3 ms·

> This will make it hard/impossible for new vendors to come into the market and only established companies will get to play, and charge, for LLMs. I don't unde

> This will make it hard/impossible for new vendors to come into the market and only established companies will get to play, and charge, for LLMs. I don't understand how you leap from "US govt. decides who gets to use GPT-5" to "limit new vendors from competing with OpenAI". Can you walk me through the logic?

freakynit 4mo ago
These frontier models are now good enough that they can assist heavily with new optimizations for future models, including, code them. Restricting their usage to a few companies takes away that advantage away from other companies, thereby, limiting new vendors from competing with OpenAI.
- andsoitis 4mo ago
  > Restricting their usage to a few companies takes away that advantage away from other companies, thereby, limiting new vendors from competing with OpenAI. wait, are you saying a competitor needs access to an OpenAI model in order to build a competing model?
  - m4x 4mo ago
    I agree with your point, but to play devils advocate: doesn't a competitor arguably need access to these beyond-frontier models to even become an effective competitor to OpenAI and Anthropic?
    - freakynit 4mo ago
      "Assisting", not "one-shotting", a better model.
  - freakynit 4mo ago
    Nothing new in that. Everyone pays to someone or the other to make their own product/life better. Most of the times these products do not compete, sometimes, they do. How many times you believe duckduckgo would have google'd stuff just to create a competitor to Google itself? I believe thousands of times.. could be more. OpenAI cannot claim the code their models produce as their own since their own models used codes from public internet during training to produce new code.
    - andsoitis 4mo ago
      That’s incoherent.
      - freakynit 4mo ago
        What part?
- 59nadir 4mo ago
  Do you have any sources for the claim that LLMs meaningfully help in the production of LLMs?
  - freakynit 4mo ago
    Read Anthropic's own blog post released sometime around last week.
    - Planktonne 4mo ago
      Do you have a source that isn't known for consistently over-stating their model's capabilities?
      - fc417fc802 4mo ago
        Common sense? Each model iteration is clearly more capable than the last and at this point the benefit to researchers of all stripes is self evident.
      - freakynit 4mo ago
        You do realize this statement about "overstating model capabilities" has nothing to do with AI assisted AI development, right?
        Planktonne 4mo ago
        It absolutely does. Because 'AI-assisted AI development' is one of the claimed but not verified capabilities.
        freakynit 4mo ago
        You could produce a completely fcked up model using AI, and also a completely awesome one using AI. So, the "overstating capabilities" part really have nothing to do with this discussion. As for the claim part around AI being used to assist further AI development, I have already told to read Anthropic's blog post. You can take help from chatgpt, if needed.
fragmede 4mo ago
> the U.S. government would initially approve who gets access to its latest new release while AI companies and the administration work out a longer-term plan for regulation on the sector. So you're a new vendor with a GPT-5.6 or Mythos class model. How do you suppose the regulations are going to work? First you need to get on the list of companies that are allowed to release models, and then you need to have a whole system for limiting access. Both are going to be hugely expensive, on top of training new models already being insanely expensive to begin with. Thus, it's not a legal limit, but a real practical limit because it's too expensive. If only OpenAI and Anthropic and Google can afford to jump through the hoops, they've effectively gotten to "limit new vendors from competing with OpenAI".
- andsoitis 4mo ago
  > So you're a new vendor with a GPT-5.6 or Mythos class model. How do you suppose the regulations are going to work? First you need to get on the list of companies that are allowed to release models, and then you need to have a whole system for limiting access. Both are going to be hugely expensive, on top of training new models already being insanely expensive to begin with. So your thesis is that someone has the resources to create a model at the level of Mythos or GPT 5.6 but bot have the resources to jump through the concomitant legal requirements? Surely that’s unlikely?
  - fragmede 4mo ago
    In a perfectly spherical cow sort of way, everything is fungible and if you can get the resources to train a Mythos-grade model then you'll also have the resources to jump through all the legal hoops. Regulatory capture is a tale almost as old as capitalism though. History bears this out. Adding barriers to entry further entrenches the incumbents.

freakynit 4mo ago

These frontier models are now good enough that they can assist heavily with new optimizations for future models, including, code them. Restricting their usage to a few companies takes away that advantage away from other companies, thereby, limiting new vendors from competing with OpenAI.

andsoitis 4mo ago
> Restricting their usage to a few companies takes away that advantage away from other companies, thereby, limiting new vendors from competing with OpenAI. wait, are you saying a competitor needs access to an OpenAI model in order to build a competing model?
- m4x 4mo ago
  I agree with your point, but to play devils advocate: doesn't a competitor arguably need access to these beyond-frontier models to even become an effective competitor to OpenAI and Anthropic?
  - freakynit 4mo ago
    "Assisting", not "one-shotting", a better model.
- freakynit 4mo ago
  Nothing new in that. Everyone pays to someone or the other to make their own product/life better. Most of the times these products do not compete, sometimes, they do. How many times you believe duckduckgo would have google'd stuff just to create a competitor to Google itself? I believe thousands of times.. could be more. OpenAI cannot claim the code their models produce as their own since their own models used codes from public internet during training to produce new code.
  - andsoitis 4mo ago
    That’s incoherent.
    - freakynit 4mo ago
      What part?
59nadir 4mo ago
Do you have any sources for the claim that LLMs meaningfully help in the production of LLMs?
- freakynit 4mo ago
  Read Anthropic's own blog post released sometime around last week.
  - Planktonne 4mo ago
    Do you have a source that isn't known for consistently over-stating their model's capabilities?
    - fc417fc802 4mo ago
      Common sense? Each model iteration is clearly more capable than the last and at this point the benefit to researchers of all stripes is self evident.
    - freakynit 4mo ago
      You do realize this statement about "overstating model capabilities" has nothing to do with AI assisted AI development, right?
      - Planktonne 4mo ago
        It absolutely does. Because 'AI-assisted AI development' is one of the claimed but not verified capabilities.
        freakynit 4mo ago
        You could produce a completely fcked up model using AI, and also a completely awesome one using AI. So, the "overstating capabilities" part really have nothing to do with this discussion. As for the claim part around AI being used to assist further AI development, I have already told to read Anthropic's blog post. You can take help from chatgpt, if needed.

fragmede 4mo ago

> the U.S. government would initially approve who gets access to its latest new release while AI companies and the administration work out a longer-term plan for regulation on the sector. So you're a new vendor with a GPT-5.6 or Mythos class model. How do you suppose the regulations are going to work? First you need to get on the list of companies that are allowed to release models, and then you need to have a whole system for limiting access. Both are going to be hugely expensive, on top of training new models already being insanely expensive to begin with. Thus, it's not a legal limit, but a real practical limit because it's too expensive. If only OpenAI and Anthropic and Google can afford to jump through the hoops, they've effectively gotten to "limit new vendors from competing with OpenAI".

andsoitis 4mo ago
> So you're a new vendor with a GPT-5.6 or Mythos class model. How do you suppose the regulations are going to work? First you need to get on the list of companies that are allowed to release models, and then you need to have a whole system for limiting access. Both are going to be hugely expensive, on top of training new models already being insanely expensive to begin with. So your thesis is that someone has the resources to create a model at the level of Mythos or GPT 5.6 but bot have the resources to jump through the concomitant legal requirements? Surely that’s unlikely?
- fragmede 4mo ago
  In a perfectly spherical cow sort of way, everything is fungible and if you can get the resources to train a Mythos-grade model then you'll also have the resources to jump through all the legal hoops. Regulatory capture is a tale almost as old as capitalism though. History bears this out. Adding barriers to entry further entrenches the incumbents.