- https://the-mad-pirate-prints.teemill.com/
@TheMadPirate @icedquinn @binkle >I think it will get more accurate with time though.
Assuming model collapse doesn't happen.
@gentoobro @icedquinn @binkle @TheMadPirate Not all LLMs are VC funded. Some LLMs like GPT4All are FOSS and user funded.
@TheMadPirate @icedquinn @binkle @gentoobro I've noticed this exact problem with ChatGPT which is why I don't trust it as a brainstorming tool.
That's because there are far more publish conversations of people agreeing than disagreeing. The latter you would only find on forums or novels, not published in nonfiction books or articles. There's little point publishing arguments in most situations, as only the final result is useful.
Of course, LLM's have no concept of truth or falsehood (or "concepts" at all for that matter) so if they were trained heavily on arguments then they would just constantly argue with you regardless of reality.
VC's are funding the infinite construction of massive datacenters to train the new whiz-bang LLM chatbot that will finally make humans obsolete and therefore somehow make money off of people who now have no jobs.
That is where the WEF "we just don't need most people" and Georgia Guidestones "maintain 500 million" philosophies come into play.
Huge population collapse is almost certainly planned, but first they have to get rid of the hardiest, most independent people who could fight back and survive.
i do indeed download and try them when people go on fanboying it. i've tried mistral, mixtral, and deepseek, and then tried them on. i think i may need to redownload mixtral because that one tends to do the one thing i can trust them to do (summarize news.)
deepseek doesn't follow instructions very well (it just tries to sidestep anything vaguely lewd, and when asking it for short or one sentence summaries of an email it gives me a list of "the summaries of all the emails" which is both wrong because its just one mail, and then wrong because its a list and not a sentence)
other people are free to expediently shoot themselves in the feet though. in fact everyone who opposes me should cripple themselves as much as possible to make my life easier. which is what these things do, as surveys are showing that they both produce worse code and the users are becoming dependent mental midgets.
deepseek came up with some odd things when i asked it for neogothic brutalist buildings. but deepseek fails to do shit thats actually useful (summarizing shit for me.)
there are some uses for these things but they're as cogs not as the whole product. and for the stuff they are actually capable of you'd be better off using a mamba model over a transformer anyway.
@icedquinn @binkle @TheMadPirate @s8n I find ChatGPT useful for generating HTML templates without including any third-party JS or CSS, but I wouldn't trust it to generate an entire site.
these things only work in a world of no consequences, where you can freely be wrong as much as you want and there's no relationship between success and accuracy
They're chatbots, and not much more.
VC helicopter money collapse will happen first. They're already super jumpy.