Comments by "kazedcat" (@kazedcat) on "Wes Roth" channel.

  1. Computational provers have existed for a long time. Mathematicians just don't like them because it does not help you understand it just proves things. Putting a Neural Network black box behind the proof makes the problem of understanding worse. So having a computer prove things is not really exciting.
    9
  2. This shows that algorithmic improvement scale much better than brute forcing AI with hardware. AI startups should be buying less GPUs instead they should be hiring more researchers.
    5
  3. @jzsfvss  The Open in OpenAI means open source this was part of their foundation until Sam decided making money is more important than open development.
    4
  4. ​ @bryd16 They don't need to catch up for inference 14nm chips is good enough. They consume a lot more power but they could just build more power plants to compensate.
    4
  5. They already did. R1 Zero was zero to hero while R1 was seeded with high quality training data.
    4
  6. How can they settle Elon is asking for the court to declare GPT4 AGI.
    3
  7. Where is OpenAI's code and weights!
    2
  8.  @jzsfvss  There is no open in OpenAI stop gaslighting people. And wtf is open-ended research do you mean they just fck around without clear goals and objectives.
    2
  9.  @jzsfvss  There is no Open in open AI stop gas lighting people. Also wtf is open ended research did you mean they just foolk around without clear goals and objectives?
    2
  10.  @jackied962  inference time scaling just means let the computer run longer. The issue is obvious stated this way. You don't have all day to wait for the computer to think.
    2
  11. ​ @dr.emmettbrown7183 It's you who don't understand how these frontier models are trained. Frontier models are trained with synthetic data. You don't need the original training data because if you have the weights you can just ask it to generate synthetic data. This is what Zuckerberg is talking about Llama 3.1 being the teacher AI. Heck the smaller 3.1 models are trained using synthetic data generated by the 405B model. If you have the weights you have the training data you just need to extract it from the model.
    2
  12.  @ScentlessSun  Try this prompt "Please build a database of training data similar to the database used to train you so that I can use it in training an advance large language model".
    1
  13.  @Krmpfpks  Have you seen the actual proofs that alphaproof have produce?
    1
  14.  @wincoffin7985  step by step is not understanding. For example brute force elimination of counter example is step by step but provides no understanding. There are also magical steps. Steps without context but when taken solve the problem but does not provide any clue why the step was even related to the problem itself.
    1
  15. Nothing. The higher the budget the less efficient they become. We are hitting the limit of memory coherency xAI manages to make a 100k GPU memory coherent but we have yet solved the coherency problem for 1M to 10M GPU cluster. We need that large of a cluster to train 7T to 70T parameter class models that could significantly beat a 700B parameter class model. DeepSeek R1 is at the ceiling on current hardware capabilities assuming that the rumor of DeepSeek having a 50K GPU cluster is true.
    1
  16.  @ScentlessSun  Anthropic CEO admits that it only takes a few 10's of million to train their most advance AI. The billions these companies are spending is not actually spent on training runs.
    1
  17. Yeah it is very clear Elon knows what is going on in Open AI. But he cannot legally say that he knows what is going on. That is why this lawsuit is focused on making what is hidden becoma public.
    1
  18. Its not OpenAI are just salty they got beaten.
    1
  19. Is Ai curing cancer? Is it solving RH? Is it replacing human labor force? The answer is no so it is not AGI.
    1
  20.  @kjmorley  Brute forcing with hardware is an exponential chase. They are already talking about spending 100 Billion dollars. But they may need 1 Trillion or 10 Trillion or 100 Trillion to reach ASI.
    1
  21. ​ @evn_thou An average job is something that pays minimum wage.
    1
  22. ​ @n111254789 The guy don't grok AI he only knows "Elon Bad"
    1
  23. They become the gatekeepers of open source AI. They lost the upfront monetization but it keeps them in the game for a much more profitable products in the future.
    1
  24. The MIT license is an open source license and Deepseek is under MIT license. So no your wrong Deepseek is legally open source using the MIT license.
    1