow became an exhibit. Those works may have their place, but they were not what captured my attention. What fascinated me were the pieces accompanied by thoughtful explanations of the questions the artists were trying to explore. They were not simply making objects. They were investigating ideas.

I remember stopping in front of one installation far longer than I expected. It showed the same nude figure in different settings: an art museum, where it was admired; a medical school, where it became anatomy; and a supermarket, where it suddenly felt inappropriate. The body never changed. Only the context did. The question on the wall wasn’t about nudity. It was about humanity: Do we admire the artwork, or are we simply responding to the frame in which it is presented?

This question lingered with me long after I left the gallery. Not because it changed my opinion about art, but because it exposed an assumption I didn’t know I was making. Scientists face the same challenge. Every field has its own frames, conventions and accepted truths. They help us make sense of the world, but they also shape which questions we think are worth asking in the first place.

Most of us spend our education learning what to think rather than how to think. We memorize facts, equations, pathways and definitions. We are rewarded for reproducing existing knowledge accurately. This approach is efficient for transmitting information, but it leaves an important skill underdeveloped: asking good questions. We become excellent at consuming established ideas without ever examining the assumptions that created them. And sure, scientists ask questions for a living. But every field also inherits assumptions so deeply embedded that they stop looking like assumptions at all and they become common sense. They determine which hypotheses sound reasonable, which experiments are worth funding and sometimes which discoveries are even imaginable.

Artists often do the opposite.

An artist asks why a chair has to be a chair. Why a portrait has to resemble its subject. Why a museum should display objects rather than questions. Most viewers either dismiss these ideas or find them fascinating. But that isn’t really the point. The point is that the artist noticed an assumption everyone else had stopped seeing.

The most compelling works in modern art rarely begin with a desire to produce something beautiful. They begin with a question. Why do we accept this convention? Why do we interpret this object in a particular way? What happens if I remove the feature everyone considers essential? What invisible rule has society been following without ever noticing it? Even when the resulting artwork is not particularly enjoyable to look at, the underlying thought process can be deeply valuable.

Walking through a modern art museum, I realized that the real exhibition is not hanging on the walls. It is happening inside your own head. Every explanation invites you to consider assumptions you did not know you were making. You leave not necessarily agreeing with the artist, but thinking differently than when you entered. After enough of those moments, you stop looking at paintings differently. You start looking at everything differently.

That habit is precisely what science needs.

Biology, like every scientific discipline, is built on accepted models and assumptions. Most of the time they are correct. They allow research to progress efficiently through careful, incremental improvements. Yet the discoveries that redefine entire fields often begin with someone asking a question that everyone else overlooked. Why should genes behave this way? What if the variable we’ve been treating as noise is the real signal? What if the accepted explanation is incomplete? What if the experiment everyone trusts is measuring the wrong thing? What if our classification system is preventing us from seeing a different pattern? These questions can sound almost unreasonable until the evidence proves otherwise.

But biology advances when someone asks an uncomfortable question. What if the accepted model is incomplete? What if everyone has been measuring the wrong variable? What if the phenomenon we’ve taken for granted isn’t actually true? Most of those questions lead nowhere. But occasionally one of them changes an entire field.

Scientific revolutions are rarely born from knowing one more fact than everyone else. They often begin with looking at the same facts and refusing to accept that the standard explanation is the only possible one.

Modern art offers an unusual form of training for this mindset. It exercises curiosity without requiring immediate answers. It rewards alternative perspectives. It encourages us to notice the invisible assumptions that quietly organize our thinking. In that sense, the value of a modern art museum is not primarily aesthetic. It is cognitive.

Perhaps we have been visiting museums for the wrong reason. Instead of asking whether we like a piece, we should ask what question the artist was trying to answer. That simple shift transforms the experience from passive observation into active inquiry.

The next breakthrough in biology will probably not emerge because someone memorized one more textbook chapter than everyone else. It will emerge because someone noticed an assumption everyone else had stopped seeing. And if modern art teaches us anything, it is that progress often begins the moment we become willing to question what everyone else quietly accepts.

The most interesting thing about the Bosphorus Strait is not that it separates Europe from Asia. It is that millions of people look at the same stretch of water and instinctively see a line that does not actually exist.

I am writing this from the rooftop of a hotel in Istanbul, looking at the bridge that has become one of Turkey’s defining landmarks. Most visitors admire the scenery. I find myself wondering why humanity became so attached to drawing boxes around things in the first place. Europe and Asia are treated as different continents, despite being part of one continuous landmass. Their boundary is a historical convention, not an unavoidable feature of nature. Yet that convention has shaped politics, culture, identity and commerce for centuries. It is a reminder that some of the most powerful forces in our lives begin as ideas that enough people decide to believe together.

