3 ms·

>ChatGPT in it's current incarnation will not improve markedly. It's an architectural limitation at this point and more data might smooth out some rough spots,

>ChatGPT in it's current incarnation will not improve markedly. It's an architectural limitation at this point and more data might smooth out some rough spots, but it won't solve the behavior that this chain is describing. Really? how do you know this? Are you an expert in this area and this is actually something experts are talking about or is this just an educated guess? Serious question.

freeone3000 4y ago
I agree with this -- there's no actual mechanism for logical influence at runtime. Logic itself is poorly modelled by studying corpuses -- text sources don't demonstrate it. In addition, GPT is currently saturated with data; it has effectively all of the data. Adding more neurons or more layers might help by adding more memorized facts and exchanges, but I'm of the opinion that our current approach of function approximators fundamentally cannot reason logically. It'll be a while before we can prove or disprove this (the proof techniques used for these are still nascent; a few dozen neurons have been solved...). But I think we'll need another model working in conjunction using a different approach, or even a scaffolding system, as neither the LLM nor the RL on top can implement or enforce logic.
- kilgnad 4y ago
  >In addition, GPT is currently saturated with data; it has effectively all of the data. No this is not true. The reason why chatGPT is better then GPT-3 is because it's trained on additional reinforcement data to determine good and bad answers. They basically outsourced and hired a bunch of Kenyans to do this. If they hire say a computer scientist to reinforce the code and rate the best coded answers then it can improve in that specific specialty. There is ALOT of room for additional sets of this type of data. Are you sure you know this as an expert in this area or you're just agreeing with him as a layman? I'm certainly a layman myself.
  - freeone3000 4y ago
    I've published a few papers and spent the last five years at MSR working in AI. I'm not particularly distinguished, but I believe I qualify as an expert. The RL model made the responses more useful by determining what is and is not a useful reply, then re-running on non-useful replies. It doesn't actually increase the knowledge or the LLM in a meaningful way; it does increase the precision, which is viewed as a form of increased performance? But the RL model cannot inject new facts, and cannot perform reasoning to a greater degree than the LLM can.
    - kilgnad 4y ago
      I see. It seems to me a lot of the issues are precision problems right? chatGPT produces answers that look correct, and if you increase the precision along the correct path in this knowledge space the lookalike answer converges into the correct answer. Let's say we layer an RL model with training that doubles the size of the current LLM. I think most of the counter examples of performance failures of chatGPT in this thread will become nill. Thanks for responding btw. Good to hear the opinion of an expert.
- threeseed 4y ago
  From my layman perspective it needs a rules engine. Over our lifetime we constantly learn rules which constrain the list of expected behaviours we expect in the world. Some of the rules are logical, scientific and mathematical others are specific to individual groups and societies. And those rules should be largely immutable where the user can't convince ChatGPT to chat them at runtime.
  - kilgnad 4y ago
    [dead]

freeone3000 4y ago

I agree with this -- there's no actual mechanism for logical influence at runtime. Logic itself is poorly modelled by studying corpuses -- text sources don't demonstrate it. In addition, GPT is currently saturated with data; it has effectively all of the data. Adding more neurons or more layers might help by adding more memorized facts and exchanges, but I'm of the opinion that our current approach of function approximators fundamentally cannot reason logically. It'll be a while before we can prove or disprove this (the proof techniques used for these are still nascent; a few dozen neurons have been solved...). But I think we'll need another model working in conjunction using a different approach, or even a scaffolding system, as neither the LLM nor the RL on top can implement or enforce logic.

kilgnad 4y ago
>In addition, GPT is currently saturated with data; it has effectively all of the data. No this is not true. The reason why chatGPT is better then GPT-3 is because it's trained on additional reinforcement data to determine good and bad answers. They basically outsourced and hired a bunch of Kenyans to do this. If they hire say a computer scientist to reinforce the code and rate the best coded answers then it can improve in that specific specialty. There is ALOT of room for additional sets of this type of data. Are you sure you know this as an expert in this area or you're just agreeing with him as a layman? I'm certainly a layman myself.
- freeone3000 4y ago
  I've published a few papers and spent the last five years at MSR working in AI. I'm not particularly distinguished, but I believe I qualify as an expert. The RL model made the responses more useful by determining what is and is not a useful reply, then re-running on non-useful replies. It doesn't actually increase the knowledge or the LLM in a meaningful way; it does increase the precision, which is viewed as a form of increased performance? But the RL model cannot inject new facts, and cannot perform reasoning to a greater degree than the LLM can.
  - kilgnad 4y ago
    I see. It seems to me a lot of the issues are precision problems right? chatGPT produces answers that look correct, and if you increase the precision along the correct path in this knowledge space the lookalike answer converges into the correct answer. Let's say we layer an RL model with training that doubles the size of the current LLM. I think most of the counter examples of performance failures of chatGPT in this thread will become nill. Thanks for responding btw. Good to hear the opinion of an expert.
threeseed 4y ago
From my layman perspective it needs a rules engine. Over our lifetime we constantly learn rules which constrain the list of expected behaviours we expect in the world. Some of the rules are logical, scientific and mathematical others are specific to individual groups and societies. And those rules should be largely immutable where the user can't convince ChatGPT to chat them at runtime.
- kilgnad 4y ago
  [dead]