1. It will get better, and in the case of language models, that could have profound impacts on society

why is that a given?

the materials research Deepmind published

these results were extremely flawed and disappointing, in a way that’s highly reminiscent of the Bell Labs replication crisis

3. There are other things being researched that are equally important that get little daylight, such as robotics and agentic AIs

these get brought up a lot in marketing, but the academic results of attempting to apply LLMs and generative AI to these fields have also been extremely disappointing

if you’re here seeking nuance, I encourage you to learn more about the history of academic fraud that occurred during the first AI boom and led directly to the AI winter. the tragedy of AI as a field is that all of the obvious fraud is and was treated with the same respect as the occasional truly useful computational technique

I also encourage you to learn more about the Rationalist cult that steers a lot of decisions around AI (and especially AI with an AGI end goal) research. the communities on this instance have a long history of sneering at the Rationalists who would (years later) go on to become key researchers at essentially every large AI company, and that history has shaped the language we use. the podcast Behind the Bastards has a couple of episodes about the Rationalist cult and its relationship with AI research, and Robert Evans definitely does a better job describing it than I can

this is the general argument in favour of cryptocurrency, with the name changed. you don't seem to have argued that the actual reality of AI we have right now is not the same problem.

I hear what you're saying, but I think it's sort of a motte-and-bailey setup:

Motte: Many functions can be probably approximately learned, even some uncomputable functions

Bailey: Consciousness, appreciation for art, useful laboring, and careful argumentation are learnable functions

There are other kinds of AI use cases, such as alphafold and the materials research Deepmind published

alphafold is mostly pattern matching on known proteins and the other bit, well google very quickly distanced themselves from these shitty results when they learned how shitty they are. i've made a post about it specifically https://discuss.tchncs.de/post/11138402 and i won't rewrite it again

non sequitur