The same pattern appears everywhere. Countries exist because people agree on borders. Money works because people trust pieces of paper and digital ledgers. Companies exist because the law says they do. Religions, ethnic identities and political movements all derive much of their power from shared belief. None of this makes them fake. Social constructs are among the most consequential inventions humanity has ever produced. They influence behaviour every day. The danger arises when we forget that they were designed by people and therefore can also be redesigned by people.

Business offers this lesson repeatedly. The companies that create disproportionate value often ignore categories that everyone else accepts without question. They borrow ideas from biology to improve logistics. They apply psychology to software. They combine finance with computer science, or manufacturing with artificial intelligence. The opportunity rarely sits inside one box. It usually appears in the empty space between several of them.

Our knowledge suffers from a similar tendency. We draw a circle around one collection of ideas and call it biology. Another becomes chemistry. Another economics. Another computer science. These divisions are immensely useful because they allow deep specialization. Yet nature itself has never heard of departments or university faculties. Cells obey chemistry. Chemistry follows physics. Intelligence emerges from biology, computation and mathematics simultaneously. Many of the most exciting scientific advances already occur where disciplines overlap, suggesting that the boundaries serve us best when they remain permeable rather than permanent.

Perhaps this habit of categorising begins early. Much of modern schooling still reflects assumptions developed during the nineteenth century, when industrial societies expanded public education to improve literacy, prepare citizens for civic life and support rapidly changing economies. Age groups, timetables, standardised curricula and examinations were sensible responses to the challenges of that era. They remain effective for some objectives today. Yet the world has become dramatically more interconnected, more digital and more interdisciplinary. It is reasonable to ask whether an educational model designed for industrial societies is sufficient for an economy increasingly driven by creativity, synthesis and continuous learning.

If we continue teaching people to think inside neatly separated subjects, we should not be surprised when many struggle to connect ideas that naturally belong together. We may celebrate the occasional genius who ignores conventional boundaries, when perhaps our institutions could produce many more such thinkers simply by making boundary crossing the norm rather than the exception.

As I finish my tea, the bridge begins to glow. Most people see symbolism. Europe meets Asia. East meets West. Civilisations connect. Those stories certainly attract visitors. Yet there is another story that fascinates me even more. Lighting the bridge makes it more beautiful. A more beautiful bridge attracts more tourists. More tourists fill nearby restaurants. Restaurants earn more revenue, employ more people and generate more tax receipts. The bridge is not merely connecting two continents. It is connecting architecture, psychology, tourism, economics and public finance into a single system.

Perhaps that is the real lesson from Istanbul. The world has always been one interconnected network. Continents, disciplines and borders are useful maps, but they are not the territory. Progress often begins the moment we stop asking which box an idea belongs in and start asking how many boxes it can connect.

I just sat through a meeting where accomplished lab scientists debated the reproducibility crisis. They spoke about statistical methods, publication pressure, incentives, and experimental design. As I listened, I became increasingly uncomfortable. Not because I disagreed with anything they said, but because I realized they were all answering the same question from inside the same frame.

I am not a scientist. I came into the room as an outsider. That may have been an advantage. What I heard was not a scientific problem. I heard a management problem. And once you see it that way, it becomes surprisingly difficult to unsee. In fact, the reproducibility crisis may be the only management crisis that scientists insist on solving as a statistics problem.

For more than a decade, the scientific community has searched for the root causes of irreproducible research. Scientists blame everything imaginable and each explanation contains an important truth. Yes, better statistical methods matter, open science matters and better incentives matter. Yet despite years of reforms, reproducibility remains stubbornly difficult. Perhaps this is because we have been trying to solve the wrong kind of problem.

A laboratory is not merely a place where experiments happen. It is an organization. It has leaders, teams, workflows, documentation, training, quality control, competing priorities, limited resources, and human error. Every experiment passes through this organizational machinery before it ever becomes a figure in a paper. If that machinery is poorly designed, the output should not be expected to be consistently reliable.

This sounds obvious, yet it is almost completely absent from discussions of reproducibility.

Imagine hiring the CEO of a hospital solely because she is the world’s best surgeon. Give her no training in leadership, budgeting, personnel management, operations, or organizational design. Then express surprise when the hospital struggles. We would recognize the absurdity immediately. Running a hospital and performing surgery are different skills.

