3 ms·

All that money and they’ve been trying to shy away from talk of AGI recently to lower expectations.

by mattgreenrocks 1y ago

CuriouslyC 1y ago
A lot of this is because there isn't a good definition of AGI. Look at Sama's recent interviews, that's how he deflects, along with the statement about the Turing test having ultimately been inconsequential. They have an internal definition of AGI that is "the model can perform the vast majority of economically viable tasks at the level of the highest humans" which isn't the story the investors are expecting when they hear AGI, so they're trying to stay mum to truthfully roadmap AGI while not blatantly lying to capital.
impossiblefork 1y ago
The AI is getting better at things like maths. I recently asked it about iterating the Burrows-Wheeler transform, and it appeared to really understand that. It's not super easy to reason about why it's reversible, etc. and I felt that it got it. This is obviously not AGI, and we're very far from AGI as we can see by trying out these LLMs on things like stories or on analyzing text, or dealing with opposing arguments, but for programming and maths performance is at a level where it's actually useful.
- rhetocj23 1y ago
  My answer to this is - so what? What are the effects in the real economy? Theres probably a good 2-3 years left of runway for substantial progress before things really fall off a cliff. There has to be some semblance of GDP (real) growth.
  - impossiblefork 1y ago
    I think it might be possible to automate development of short programs. I also think it might be possible to reduce the confusion, misunderstandings and cliches, so that the models can work for longer stretches. But people probably expect to get the next version for what they pay in subscriptions now, so I can't imagine much more revenue growth for the model companies.

3 ms·

All that money and they’ve been trying to shy away from talk of AGI recently to lower expectations.

by mattgreenrocks 1y ago

All that money and they’ve been trying to shy away from talk of AGI recently to lower expectations.

CuriouslyC 1y ago
A lot of this is because there isn't a good definition of AGI. Look at Sama's recent interviews, that's how he deflects, along with the statement about the Turing test having ultimately been inconsequential. They have an internal definition of AGI that is "the model can perform the vast majority of economically viable tasks at the level of the highest humans" which isn't the story the investors are expecting when they hear AGI, so they're trying to stay mum to truthfully roadmap AGI while not blatantly lying to capital.
impossiblefork 1y ago
The AI is getting better at things like maths. I recently asked it about iterating the Burrows-Wheeler transform, and it appeared to really understand that. It's not super easy to reason about why it's reversible, etc. and I felt that it got it. This is obviously not AGI, and we're very far from AGI as we can see by trying out these LLMs on things like stories or on analyzing text, or dealing with opposing arguments, but for programming and maths performance is at a level where it's actually useful.
- rhetocj23 1y ago
  My answer to this is - so what? What are the effects in the real economy? Theres probably a good 2-3 years left of runway for substantial progress before things really fall off a cliff. There has to be some semblance of GDP (real) growth.
  - impossiblefork 1y ago
    I think it might be possible to automate development of short programs. I also think it might be possible to reduce the confusion, misunderstandings and cliches, so that the models can work for longer stretches. But people probably expect to get the next version for what they pay in subscriptions now, so I can't imagine much more revenue growth for the model companies.