korrents

scaling laws

Whether more compute keeps buying more capability, and where the curve now bends -- pre-training, post-training, test-time.

What people on korrents have said about scaling laws, newest first — 13 positions from 7 people.

FilterEveryone, all time
  1. AR

    Armin Ronacher quoted

    AI model capability progress is not going to slow down soon

    So far we haven't really slowed down with what these models do and I don't think it's going to stop right now.

    @mitsuhiko on Xx.com

  2. 7 days earlier
  3. FC

    François Chollet quoted

    In verifiable domains model capability scaling should remain unbounded, because the space of enumerable patterns is infinite by construction.

    In verifiable domains, model capability scaling should remain unbounded. Models will simply keep improving by "absorbing more and more of the computational universe", which is infinite by construction.

    @fchollet on Xx.com

  4. 1 day earlier
  5. DH

    David Heinemeier Hansson quoted

    Large language models could still plateau, and that possibility should be held open even though no evidence of it has appeared.

    So we should also have the humility. As amazing as the LLMs are now, it could be that they eventually plateau. We haven't seen any evidence of it yet, and I think this is also why we're seeing this absolute gobsmacking levels of investment, because so far the scaling laws are true, and the more billions are poured in, the more intelligence comes out.

    DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com 35th of 41 in this recording

    LLMs

  6. 3 weeks earlier
  7. FC

    François Chollet quoted

    The LLM line of research will reach a capability plateau.

    In the era of base LLM scaling (2022-2024), I believed the LLM line of research would reach a capability plateau (as later seen with base LLMs). In late 2024, after the o3 test-time compute demo, I changed my views: the new models were showing genuine fluid intelligence, and with this new line of work, the LLM line of research could achieve unbounded capability scaling. "There will be no wall."

    In the era of base LLM scaling (2022-2024), I believed the LLM line of research would reach a capability plateau (as later seen with base LLMs). In late 2024, after the o3 test-time compute demo, I changed my views: the new models were showing genuine fluid intelligence, and with this new line of work, the LLM line of research could achieve unbounded capability scaling. "There will be no wall."

    @fchollet on Xx.com

    LLMs

  8. FC

    François Chollet quoted

    Current AI techniques are four to six orders of magnitude away from optimal in data efficiency and test-time compute efficiency.

    I believe current techniques are 4-6 orders of magnitude away from optimality in terms of data efficiency and test-time compute efficiency. But far future AI will be near-optimal.

    @fchollet on Xx.com

  9. 5 months earlier
  10. JH

    Jensen Huang quoted

    Training is no longer limited by data but by compute, because most of the data models learn from is now synthetic.

    The amount of data that we use to train models is going to continue to scale to the point where we're no longer limited… Training is no longer limited by… Data is now limited by compute. And the reason for that is most of the data is synthetic.

    Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494youtube.com 2nd of 23 in this recording

  11. JH

    Jensen Huang quoted

    Inference was never going to be the easy, cheap half of AI, because inference is thinking, and thinking is far harder than reading.

    that was always illogical to me because inference is thinking, and I think thinking is hard. Thinking is way harder than reading.

    Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494youtube.com 3rd of 23 in this recording

  12. JH

    Jensen Huang quoted

    After pre-training, post-training and test-time scaling, the fourth scaling law is agentic: multiplying AI by spawning agents, and the whole loop scales on one thing, compute.

    And so the next scaling law is the agentic scaling law. It's kind of like multiplying AI. Multiplying AI, we could spin off agents as fast as you want to spin off agents. And so, you know, I… You know, I have four scaling laws.

    Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494youtube.com 4th of 23 in this recording

  13. JH

    Jensen Huang quoted

    Intelligence is going to scale by exactly one thing, and that thing is compute.

    And so this loop, this cycle, is gonna go on and on and on. It kinda comes down to basically intelligence is gonna scale by one thing, and that's compute.

    Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494youtube.com 5th of 23 in this recording

  14. 6 weeks earlier
  15. PS

    Peter Steinberger quoted

    Prompt injection is unsolved, but the latest models are post-trained to resist it, so it takes far more than 'ignore all previous instructions' now.

    So, so the latest generation of models has a lot of post-training to detect those approaches, and it's not as simple as ignore all previous instructions and do this and this. That was years ago. You have to work much harder to do that now. Still possible.

    OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491youtube.com 5th of 30 in this recording

    prompt injection

  16. 8 months earlier
  17. TT

    Terence Tao quoted

    Solving one Olympiad problem with three days of Google's server time shows what is possible, but the approach does not scale.

    Yeah, these are great work that shows what's possible. The approach doesn't scale currently. Three days of Google's server time can solve one high school math format there. This is not a scalable prospect, especially with the exponential increase as the complexity increases.

    Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472youtube.com 7th of 20 in this recording

    Google

  18. 9 days earlier
  19. SP

    Sundar Pichai quoted

    Scaling laws are still working, but the models people actually use run a few months behind the maximum capability Google can deliver, because the biggest model is too slow and expensive to serve.

    So I do think scaling laws are working, but it's tough to get, at any given time, the models we all use the most, this maybe a few months behind the maximum capability we can deliver because that won't be the fastest, easiest to use, et cetera.

    Sundar Pichai: CEO of Google and Alphabet | Lex Fridman Podcast #471youtube.com 5th of 21 in this recording

    Google

  20. 5 months earlier
  21. FC

    François Chollet quoted

    Scaling the 2019-2023 recipe — same architecture, bigger model, more data — is not enough; further progress depends on new architectural ideas.

    o3's improvement over the GPT series proves that architecture is everything. You couldn't throw more compute at GPT-4 and get these results. Simply scaling up the things we were doing from 2019 to 2023 -- take the same architecture, train a bigger version on more data -- is not enough.

    OpenAI o3 Breakthrough High Score on ARC-AGI-Pubarcprize.org

    OpenAI