How close are LLMs to AGI?

This is already AGI, as it is not tailored to a specific purpose (unlike all previous "AI"s were). It can do anything (given multi-modality), it is general. Humans are still better at specific things, but LLMs easily beat humans in generality. For example, GPT-4 can surely write a better poetry than an average human (despite that it is not "the best of the best" poetry), and so on.

It is AGI.

There is no point to rant about "oh but there is no long-term memory, no goal-setting, no planning and such" — it is really easy to augment it with all that, see Auto-GPTs and similar projects. Just loop it through itself and give it an interface to the external world, and it could do anything. Incrementally improve the model, and you'll get to superhuman performance in all tasks.

Tbh, chatGPT caught the world by surprise. Especially the fact that it can do in-context learning so well is a big surprise even for experts in the field. You can see these from many interviews of experts in the fields for example https://www.therobotbrains.ai/copy-of-raluca-ada-popa. So no one really knows how close LLM to AGI. Most of ppl who has been working on AGI thinks that this is no where near. But again, if they are still processing what chatGPT is capable of, how would they know for sure.

This depends crucially on your definition of AGI. LLMs are more like a mathematical function than they are a conscious being. If, in your opinion, AGI could be realized as a input/output function with no changing internal state, like a lens or a lookup table, then we could say LLMs could be an element of AGI. But if, like many of us, you believe an AGI needs a changing internal representation of the world, and an ability to mull over prior knowledge and incorporate new inputs, then LLMs are at best only a useful component of AGI. Maybe like the retina of the human eye plays some role in human visual intelligence.

The fact that LLMs appear so intelligent to humans is really a reflection of our inability to imagine effects of scale. We can understand simple linear predictions as trivial calculations, but when language-based pattern discovery is many layers deep and those patterns are combined in nontrivial (but non-intelligent) ways, we project intelligence onto the result.

https://www.youtube.com/watch?v=HKI8z3Nm5hg

In my opinion, based on what I have seen and understood thus far, not close at all.

But they can, in some scenarios, give the appearance of being close, which might be enough to be useful for some AGI purposes, whatever those might be.

LLM's are about as close to AGI as an apple is to a fully matured apple tree.

LLM's are one cell

come on