In the last step of training LLMs, reinforcement learning from verified rewards, LLMs are trained to maximize the probability of solving problems using their own output, depending on a reward signal akin to winning in Go. It's not just imitating human written text.

Fwiw, I agree that world models and some kind of learning from interacting with physical reality, rather than massive amounts of digitized gym environments is likely necessary for a breakthrough for AGI.

Human voices don't take 30 seconds to think, retrieve, research, and summarize a high quality answer. Humans are calibrated in their knowledge, they know what they understand and what they don't. They can converse in real time without bullshitting.

Frontier real time-ish LLM generated voice systems are still plagued by 2024 era LLM nonsense, like the inability to count Rs in strawberry. [1]

I'd personally love a voice interface that, constrained by the technology of today, takes the latency hit to deliver quality.

[1] https://www.instagram.com/reel/DTYBpa7AHSJ/?igsh=MzRlODBiNWF...

The intent is still great, but now we should think about writing good general rules for building programs that are essentially all AI generated. What generic wisdom leads to flexible, auditable, composable and robust apps and systems?

Comments URL: https://news.ycombinator.com/item?id=47204331

Points: 1

# Comments: 2

If you can get 99 percent of the quality for 50 percent of the cost, that is most times a good tradeoff.

It seems completely implausible.

I could believe that if a $20 sub used every possible token granted, it would cost $250. But certainly almost no one was completely milking their subscription. In the same way that no one is streaming netflix literally 24/7.

Right now I’m struggling to beat a baseline LightGBM model trained on hand-engineered expert features. My attempts at using a win probability head on top of nanoGPT, treating events as tokens, have been significantly worse. I am seeing about 65% accuracy compared to the LightGBM’s 70%. That 5% gap is huge given how stochastic the early game is, and the Transformer is easily 4 OOM more expensive to train.

To bridge the gap, I’m moving to a hybrid approach. I’m feeding those expert features back in as additional tokens or auxiliary loss heads, and I am using the LightGBM model as a teacher for knowledge distillation to provide smoother gradients.

The main priority here is personalized post-game feedback. By tracking sharp swings in win probability, or $\Delta WP$, you can automatically generate high or low-light reels right after a match. It helps players see the exact moment a play was either effective or catastrophic.

There is also a clear application for automated content creation. You can use $\Delta WP$ as a heuristic to identify the actual turning points of a match for YouTube summaries without needing to manually scrub through hours of Twitch footage.

Now I am trying to use that model to make:

1. A post game instant replay that shows the most important/pivotal moments from the most recently finished game. Some arcades have a seperate display for observers, it could work well there, or as good filler between matches on twitch streams.

2. A personalized per tournament/yearly highlights recap.

If it works well, it might be a kind of tool that generalizes well for summarizing long twitch streams for Youtube.

https://github.com/rrenaud/kq_stream_highlights

https://bmjopen.bmj.com/content/bmjopen/5/7/e008155.full.pdf

It felt pretty easy to read and write, had minimal surprises, and it made writing simple programs easy. The batteries-includedness was great.

It felt like it was designed by a smart guy for practical programming, rather than by a brilliant academic who was wed to purity for maximum elegance. It accepted some warts, but them in mostly ergonomic places.

It was dictated by people who could relegate map, filter, and reduce from builtins to the library. As much as I personally even liked map in particular (I am not super anti functional programming), it's nice to realize the designer had the taste to prefer longer but more explicit programs.

    ys = [f(x) for x in xs]
    ys = map(f, xs)

If you show both of those lines of code to a person in cs101 who has studied Java for a couple months but hasn't seen Python, I am pretty sure they are gonna understand the first line way quicker. Mostly consistent decision making like that leads to ergonomic, practical languages. And they win.

It fits in people's brains better. See also, pytorch vs tensorflow.

Summarize recent working arxiv url

And then it tells me the date is from the future and it simply refuses to fetch the URL.

Because AlphaGo can only do one.

AI could very well be better at formal theorem proving than fields medalists pretty soon. It will not have taste, ability to see the beauty in math, or pick problems and set directions for the field. But given a well specified problem, it can bruteforce search through lean tactics space at an extremely superhuman pace. What is lacks in intuition and brilliance, it will make up in being able to explore in parallel.

There is a quality/quantity tradeoff in search with a verifier. A superhuman artificial theorem prover can be generating much worse ideas on average than a top mathematician, and make up for it by trying many more of them.

It's Kasparov vs DeepBlue and Sedol vs AlphaGo all over.

It's also nowhere near AGI. Embodiment and the real world is super messy. See Moravec's paradox.

Practical programs deal with the outside world, they are underspecified, their utility depends on the changing whims of people. The formal specification of a math problem is self contained.

Low level, automated theorem providing is going to fall way quicker than most expected, like AlphaGo, precisely because an MCTS++ search over lean proofs is scalable/amendable to self play/relevant to a significant chunk of professional math.

Legit, I almost wish the US and China would sign a Formal Mathematics Profileration Treaty, as a sign of good will between very powerful parties who have much to gain from each other. When your theorem prover is sufficiently better than most Fields medalists alive, you share your arch/algorithms/process with the world. So Mathematics stays in the shared realm of human culture, and it doesn't just happen to belong to DeepMind, OpenAI, or Deepseek.