No Priors

Max Jaderberg on why drug discovery must get out of the real world

Max Jaderberg· President of Isomorphic Labs at Isomorphic Labs
·~39 min·English·No Priors
AI CompanyTraining
TL;DR

Max Jaderberg argues drug discovery stalled because it is still an empirical, real-world trial-and-error process — and the fix is building simulation models accurate enough that candidates are designed and iterated on a computer, leaving the lab only to validate.

01The Paradox

Every Metric Got Worse as Knowledge Exploded

<strong>Every metric got worse while approvals stayed flat</strong> — cost per drug rose from a few hundred million to ~$1.5B and timelines from 8–9 to 15 years, even as scientific papers multiplied 10–20×.

it's crazy to see just how it keeps getting harder and harder to make new medicines, more and more expensive, longer and longer.

— Max Jaderberg, No Priors
Key Insight
The implication Jaderberg doesn't spell out: if more knowledge made things worse, the bottleneck was never information — it's the method that turns information into medicines. You can't fix a non-scaling process by feeding it more inputs.
02Why It's Hard

We Didn't Know How Aspirin Worked

<strong>Medicine has always been empirical</strong> — we adopted drugs like aspirin and paracetamol without knowing how they worked, and much of the industry still designs by making and testing molecules in the real world.

aspirin we didn't know how it worked. Paracetamol we didn't know how it worked.

— Max Jaderberg, No Priors
Key Insight
Trial-and-error has a built-in bias: it finds the easy wins first. So the drugs left undiscovered aren't random — they're exactly the high-complexity targets the empirical loop structurally cannot reach in finite time or money.
03The Core Bet

Get Out of the Real World

<strong>The north star is to move drug discovery onto a computer</strong>, using the real world only to validate a finished candidate — not to learn and iterate.

We still need to go into the real world to validate because we're coming up with novel, you know, scientific hypotheses and we want to test them out in people at the end of the day.

— Max Jaderberg, No Priors
Key Insight
Note what he is NOT claiming: the real world never disappears. He's demoting it — from the place you discover (slow, expensive, many loops) to the place you confirm (once, at the end). The bet is on driving the loop count down, not to zero.
04The Mechanism

Experimental Accuracy Is the Dotted Line

<strong>Experimental accuracy is the bar</strong>: once a simulation model is as accurate as a real experiment, you no longer need to run the experiment.

to get to a point where it's experimental accuracy. At that point, you don't need to go into the real world.

— Max Jaderberg, No Priors
Key Insight
Experimental accuracy reframes what an AI model is even for: not to top a benchmark, but to match what the wet-lab experiment would have told you — at which point running the experiment is optional. The bar is set by reality, not by a leaderboard.
05Where It Fails

Drugs Really Fail in the Clinic

<strong>Drugs really fail in the clinic</strong> — toxicity can stay invisible until a candidate reaches real patients, long after the cheaper early work is done.

Where things really fail is in the clinic because we never could see the problems until we get to the point where we put these drugs in patients and we say, "Oh no, like there's toxicity that we didn't account for."

— Max Jaderberg, No Priors
Key Insight
The leverage isn't the cheap early lab work — it's seeing the clinical problems before patients do. Everything Isomorphic builds aims to pull that understanding earlier, where a failure is still a prediction to fix rather than a trial that already failed.
06The Broken Proxy

A Rat Is 30% Predictive

<strong>Animal toxicity testing is only ~30% predictive</strong> of human clinical tox — a 50–70-year regulatory proxy Jaderberg says an ML team would reject on sight.

a rat isn't actually a human. It's not actually very predictive.

— Max Jaderberg, No Priors
Key Insight
Jaderberg's tell is that he judges the industry's gold standard by his own field's bar: a 30%-predictive model is one an ML team would reject on sight. The regulatory proxy everyone trusts wouldn't pass review inside his own company.
07What It Unlocks

Exquisite Molecules, Then Cocktails

<strong>Imprecise molecules are hard to combine safely</strong>, which limits cocktails — so precise design is the key to modulating a whole pathway at once.

it's actually because we're not great at engineering exquisite molecules and medicines that means that we can't combine them together in a safe way for patients.

— Max Jaderberg, No Priors
Key Insight
The subtle point: safety and ambition aren't opposites here. Precisely because molecules can be engineered to be safer, you can afford to be more aggressive — combining several at once to modulate a whole pathway, which single imprecise drugs can't risk.