I'm surprised no one has mentioned GPT-4 and the wild speculation about what capabilities it will have. The blog post we're discussing says things like:
> We are excited to carry the lessons from this release into the deployment of more capable systems
so I immediately assumed that ChatGPT is basically a ruse for generating training data that will be used to ensure GPT-4 passes the Turing Test.
It's unlikely that OpenAI would have set out with that as their primary goal for GPT-4, but it's been two and a half years since GPT-3 came out (twice the time span between GPT-2 and GPT-3) so it's possible they realised while developing GPT-4 that it would be theoretically capable of passing such a test, and if they didn't aim for that milestone, then one of their competitors would beat them to it.
I know that a lot of commentators and researchers are becoming increasingly cynical about the value of passing a Turing Test, especially as it requires the AI to be capable of deception (which has various ethical and safety implications), but I still think that OpenAI would be tempted to try to pass it if they thought they were close with GPT-4. In particular, I think that the incentives point towards OpenAI creating a carefully stage-managed version of the test, using unsophisticated judges, such as school children (or rogue Google engineers?).
Even with all the caveats of such a test, I think GPT-4 could be meaningfully described as "AGI" if it ends up being able to hold interesting conversations and solve mathematical problems and produce digital art and play Atari games and write code snippets, all using the same internal model.
I suspect they will conquer multiple modalities, or at least make decent progress there. They have a good foundation model for audio now (Whisper), and had great results with text/image on both CLIP (search) and DALLE (generation). There are also lots of modes throughout NLP that can be articulated (translation between languages, formats, "downstream tasks"). I bet GPT4 makes use of audio/image and text.
But no, I don't think it will be "general AI". It's more like the semantic engine we develop as humans to be able to strategize with "seemingly unrelated" concepts. We attach some arbitrary meaning to it, then form associations to other things we have seen previously, if any exist. GPT3, DALLE, CLIP, etc. make sense to use a similar sort of "backend" of semantics. But, they crucially don't possess the ability to do short or long term "planning", and as a sibling comment mentions - they aren't "online" after tranining and can't do gradient descent "on-the-fly". Proposals for AGI tend to include a "module" for planning (although no one has articulated how to do that, or what it even means), as well as the ability to run end-to-end gradient descent on-the-fly (in a finite amount of time).
The systems we have currently do seem to "generalize" to the distribution of data rather than the data itself. Indeed, this is part of what makes them useful - they can be used to create content not previously seen that is still accurate to some known modality. "an armchair in the form of a [blarn]". The human still does all the planning/prompt-engineering, however.
For all the potential power of GPT-4, If it cant "learn on the go"... can't permanently incorporate new knowledge based on current conversations/inputs... remember things and use those as part of generating responses to things, going beyond the extent of its original data... then its not going to be much of a general intelligence.
Looping that is trivial. What if it convinces someone to run some code that hacks his computer, escape the sandbox (by bootstrapping itself) and running in "live" loop?
> We are excited to carry the lessons from this release into the deployment of more capable systems
so I immediately assumed that ChatGPT is basically a ruse for generating training data that will be used to ensure GPT-4 passes the Turing Test.
It's unlikely that OpenAI would have set out with that as their primary goal for GPT-4, but it's been two and a half years since GPT-3 came out (twice the time span between GPT-2 and GPT-3) so it's possible they realised while developing GPT-4 that it would be theoretically capable of passing such a test, and if they didn't aim for that milestone, then one of their competitors would beat them to it.
I know that a lot of commentators and researchers are becoming increasingly cynical about the value of passing a Turing Test, especially as it requires the AI to be capable of deception (which has various ethical and safety implications), but I still think that OpenAI would be tempted to try to pass it if they thought they were close with GPT-4. In particular, I think that the incentives point towards OpenAI creating a carefully stage-managed version of the test, using unsophisticated judges, such as school children (or rogue Google engineers?).
Even with all the caveats of such a test, I think GPT-4 could be meaningfully described as "AGI" if it ends up being able to hold interesting conversations and solve mathematical problems and produce digital art and play Atari games and write code snippets, all using the same internal model.