Acceleratable Science

Calibrating scientific AI friction and potential resolutions

Context

Much as GenAI software engineers, interns, note takers and the like have garnered publicity for potential cost avoidance futures - the single biggest needle mover societally would be if GenAI accelerated science, a field that has been hitherto held hostage to Eroom’s Law for 2-3 decades now. But my more recent years of spending time adjacent to scientists makes me wonder not so much - can science be Accelerated by AI, but how acceleratable are sciences given the nature of the organizations they reside in? The answer to the former question is definitive, and the latter - conditional.

The subject of this blog then is:

• how do we calibrate acceleratability on a granular basis?

• what are the battles worth fighting, even if the war is daunting?

• is the steady state an impasse, if not how does it resolve?

People

I deliberately used 'Science’ as a catch all term - but clearly scientific tribes i.e.(sub-)disciplines across the sciences are widely variant in their risk appetite. Some of this is the nature of the people who self-select into a field, and some the weltan-schauung (more apropos than the english ‘culture’) imposed by the particular scientific field.

I won’t go thru justifying each one of these depiction - but concentrate on the polarities. In physics, by definition - crazy is good, for any investigation into the nature of reality is likely to substantially intersect the feelings of an LSD induced ‘trip’ (or at least as communicated by Aldous Huxley!). Einstein’s theories were as kooky as they came - and that seems to drive a mindset of being open to wide aperture thinking and experimentation.

Conversely biology is the other end of the spectrum, where a strong orthodoxy keeps any gyrations in check. One slightly provocative theory proven not entirely correct, and you might end up in a career bagging groceries in Idaho (to quote Steve Jobs’es ostensible comment about John Sculley). Clinical is substantially worse as ‘Biology meets billions of dollars’. No one in clinical is penalized for taking a year or two (or five or ten) longer. But one (seeming) mistake that leads to anything resembling litigation, and your future career might be market research. So Biology marches on, with a very foggy mechanistic understanding of the human body and a vague ‘it is all very complicated, you (a non-biologist) wouldn’t understand’ mantra. Not to in any way cast aspersions on the notion that the human body is complicated, but this kind of causal hole in a science would cause most other sciences to spend all their energies in resolving it vs celebrating statistical hell.

Clinical science makes biologists look like break dancers in agility. The only time in my career where I have seen 30 year old software technologies represent SOTA in the science was in Satellites (another area where any mistake gets you on headline news, and extreme conservatism is celebrated as justified prudence). I was talking to a Mayo Clinic Doctor on a long flight about the cost structure of a genomic drug for a specific malady. His numbers were ‘under $10K to make the drug in the lab and hand it to a patient willing to take the risk, will eventually hit the market at $3-8 million out of pocket per patient(!!!)”. Why I asked, and the answer was ‘clinical process’.

Even Chemistry, as the somewhat faster space in science due to inroads in Computational Chemistry - is a space where exploring 200 variants of a molecule in a year is consider a pretty hard year’s work. And we aren’t talking the discovery of Penicillin, we’re talking ‘industrial optimization’ as in cost-benefit tradeoffs in and around non Nobel Prize threatening molecules. Exploring just 200 alternatives in most technological fields would be viewed as a punishable ‘sleeping on your job’.

To the extent that the output of science is a function of the per capita innovation velocity of the participating scientist, the question to be asked is - when will the scientist of ilk X convert to a ‘hunger of a starving warrior’ mindset that physicists have (something that tech companies without the ability to bail $1M salaries are eternally thankful for).

Enterprises

The other more complex question is - who is paid to innovate in the innovation delivery ecosystem. This can end up being a magnum opus, in and of itself, but let me take Medical Devices as a data point.

Turns out that the answer varies depending on the specific medical discipline . Doctors and Hospitals are on opposite sides of the coin, and sales guys have a keen understanding and skill to partner with the likely winner! In areas where skill/profitability are high (cardio, neuro, ortho to name a few) - the physician holds the power and wants SOTA technologies to maximize patient benefit. The hospital (aka ‘Integrated Delivery Network’ - IDN) is willing to swallow pride and bite the bullet and ingest significant $ to pad their pockets. In other spaces where reimbursement is IDN controlled, cost indexed innovation is the key. This isn’t restricted to industry ecosystems - after all professors in hard to get employed fields will innovate what yields tenure, ‘starving artist’ scientists (who need 2 postdocs to get a blue chip industry R&D job) will tow the line to short-term profit motives or NIH politics. While most fields eventually benefit from Darwinian innovation, that is not a given before the innovation actually happens.

Regulatory

Most industries are built on Lindy effect truths (something is true because it has 1. been believed to be true for a while and 2. there is no money to be made being devil’s advocate to that truth). A prime example of Lindy effect truth was the Nielsen rating for TV programs - hugely flawed and highly profitable to all concerned. FDA regulations are similar - the major stakeholders dealing with it realize that while it is painful and contorted, effectively dealing with such regulation is itself an industry moat. Dealing with wireless antennas and radiation effects was a moat in the mobile industry, until Apple made antennas obsolete by obsoleting voice calls as a money maker on phones. All this is to say - accelerating science is irrelevant to many money making scientific enterprises, unless market windows impact their business in material ways. In markets that are effectively regulated monopolies (and you’d be surprised as to how many mundane industries fall under this category), science is a ‘page 2’ topic, not a headliner.

Futures - Boost, Bend or Break?

I will counterpoint this rather dour portrayal with the more (dubiously) cheerful - the future doesn’t give a damn what you think, it asserts itself sooner or later. The real question is how the future asserts itself over the present - boost, bend or break?

Innocuous futures with low friction will be boosters (the way terminal based machines were replaced by one’s that supported windows) - but despite the rhetoric about AI-based CoPilots, boosting futures that matter, are rare. Bend is a more painful experience, where innovations hurt the feelings of incumbent innovators but are eventually mostly additive. Quantum mechanics impact on physicists and physics departments might fall under that category.

Break is where an industry or science furiously fights discomfort (perhaps correctly in the short term, but almost always wrong in the longer term) and builds strong incumbent gatekeepers. It therefore prevents almost all likely disruptors .. and falls hard to an unlikely one. As a data point I am intimately and painfully familiar with - Exhibit A is how mobile wireless went from mystical to mundane. In fact, so mundane that a GenZ reader would be mystified that designing a mobile wireless device was ever mystical.

Despite the billions being poured into synthetic biology, clinical and AI accelerate Biology/Chemistry, it is my expectation that this space will be in the break category. The ‘reactions’ to the recent announcement of the Sakana AI Scientist were instructive! As also the celebratory atmosphere when AI-led companies like Recursion failed their first FDA milestone with toxicology results that didn’t quite pass the test.

And...

All this suggests that (non-physics led) science fields like drug development will focus on a glass half empty (even legitimately short on accelerating the in vivo bits) science population, until AI-led superiority is proven beyond all reasonable doubt. There will be a Recursion 2.0, probably created in China where the clinical dictatorship isn’t quite so stifling, the litigation landscape not quite so threatening, the science incumbents not quite so entrenched. I can further hypothesize that cellular physics will be the universal ‘breaker’ of what are to this point empirical sciences .. but I suspect Ray Kurzweil’s next book will do a far better job, so will leave him to it!