Yet this is remarkably similar to how we build academic science. Principal investigators become laboratory leaders because they are outstanding scientists. They are selected for their publications, grants, discoveries, and technical expertise. Almost nowhere are they evaluated on their ability to build healthy teams, design reliable processes, mentor people effectively, manage projects, resolve conflict, or create systems that consistently produce high-quality science. These are simply assumed to emerge with experience.

Sometimes they do. Often they don’t.

The consequences are easy to recognize, although we rarely describe them as management failures. Experiments are documented inconsistently. Protocols evolve without being recorded. Students inherit undocumented techniques. Knowledge exists inside someone’s head instead of inside the laboratory’s systems. Projects drift because priorities change faster than plans. New researchers spend months rediscovering what someone else already learned. Every scientist has lived through some version of this. We simply call it “how research works.”

But what if it is simply poor management?

This is not an accusation against principal investigators. On the contrary, it is remarkable that so many laboratories function as well as they do. Modern academic scientists are expected to be researchers, fundraisers, teachers, mentors, recruiters, administrators, communicators, and entrepreneurs, often simultaneously. They are asked to run increasingly complex organizations with little formal preparation for doing so. In almost any other profession, this would be recognized as a structural flaw rather than an individual failing.

This is also why incentives, although important, cannot carry the entire burden. Incentives influence behavior, but they cannot create skills that people have never been taught. A grant can reward data sharing, but it cannot teach someone how to build reproducible workflows. A journal can require better reporting, but it cannot fix a laboratory where documentation has never been part of the culture. Better incentives are necessary. They are unlikely to be sufficient.

Perhaps the most revealing question is not why reproducibility is difficult, but why management is so rarely mentioned. We routinely discuss experimental design, statistics, peer review, funding models, artificial intelligence, and publishing. We devote conferences, journals, and entire careers to improving scientific methodology. Yet the discipline of managing scientific organizations is treated as peripheral, almost administrative, despite determining how science is actually produced.

That blind spot has consequences.

Every laboratory is, in effect, a small company whose product is knowledge. We would never expect a biotechnology company to ignore management while obsessing over analytical methods. We would never advise a pharmaceutical manufacturer to improve statistical analysis while neglecting process control. We understand instinctively that reliable outcomes require reliable organizations. Somehow, that intuition evaporates when the organization is an academic laboratory.

The reproducibility crisis has been misnamed. Reproducibility is not the disease. It’s only a symptom. The deeper problem is that we have built one of society’s most important knowledge-producing enterprises on the assumption that exceptional scientists will somehow become exceptional organizational leaders. Some do. Many learn through years of trial and error. Others never receive the support they need. The remarkable thing is not that reproducibility remains challenging but that we expected a different outcome.

Imagine a laboratory where documentation is rigorous, onboarding is systematic, protocols are treated as living documents, mentoring is intentional, workflows are standardized where appropriate, and continuous improvement is part of the culture. Would every experiment become reproducible? Of course not. Biology is too complex, measurement is imperfect, and uncertainty is part of discovery. But would reproducibility still dominate conversations about scientific quality?

If your answer is “probably not,” then perhaps we have spent years treating the symptom while overlooking one of its most important causes.

The reproducibility crisis may not primarily be a crisis of methods. It may be a crisis of leadership.

Happy Independence Day, America!

Yes, it’s the day we celebrate American patriotism by shooting fireworks made in China. It’s the day we eat hot dogs and drink beer so we can have more reasons to complain about the healthcare system. It’s the launchpad of finger pointing to asshole billionaires, whose wealth is largely imagined on paper, but somehow robbing them of this imagined wealth will fix the healthcare system (presumably hot dogs will no longer cause heart attacks).

And for those who love to criticize America - blind celebration is not an American tradition. It’s deeply human. People around the world celebrate Labor Day. An average person will tell you it celebrates labor, while in reality in celebrates labor unions. I am not sure there is much to celebrate though. Yes, labor unions gave us weekends, holidays, and paid time off, but many employers would argue that today’s labor unions are just legalized, well-organized, mass blackmail - a revolt against the feeding hand, with similarities to how the British viewed the US independence and how California billionaires view taxes on their imaginary vacation homes.

Humans don’t know what they celebrate. An average person can’t tell you why humans chop millions of green trees, hang lights on them, and go on an annual overconsumption shopping spree to celebrate the birth of Jesus. An average teenager views Thanksgiving with the same level of confusion they have when they try to figure out how to send a text message from a rotary phone or when you mention cable TV to them.

