It struck me as a weird article. For one thing it's not a healthy or rational "market" at this time in that investment is flowing into it far beyond the actual revenue or realized benefits. It continues to defy all the principles of microeconomics, thus I'd avoid using that word in that situation.
It's also a little weird to me that he goes to lesswrong for answers (I'd trust copilot more) and also that lesswrong caught up with my opinions from a year and a half ago.
That is, OpenAI has no sustainable advantage or "moat" for ChatGPT-4. Adding more coefficients won't improve performance dramatically because LLMs are structurally inadequate to solve many problems. (Read an introductory algorithms book) On the other hand we didn't understand how these things work at all a year and a half ago (we understand a lot more now but less than we will in another year and a half) and there is a huge amount to do in terms of space/time performance optimization. So it was clear to me that models from the low end were going to catch up.
The "structurally inadequate" problem can be attacked by adding elements to the architecture such as tools (let it write a Python program if it needs to sort a list) and RAG but the book Godel, Escher, Bach points out that sticking together multiple problem solving strategies won't fundamentally solve certain problems that are intrinsic to problems such as the intractability of logic + first-order arithmetic.
1 comment
[ 2.5 ms ] story [ 10.0 ms ] threadIt's also a little weird to me that he goes to lesswrong for answers (I'd trust copilot more) and also that lesswrong caught up with my opinions from a year and a half ago.
That is, OpenAI has no sustainable advantage or "moat" for ChatGPT-4. Adding more coefficients won't improve performance dramatically because LLMs are structurally inadequate to solve many problems. (Read an introductory algorithms book) On the other hand we didn't understand how these things work at all a year and a half ago (we understand a lot more now but less than we will in another year and a half) and there is a huge amount to do in terms of space/time performance optimization. So it was clear to me that models from the low end were going to catch up.
The "structurally inadequate" problem can be attacked by adding elements to the architecture such as tools (let it write a Python program if it needs to sort a list) and RAG but the book Godel, Escher, Bach points out that sticking together multiple problem solving strategies won't fundamentally solve certain problems that are intrinsic to problems such as the intractability of logic + first-order arithmetic.