We can barely inspect much of it, let alone fully understand it.
I'm not sure I see one is clearly more difficult than the other.
All it takes is to identify the syntax, so to say.
If it is indeed HIGHLY analogous to programming, we would then expect LLMs/future systems to be HIGHLY proficient at accurate ex-vivo gene [or enzyme/protein] modification/construction
But the hard part is really nothing is annotated or defined. We have annotated and defined some things but its tricky work and so much left to describe. Its like you have entered a house and have no idea what each room is for, or what the light switches do, or even what even is a light switch, or a room for that matter. Maybe you identify a repeated plastic switch through the building that seems to be nearby doorways, you call this the light switch. What does it do exactly? Have to flip it and hope you can detect what changed. Hopefully when you flip it the whole house doesn't just die in the womb, but actually limps along in some way where you can say "this switch controls the garage developing as an attached structure or detached in the back yard" Even more fun when the switch is just one piece of the circuit of a dozen plus switches that all have to flip a certain way in a certain order over a certain time for some function.
And a [YouTube talk by the author](https://www.youtube.com/watch?v=984vm12HUF0).
Sorry about this "not A but B", now is one of those situations where is needed.
Is not:
- cellular automata, - Turing machines
implemented chemically.
The goal, first of UPIM, then chemlambda or chemSKI, is simply to: - find chemical complexes, - or to make them
(though I suspect that we shall discover them in our cells)
so that they enter in random chemical reactions which are akin the graph rewriting inspired by lambda calculus or SKI combinators or Interaction Combinators.
The thesis is that this chemical translation still can do "anything" despite the lack of control of reactions or the combinatorial explosion of possible reaction networks.
Under this thesis we are graph quines.
Models make progress on coding and math because they can write tests and proofs to an extent. Many industries that are more 'physical' and require performing experiments lack that instant feedback loop. Find a way to close that loop and AI begins to look useful.
But try and convince companies to invest on closing that loop just to see if the current models work well on their problems or not? Tough sell. So Anthropic just shows them, hey look, this is possible and if you don't do it I will.. so they fold.
This is basically what they targeted with this approach. They can't automate the experiments since they are often bespoke towards certain goals or even feelings and assumptions based on sage technician knowledge that isn't really taught in any one place. Instead, they tried to automate the process of searching for candidate targets to then test in downstream lab experiments.
Seems exciting, but this sort of thing has been done for a while with just about every single ml classifier method out there for all sorts of biological data. Just yet another way to slice the pie.