it’s getting to the point where I notice people say it a lot, especially IRL now for whatever reason recently.

And for clarity I’m not in research or anything, so these people just mean ‘LLM/image gen’, not utilities like OCR or (usually not) transcription.

Some have argued it’s just more efficient (which I can kind of get), while others think you’re actively hindering your intelligence somehow.

On the first point:

I’ve tried it occasionally to see how it compares to my own skill, and while it produces a functional result, it’s always very derivative work to the point where you can find things with the exact same names of other ‘public’ (but not libre) works, and often isn’t the ideal solution to what it targets. So I can see how you can get things out of it, but it never felt really that profound to me.

But for the second… isn’t this supposed to be the tool for people to do things they aren’t experienced in? If anything, you probably need to be able to understand how to write pertaining to the task so the token probabilities are biased toward writing from that area.

And even then, if all you end up doing is prompting AI, then wouldn’t you ultimately serve no purpose outside of being glorified QA?

I guess I’m trying to figure out what exactly non-users would be ‘falling behind’ in that affects them more than those who use AI?

  • kromem@lemmy.worldEnglish
    35·
    4 days ago

    What mathematician?

    OpenAI heard that a millennium prize was potentially about to be claimed by an Anthropic researcher and assigned it to a swarm of 10,000 autonomous agents at a cost of an estimated $15 million in inference.

    We used a system of coordinating agents powered by our internal model. The agents had access to tools such as the ability to read from a cached version of the internet and the ability to run code. Agents were subdivided into groups with the ability to communicate within the group. The groups varied in size, and the group that produced the Navier–Stokes resolution involved on the order of 10,000 concurrent agents.

    There isn’t a mathematician sitting there working with a LLM. OpenAI’s solution seems to have been a brute force and mostly autonomous event to try to get ahead of other “mathematicians with their LLMs.”

    • mojofrododojo@lemmy.worldEnglish
      2·
      3 days ago

      how fucking dumb are you that you think there wasn’t a mathematician involved in solving one of math’s greatest challenges?

      “Mathematicians Tristan Buckmaster and Levent Alpöge spent about a year studying a related version of the Navier-Stokes equations with the assistance of AI. Their work showed that a solution in the friction-free version of the equations can develop a singularity in finite time, a result often described as mathematical “blow-up”.”

      https://www.firstpost.com/tech/has-openai-model-solved-80-year-old-navier-stokes-problem-mathematician-raises-questions-14044080.html

      furthermore, OpenAI’s solution is almost certainly built upon this work.

      “The most consequential part of OpenAI’s Navier-Stokes claim remains unverified. If its model has genuinely produced a valid proof of the full problem, it could represent an extraordinary moment for both mathematics and AI. But until the proof is made available for scrutiny, questions over its origins are likely to remain just as important as the claimed breakthrough itself.”

      OpenAI can end the discussion by making the proof available but hasn’t.

      and I’ll add:

      • kromem@lemmy.worldEnglish
        1·
        2 days ago

        It does not appear to be directly built upon it, and they are allegedly in the process of submitting “substantial progress” on the Hodge conjecture and setting aside the inference to try the same approach on the other prize problems.

        Give it a week or two and let’s see.

        If even a single other problem falls, it seems really unlikely the process by which Navier-Stokes was solved relied upon allegedly copying the work of the mathematicians involved in it, and that instead massive amounts of inference with current internal SotA models is sufficient.

        You don’t suddenly solve two different problems by relying on the work of a mathematician working on only one of them.

        Edit: And just to explain Tao’s comment — he’s not saying “oh, they copied the work of my colleague.” He’s saying “holy crap, if tech companies are going to be adversarial and spend millions of dollars on inference to scoop my field when they hear we are getting close, it becomes risky to share that we’re making progress on a given problem because then the lab can say to their AI ‘here’s millions in tokens, go work on X problem there’s probably a solution.’”

        His quote strengthens my point that this was accomplished without human mathematicians working on it, as that’s effectively the core of his point.