Pure LLMs haven’t been the only game in town for state of the art models for a long time. In modern AI workflows, LLMs form one stage of a multi-stage process. I’d analogize it to PageRank vs Google Search, PageRank has been a foundational algorithm in Google Search since inception but Google has augmented PageRank with significant pre & post processing steps & introduced new modalities where pagerank is far less important like image search.
IMHO, far more important than paying attention to the algorithm is paying attention to the data. LLMs got an early win by ingesting a giant bolus of data which was all of the public internet. There’s likely not any corpus of data that cheap and massive that LLMs will ever get to train on again and so new data will be either extremely more expensive or lower quality. Up til now, AI doesn’t know anything about the world, AI only knows about what humans tell it about the world. New innovations in AI will come from, eg: hooking it up to every security camera footage in the world and building algorithms to process and understand the world from unprocessed, raw video or giving an AI robotic manipulators and vision systems so it can derive physics from first principles.
Sure, but is there any widespread (!) example where it isn’t both essential and ultimately limiting to the whole stack? A car might not be just an engine, but without an engine the wheels aren’t getting you anywhere.
Yeah, all the image and video generation models aren’t LLMs because they aren’t languages. A lot of newer AI applications don’t have anything to do with languages anymore, eg: dealing with building environmental optimization or other 3d space tools, industrial manufacturing & supply chain optimization etc.
A recent startup I saw is tapping into security cameras all around a construction site and real time notifying people when the AI believes an OSHA violation might have occured. They’re building their own custom models, no LLMs anywhere near it.
OK, I was under the impression that any text-to-image model is still going to use a transformer-based LLM in decoder mode to parse the prompt, like e.g DALL-E does; if that isn’t the case, I’d appreciate clarification.
As for those other applications, nothing about my point is lost broadening to attention-based transformer models rather than LLMs narrowly understood. Although I still admit I would be surprised if, on balance, the bulk of the value proposition of AI for the economy isn’t in the promise of LLMs to make office workers more productive (or replace them entirely).
I mean that’s horrific and all, and hopefully would itself be a violation of workplace regulations at least here in Germany, but I struggle to see that the current AI economy could be supported on such a use case.
I have NOT read the whole thread but will report my recent experiences.
I was NOT a fan of AI but of course get Google’s Gemini AI “for free” when I Search.
One of its stupider results equated the 18th century with the 1800’s. I suppose a lot of
High School students make this mistake and its LLM was just copying their ignorance.
But AI is getting better and I’ve been experimenting with ChatGPT! Archimedes’ test whether
his king’s crown was pure gold is an interesting matter where the wrong answer is almost universally given! I learned of this a few years ago, and that Galileo finally figured out Archimedes’ ACTUAL solution. ChatGPT (VERY paraphrased) produced “Blah blah blah … blah” and finally “But that’s the wrong answer. Would you like the right answer?” After further coaxing,
ChatGPT told me what I already knew and added that Jerome Cardan wrote up the correct solution half a century before Galileo was even born.
Just now I asked it about a Wordle I had just solved. (I wondered what possible words I had overlooked.) I will spoiler-tag the exact question in case a Wordle player reads this.
I could solve this myself with a long pipe of ‘grep’ and ‘grep -v’ that would take me 2 minutes to type and 2 seconds for computer to solve. (I do NOT do that when playing Wordle.) ChatGPT spent several minutes, telling me the names of various scripts and .exe’s it was running. (those mentions are not saved to the transcript.
Finally, after several minutes and LOTS of scripts ChatGPT
produced "Babble Babble Babble … Babble
Short answer: LIMBO — it’s the only Wordle answer.
Followed by much more babble babble babble.
Then:
You said:
Did you overlook BROIL?
ChatGPT said:
You’re right — I missed BROIL. My earlier claim that LIMBO was the only possibility was a mistake.
More babble babble and finally, the info I was actually curious about:
The complete set of answers from the official Wordle/NYT answer list that match those two feedbacks is:
BROIL
LIMBO
IGLOO
LINGO
I am NOT impressed that it couldn’t find the trivial but long pipe of ‘grep’ I’d have used. Or at least after HUGE cogitation achieved a correct answer.
But ChatGPT and I are VERY polite to each other, and it did guide me through the procedure to use some of its features.
Will AI take over the world in a few years? Or is NVIDIA in a bubble that will cost investors Trillions? Or BOTH???
I dunno; just thought the anecdote might be interesting.
I’ve been reading a number of articles theorizing where we are rapidly reaching a point where the AI business model will no longer scale. AI and LLMs have reached a point where they are “pretty good” but really not good enough to wholesale replace everyone’s job as promised (as supported by anecdotal comments in this thread). Improving AI requires more investment and higher costs for data centers and the energy to run them. Those costs have apparently not been passed on to the consumer yet.
So there does seem to be growing concern about all this crashing down at some point with companies left holding the bag having shed all their institutional knowledge in the people they laid off.
Also anecdotally, my social media feeds seem to be flooded with so-called experts and coaches offering their services to get people up to speed so they don’t get left behind racing to be the first person to build a billion dollar AI business with one employee!
Meta’s (META) chief artificial intelligence scientist, Yann LeCun, plans to leave the company to launch his own AI start-up, marking a major shift inside Meta
[…]
LeCun has publicly disagreed with Zuckerberg’s heavy reliance on LLMs, calling them “useful but fundamentally limited” in their ability to reason and plan like humans. His upcoming start-up will extend his FAIR research into “world models” that could ultimately enable machines to think more like people.