AI is here to stay in academia and engineering without a doubt. In my last post I made the audacious claim that AI accelerated research is the future, and anyone opposing it is a fool. I think my evaluation of the offended academics in that post hit the nail on the head : [CALLED IT]. (The plagiarism allegations might hold up though. Like I said, these corporations aren’t exactly humanitarian aid agencies.)

In my research news : Claude code grinded down my distributed speculative truncated sparse blossom idea [SpecMatching]. On top of it, in collaboration with Claude I solved (to the best of knowledge for the first time) a bottleneck that was present in implementing a particular mode of fault tolerant quantum computation in real time : namely magic state preparation for Clifford+T computation. Preparing high fidelity T states on rotated surface code-ish patches requires postselecting on something called the decoder “soft output” which quantifies the confidence of the decoder’s belief. For matching based decoders the complementary gap is a good proxy for it – which is the minimum cost to flip the observable given a minimal matching. It only took about a week’s worth of prompts, discussions and experimentations to arrive at a pretty elegant solution. With some more benchmarks on how it performs on “yoking” or other kinds of code concatenation, it’s easily a top tier paper. A caveat is that it requires sufficiently low error rates like one in a thousand. But that’s reasonable. I guess.

It’s safe to say that it is downright disingenious to claim full credit for it. I’ll gladly share my chats and prompts if asked. In the spirit of open science, I’ve made the code available on GitHub [fastgap] and the “Papers and talks” section hosts some slides. I don’t even care if I’m “scooped” by someone who reads this and posts it on arxiv. LMAO just check the GitHub commit timestamp at this point. You hearing me boomers? This is how you do scientific attribution and credit assignment from now on. Deal with it! sunglasses emoji

I’m already working on two other ideas regarding architecting fault tolerant quantum computers, including near term partially fault tolerant ones. Currently, the rate limiter is just my imagination and my advisor’s attention.

However, none of the above means that there won’t be any problems arising in doing AI-assisted research, or even engineering. Especially software engineering. My last post glazed LLMs too much, so I’m here to course correct a little.

It comes down to a point made, ironically, by future unemployed mathematicians et. al. They say that mathematics is as much about thinking as it is about knowledge discovery. And if we eventually become lethargic and complacent in our analytical thinking skills, it might have unforseen consequences. Only now have I realized the full gravity of this situation : it’s not just about analytical thinking skills. It’s about the thought process in general. People first gave away writing English prose to AI. (I swear to god if I hear another news broadcast or read some essay that says some distinction matters I’m going to lose it.) Then we lost artistic pursuits to AI. Then software programming. And now mathematics. Soon AI will start generating everything on its own, while still (presumably) being controlled by humans. And the folks who have nothing to do will become useless, exactly like the inflated humanoids in Wall-E.

The question will then become: are they also worthless? That discussion goes beyond the scope of this philosophical shitpost and into the territory of various social sciences. But I will say: the only way out I see is to adapt. Laziness and complacency should be viewed as a choice. Humans have overcome all sorts of crises by virtue of sheer intelligence. Whether we will be able to overcome a crisis of intelligence itself, and make a conscious decision to not give away what make us human is a genuine open question.

Another possibility is that these models won’t even be able to generalize to basic tasks? Who knows? [link] One of the concerns I raised in my last post was the slop. Another aspect which will confound anyone who verifies AI generated output is the absolutely enormous volume of it. Be it mathematical proofs or code, anyone who doesn’t bother to get into the weeds is bound to get lost in the sea of tokens spit out by AI. This will burden servers who host this mess [link], reviewers who read these papers, and committees responsible for any kind of resource allocation. It will also raise competition between researchers, and inevitably the bad behavior that accompanies it. Maybe this will incentivize deduplication of efforts across labs and industries working on the same goals. I sincerely hope these problems will be mitigated and won’t stifle progress.

In summary, I really like my code go brr machine. Don’t make it the instrument of slop and enshittification!

Shirgure out.