If by AI you mean the current LLM's using BigData like OpenAI then I doubt this will happen. All iterations of that confidently spew lies, half-truths and highly biased output. Should someone hire that thing to write software or control anything important then I would expect anything it works on to go horribly sideways, especially if the assigned project is complex.
But I could be wrong. If I am wrong then my hope is everyone has enough free time to to watch all the episodes of M.A.S.H. while simultaneously retraining in something else, whatever the next something else may be.
Have any of them successfully completed any complex projects to date that would have required critical thinking on the part of a human? Are we using any software or complex machinery created by AGI's yet? Do we own any products created by them or is everything they do still PoC/experimental at this point? Are any of the AGI prototypes currently being adopted and utilized by retail or industrial sectors?
The reason I ask is that AFAIK everything I have ever used was created either by humans and/or simple robots that have existed for quite some time.
in that it hijacks people's perception of others and desire for closure to make people perceive it has a personality when it really doesn't. In particular it seems to have a hypnotic ability that comes out of its "highest probability" approach to text generation and how HFRL trains it to write text that pleases people -- ChatGPT's bullshitting goes right past many people's critical thinking capability and if anything is dangerous about it, it is that.
Those people hypnotized by it inevitability think that adding more parameters and more processing power to it will make it overcome any and all limitations but that's likely to be wrong, instead it will reach some asymptote where you feed it more resources and it only improves marginally. That's the most dangerous place in technology development, it's like Uber putting in $25 billion into subsidizing half priced taxi rides and thinking the next $1 billion will make it a profitable business.
On the other hand if you looked at arXiv instead of blog postings on HN you'd see that people frequently do "zero-shot" prompting, get the same 70% accuracy that the blog posters do, then fine-tune a similar model on a modest sized training set and get 95% accuracy. The way I'd put it is that there are 13 users of transformer models listed here
and y'all are obsessed with just one of them. With fine-tuning a model with fewer parameters and a modest amount of training data will thoroughly trash ChatGPT at specific tasks most of the time. The people who download models from huggingface and learn to use them will be putting applications in front of customers while people messing around randomly w/ GPT-N will be forever pushing bubbles around under the rug.
However, that kind of generative model can be a lot more powerful when it is combined with systems that can do the things it can't do. For instance having it "shell out" to Python or Wolfram Alpha. People are perplexed when they give it instructions which it inevitably quits following after a while, but in addition to not having the correct structure to obey programming language like instructions, the instructions will inevitably fall out of the attention window and disappear and it's no wonder it loses track. See
Now, if there was some program surrounding ChatGPT that implemented executive control (what that guy really wants) that would rewrite the prompt, keep the instructions in the window (much shorter because they only concern ChatGPT's area of competence, and move all the things ordinary software does well (scorekeeping) into ordinary software.
We will eventually see things that work like an SAT/SMT solver or theorem prover that is integrated w/ an LLM, much like the way Alpha Go integrates "old AI" heuristic search with a neural network heuristic.
Real advances in the capability of those chatbots will involve approaches that avoid the structural weaknesses of chatbots, not an appeal to Moore's Law. The asymptote of naked chatbots is already in sight I think, but the asymptote of hybrid systems is further away in both time and capability.
But I could be wrong. If I am wrong then my hope is everyone has enough free time to to watch all the episodes of M.A.S.H. while simultaneously retraining in something else, whatever the next something else may be.