Humanity is not a machine that prints genius. Our system prints “bricks in the wall” that blindly comply. And god forbid if someone falls out of group think. They are labeled as mentally ill, the crazy, or the misfits. But the most persistent ones succeed and fly to Mars, at which point the narratives switch from crazy to genius and their invention becomes obvious. And then thousands of people claim they had the same idea, which makes them geniuses as well, because ideas have nothing to do with execution of course. And those who actually made it work are now labeled as billionaire assholes who now have to fix the healthcare system that rewards hot dogs and beer consumption with a dopamine rush and a proud label of patriotism.

But with all its faults, humanity is still great. Yes. We destroy. We do stupid things. But there is something in this craziness that reminds me of a Picasso painting - misunderstood but somehow beautiful and needs to be preserved.

What if the biggest breakthrough in drug discovery isn’t a molecule?

Sounds absurd, doesn’t it? After all, pharmaceutical companies employ some of the brightest scientists on Earth. They have access to extraordinary technologies, oceans of biological data, artificial intelligence, and research budgets that would make most industries blush.

But what if the next breakthrough isn’t hiding in biology at all? What if it’s hiding in management?

That is not a criticism of science. Biology remains brutally difficult. Most drug candidates fail because nature refuses to cooperate. No management framework can negotiate with a protein or convince a mouse to stop being a mouse.

Yet every business decision is filtered through management frameworks, many justified by phrases like “Peter Drucker,” “industry best practices,” or simply “common sense.” Peter Drucker unquestionably transformed management, and many of these practices have stood the test of time. But others have not. Some are relics of technologies that no longer exist. Some were workarounds for problems we have already solved. Some survive because no one remembers why they were created in the first place. And some have simply become traditions disguised as best practices.

Stage gates were introduced to improve investment decisions. In many organizations they also changed incentives, diluted accountability and postponed difficult decisions. Teams are rewarded for moving programs forward, not for terminating them early. Responsibility becomes distributed across committees, making it easier to postpone difficult decisions than to make them. The result is not that science becomes worse. The result is that some programs consume more time, talent, and capital than they would in a system that rewarded rapid learning instead of successful gate progression.

Pharma created committees to improve decisions. Sometimes they do. They also created a new definition of speed that makes turtles wonder whether everyone is okay.

Pharma strengthened security to protect its intellectual property. While this was the the right decision, it introduced another layer of slowness, as every additional approval, every restriction, and every review introduces friction. Security teams are simply doing the job they were asked to do. The question is whether the overall system is optimized for balancing risk with scientific speed or simply minimizing risk at any cost.

None of these observations are particularly controversial. Almost everyone in pharma recognizes them. Almost no one questions them. Why? Because organizations are remarkably good at preserving themselves. Every process, committee, approval, and policy was created to solve a real problem. Over time, those solutions become part of the culture. They stop being questioned. They become “the way we do things.”

Organizations rarely wake up one morning and decide to become bureaucratic. They simply never stop adding. Very little is ever removed. Management becomes like software. Every year we install another update. Almost nobody uninstalls anything. Eventually, the operating system itself becomes slower than the work it was built to support.

Maybe the real problem isn’t bureaucracy as much as the lack of curiosity.

We continuously ask how we can improve a process but almost never ask why this process exists in the first place. That’s a completely different question. And the world is full of other questions we never ask: Why do we have this stage gate? Why do fifteen people approve this decision? Why does this committee exist? What problem was this solving twenty years ago? Does that problem still exist today?

Those questions make people uncomfortable because they challenge assumptions that have quietly become facts. History suggests that every transformational company eventually asks them. Toyota questioned manufacturing. Amazon questioned retail. SpaceX questioned aerospace. Apple questioned music.

Every breakthrough starts with someone refusing to accept, “That’s just how it’s done.”

Pharma has spent decades reinventing medicine. Perhaps it’s time to reinvent management. Not through another transformation program, through another committee, or through another layer of governance, but by developing the courage to question the status quo from first principles.

Imagine if every process had to earn the right to exist. Imagine if every committee had to prove that it accelerated innovation instead of slowing it down. Imagine if every stage gate had to demonstrate that it increased learning rather than merely documenting it. Imagine if every leader asked one simple question before approving another meeting: “Will this help patients receive better medicines sooner?” If the answer isn’t obvious, perhaps the meeting shouldn’t happen.

This isn’t about eliminating discipline but about making discipline serve discovery. Biology will always determine which drugs succeed. Management determines how quickly we discover that.

The pharmaceutical industry doesn’t need fewer brilliant scientists. It needs more people willing to ask uncomfortable questions:

“Why do we do it this way?”

“What assumptions are we protecting?”

If we were building this company from scratch today, knowing everything we know now, would we design it this way? This is not a question of efficiency. It’s a question of imagination and imagination has always been where every great breakthrough begins.