AI does not abstract, so your job is not to check the code
Cronin says AI is just a fast shallow explorer: the past is discrete and machine-learnable, the future is continuous and in principle not machine-learnable, and humans can conjure a new space out of nothing and jump to it.
The argument · tap a timestamp to hear it
Computational chemistry simulates; computation builds
Cronin splits apart two words that get conflated: computational chemistry uses first principles to calculate whether a molecule is stable, whether two reagents will react, what isomers exist — the objects are hypothetical molecules and reactions. Computation is something else: can you normalise a series of recipes onto one machine and then run them out on standard hardware. He admits the leap sounded like nonsense at the time, so he borrows Turing's move — first assume a robot exists that can build any molecule, then ask how to instantiate chemistry onto it.
— Lee CroninChemistry has only four primitives
It took him ten years to work out: chemistry has only four primitives — add matter, remove matter, add energy, remove energy. So you can build a tape like a Turing machine's, except each cell is not 0 or 1 but reagent, reaction, waste or product; and the read-write head does not read and write bits but injects reagents, removes products, removes waste. Stack error correction and anomaly detection on top and you get a universal state machine for chemistry. He says it took a decade to get there, which is a bit embarrassing.
— Lee CroninChemify wants to be the AWS of chemistry
Chemify is four years old, 230 people, doing template-based research-grade molecule manufacturing plus discovery. Cronin's judgment is that within twenty years every molecule invented on Earth will go through this kind of state machine. He criticises today's digital chemistry as mostly machine learning inventing molecules and pretending they can be made, because there is no state machine, and the result is "people hallucinating molecules they cannot make". He positions the company as the AWS of chemistry, and says the spiritual goal is to help discover the origin of life — that is its Mars — while what feeds the company is inventing molecules for people.
— Lee CroninThe chemical computer is stuck on the compiler
The team does chemical computation with the Belousov-Zhabotinsky oscillating reaction: a 7×7 array of 49 fluidically connected cells, each able to oscillate between red and blue, coupled by magnetic stirring switches so they process information like transistors. But like quantum computing, it is stuck on finding a suitable compiler to compile problems onto the hardware. He also draws a line: the brain is a 20-watt, 2.5-kilogram wet information processor, but he does not think the brain computes in the Church-Turing sense — that would require a labelling scheme, transition rules and error correction.
— Lee CroninAI does not abstract, and checking code is not the point
Cronin says the most common misjudgment in AI anxiety is "AI writes code fast, so from now on we just check the code". He points out that checking code is not the point; the point is re-architecting and working out what is going on, because AIs do not abstract. He likens the current situation to launching satellites before relativity existed: satellites frame-drag, they lose time, GPS is off by about 6 miles a day, and we can only patch it after the fact. Machine learning lets us see phenomena we cannot explain because we do not yet have the right theory of mind, intelligence and creativity.
— Lee CroninThe past is machine-learnable, the future in principle is not
Cronin's core argument: the future is not like the past. The past is discrete, boxed and labelled by us, so it can be searched; the future is continuous and infinite — how do you sample it? With what sensor — visual, acoustic, something else? You label it according to the error and resolution of your sensing system, and build a picture of the world from that. Hence "we can machine-learn the past, but in principle we cannot machine-learn the future". This is also why he thinks the universe looks infinitely creative.
— Lee CroninOrigin of life wants reproducible irreproducibility
Cronin's lab is starting to see the emergence of persistence in Miller-Urey-type experiments: let the reaction run long enough and small molecules help each other, building a ladder of complexity. One of his postdocs found tyrosine, and also found a dipeptide, but that is below assembly index 15 and still within combinatorial explosion. He tells his team that what you want to see in any origin-of-life experiment is "reproducible irreproducibility". The team asks how that is distinguished from a badly designed experiment, and his answer is that there will be shared signatures: first a reproducible gamish, and then, as it moves toward non-determinism, you should be able to manipulate it. He admits this amounts to admitting that humans are adding information and steering it, and you can never escape that.
— Lee CroninThe chemical space at assembly index 25 is 10 to the 117
Amino acids have assembly indices between 3 and 11, and each step up the number grows the chemical space double-exponentially. Cronin's team recomputed the drug-like chemical space: the old estimate by molecular weight was about 10 to the 60 molecules; switching to assembly index, the space of molecules at index 25 is 10 to the 117. For reference, the number of atoms in the universe is 10 to the 80. He says this is exactly why chemistry is special and why intelligence and creativity come from chemistry — the space of possibilities is simply there.
— Lee CroninIn their own words · checked verbatim
I found was realized that chemistry only has four primitives. It's kind of embarrassing it took a decade to get there, but the four primitives are add matter, subtract matter, add energy, subtract energy.
Lee Cronin10:20
So we're in this ridiculous situation where everyone's hallucinating molecules they can't make.
Lee Cronin15:21
Drawing a molecule on a computer is not proof that it exists.
Lee Cronin16:21
AIs today are what I call shallow explorers. They're fast shallow explorers.
Lee Cronin43:45
we can machine learn the past, but we cannot in principle machine learn the future because the future is continuous and the past is discrete
Lee Cronin47:45
what you want to see in any origin of life experiment is reproducible irreproducibility
Lee Cronin1:08:56
There is no way you can create that molecule randomly in any abundance without an evolutionary process. Zero. It's like finding a million Mona Lisas on Mars.
Lee Cronin1:18:58
I think there's probably life everywhere in the universe. I think that it's a fairly common event, but I have absolutely no data to back that up.
Lee Cronin1:39:07
Figures
| Chemify employees | 230 | 14:21 |
| Chemify years since founding | 4 years | 14:21 |
| BZ reaction computing array | 7×7, 49 fluidically connected cells | 22:26 |
| Brain power and mass | 20 watts, 2.5 kilograms | 23:27 |
| GPS daily drift (under the assumption of no relativistic correction) | about 6 miles | 39:41 |
| Number of parts in a 747 | about one million | 43:45 |
| Cronin's confidence in denying that silicon-based AI is truly creative | 99.99999% | 49:46 |
| Assembly index threshold for objects from living systems | greater than 15 | 1:13:57 |
| Molecular weight of taxol | 833 | 1:18:58 |
| Number of molecules in the chemical space at assembly index 25 | 10 to the 117 | 1:33:04 |
Glossary
- Chemputation
- Executing normalised chemical reactions on a universal state machine rather than simulating molecules on a computer.
- Assembly Theory
- Characterising an object's complexity by the minimum number of assembly steps (assembly index) and its copy number, to judge whether it came from an evolutionary process.
- assembly index
- The minimum number of steps needed to build an object from a fixed set of atomic building blocks; products of living systems are usually above 15.
- reproducible irreproducibility
- The signature you hope to see in origin-of-life experiments: reproducible overall, but the specific products differ each run.
- shallow explorers
- Cronin's name for today's AI: fast search within an existing dataset, but no creative leaps.
How to listen
Engineers and founders working on AI applications, foundation models and scientific computing, especially anyone who cares about the judgment "can AI do original research".
From 1:11:56, the argument over whether entropy is subjective — skippable.