Clarity

Generally, I appre­ciate this call for precision from Jeremy Keith:

When I talk about large lan­guage models, I make sure to call them large lan­guage models, not “AI”. I know it’s a lost battle, but the ter­mi­nology mat­ters to me.

However, I want to offer a clarification. Citing a post of mine, he writes:

Con­flating these dif­ferent tech­nolo­gies is the fal­lacy at the heart of Robin Sloan’s faulty logic [ … ]

I’m sym­pa­thetic to this crit­i­cism because, as Jeremy observes, there IS plenty of stolen valor, in which the suc­cess of a special-purpose soft­ware tool is trum­peted as a gen­eral vin­di­ca­tion of “AI”. It’s like: “Eye-wateringly spe­cific machine learning algo­rithm esti­mates mass of dis­tant exoplanet … slop must be good!!”

But this is not a mis­take I have made, and it’s not a mis­take — not even a strategic conflation — that the most serious pur­suers of AI super sci­ence are making, either.

First, I’ll point out that I employ the precision Jeremy requests. Note “lan­guage models” here, not “AI”:

It’s pos­sible that lan­guage models could go on broad­ening and deep­ening in this way, and even­tu­ally become valu­able aids to sci­ence and technology, to med­i­cine and more.

Second, the bull case for AI super sci­ence revolves around LLMs, or their descendants, doing LLM-ish things in the lin­guistic domain: receiving instructions, for­mu­lating plans, pur­suing investigations, pos­sibly even chore­o­graphing lab work (which sounds insanely dan­gerous to me, but, that’s the storyline). When Dario Amodei writes about “a country of geniuses in a datacenter” he is imag­ining Claude-alikes, not AlphaFold-alikes.

In this scheme, a souped-up (“superintelligent”??) Claude Code receives its research task, starts working, and comes back in a day — in a year? — with the recipe for a room-temperature superconductor. Maybe Super­Claude develops new AI tools along the way — AlphaCharge, the better to sim­u­late the prop­er­ties of exotic materials! — but the heart of this scheme is the lin­guistic agent dri­ving the work.

For my part, I think the odds that this can work are low; in fact, I would say they have decreased since I made my ini­tial assessment. Yet the capa­bil­i­ties of LLMs have been sur­prising enough to merit, at this point, curiosity and patience.

Anyway, that’s all to say, while Jeremy’s AI coat­tails are a real thing, there is a sincere, uncon­fused case for explic­itly LLM-powered sci­ence that is just as real.