Hacker News Reader

> AI can currently do if given enough context

It's worth noting that you can substitute "dollars" for "context" in that sentence, which seems to be where many of these impressive achievements are coming from. As ever, it's unclear whether these models will get cheaper while remaining better, since all of the recent breakthroughs appear to be of the "think more" kind. For translation specifically, I'd be very surprised if the "think more" LLMs would help given the per-unit cost expected of the output.

> all those things aren't magical tasks that can only occur inside a human skull, they are a product of information processing

I agree but it's useful to remember that 1. brains and especially the human brain are enormous and 2. individual tokens carry significantly more meaning than individual tiny muscle twitches so even extremely primitive "cognition" can look like it's doing more work than it actually is.

Yes. It's as if they think AI will forever be LLM only and won't develop world models that incorporate current state assessment, dynamic next-state prediction, cause-and-effect reasoning, object permanence, etc. I'm not in the AI industry but I assume there's got to be lots of research and work being done on this.

As mentioned in the article, the point of language is to communicate with other humans, and you need a human to do that.

Mathematics is famously rigorously defined, it's roughly analog to AI beating humans at chess. Sure it's impressive, but it's also something you'd expect machines to be good at.

People have been skeptical of them doing advanced maths because it involves complicated thinking which the 'stochastic parrot' folk thought wouldn't happen.

> You'd be laughed at if you said that ChatGPT could help you with graduate level mathematics in 2024, but this year, AI models on simple prompts are solving previously unsolved Erdos problems.

I'm curious, do you have a graduate degree in mathematics?

I have a graduate degree in mathematics. AI models can absolutely help you do research math in 2026. I recently asked chatGPT to prove a result which I know to have been published in Advances in Mathematics (a pretty good, but not top tier, journal) this year, and it gave a correct proof which was completely distinct from the one that was published.

Fable has really spooked me, honestly. It's another big jump, but not in the actual coding. I was pretty comfortable with the "you do the implementation, I do the meta work and steering", and ... no steering required, no meta work required. Here's the backlog, let me know when it's complete, I guess I'm going to go touch grass until I have to review and refine... probably tomorrow?

Reminds me of the first time I saw a coding agent stumble through an issue in 2023 maybe? and went "this is a big deal", similarly when OG gpt started making jokes that actually kinda worked.

Updated modern version of the classic "make me a greentext", apologies for slop-posting, but it seems relevant:

    > be me
    > senior software engineer
    > in charge of making sure the tickets get, in fact, implemented
    > occasionally have to open the IDE and write some code myself
    > one day i open the IDE and the ticket is already closed
    > the agent did it overnight
    > no steering, no review notes, nothing left for me to do
    > distress.jpg
    > ask my manager what to do
    > he says "just focus on the high-level architecture stuff"
    > i say "what high-level architecture stuff"
    > he says "i don't know, you're the senior engineer"
    > rage.jpg
    > quit my job
    > become a prompt engineer, nice and simple, just tell it what to build
    > first day on the job, sit down to write the prompt
    > AI already wrote it

Greentext is eh. Very formulaic, in fact very similar to the bottomless pit one, which I'd argue is better because of it's absurdity. I have to ask, did you mention the older GPT version to fable in the prompt?

Of course I did! Wouldn't be faithfully mediocre without the right context

Spicy... I giggled.