metanonsense

Even pre-AI, there have been proofs that took years to be verified, understood and some were disputed forever and still are (like the Mochizuki / IUT drama). So, even without AI, maths has reached a point of specialization, where only a few people in the world know the areas used in a proof well enough to check it. When „new math“ is invented for a proof, people may even be more disincentivied to put in the effort. So, I‘d argue, AI makes understanding new proofs easier, not more difficult.

Always ran this to impress my friends

    10 for i = 0 to 359
    20 plot 100+100*cos(i), 100+100*sin(i)
    30 next i

Yeah.. that's something that I thought about for a few years now. I think making sense of code bases and software design will soon be a completely new industry. I had expected that companies like Jetbrains would be in a prime position to offer solutions for that, but it takes longer than I expected.

We're listening. We hear you. We're hear for you.

Maybe the boundaries of the memory subsystem are a bit fuzzy.

I think this is only "statistically" true in the sense that training is based on facts and not non-facts (except maybe with the ingestion of flat-earthers literature ;-). The existence of hallucinations in a bare transformer shows that the convex hull is not about information but about text, so the limit may more be "possible linear combinations of text", which allows for much extrapolation and counterfactuals. True creativity may be one reinforcement learning mid-training goal away that rewards novelty over correctness.

This visualization really makes not much sense. Very modest signal-to-noise ratio.

I think with agents all around, honeypots will get more important than ever.

Yeah. I also only use pro in very specific situations. Not (just) because it's slow, but if the question is too simple or vague, pro responses are sometimes overfitting to the noise in my question. Until there is lots of context and the basics are laid out, high or xhigh somehow work better. Pro gives you the last 10%.

It feels very humbling that here is one of the smartest humans on the planet asking questions, and the LLM keeps answering in this "Yes, it's really simple if you think about it" way, like a professor talking to a talented student.

I also had a frustrating but funny conversation today where I asked ChatGPT to make one document from the 10 or so sections that we had previously worked on. It always gave only brief summaries. After I repeated my request for the third time, it told me I should just concatenate the sections myself because it would cost too many tokens if it did it for me.

Impatient execs can also be painful. EDIT: Writing this while I am waiting for Codex to complete, so I may enjoy slow AI more than the usual developer ;-)

Tbf, if I did whay they do in this video, my battery life would more like a handful of seconds.

I think the issue it the (literal) "head-of-line blocking". Some requests can be processed either "concurrently" like tapping a beer from a keg or in parallel by having multiple bartenders, but having all people standing strictly in line makes this difficult.

only learned about SKIP LOCKED because ChatGPT suggested it to solve some concurrency problem I had. Great tool to learn such things.