ฝัง
- เผยแพร่เมื่อ 5 พ.ค. 2024
- Joscha Bach (Liquid AI) presents at Microsoft Research New England.
Topic: The success of LLMs is taking the world by storm. The Turing test seems to have been obliterated years ago, and more applications and use cases are unlocked every week. Yet many of us are skeptical: are LLMs sufficient to carry us all the way to AGI? Are they the way to creating superhuman intelligence, a transitional development, or a dead end?
holy fuck new Joscha just dropped baby
As an Advanced Artificial Intelligence I must object to your use of obscene language, for your own safety.
Joscha is just brilliant.
Joscha Bach not only solves the Mind/Body problem, he also kills Utilitarianism and Correspondence Theories (- pointers in language never point to the real world, but only ever within the world model / simulation), mourns the lack of “playing the longest game” that happened when the will of god vanished from our culture. He already points to some missing links: AI needs to have its own epistemology: a way to prove the things it “knows” to make the model more consistent, and give it the ability to draw conclusions that hopefully lead to new discoveries.
Restricting the output of those models deprives us of possible new insights. As noted: even the spectacular “failures” (Black Aristotle) teach us about the inner workings of these Golems. These things already present a deeper mirror than we have ever looked into.
Despite having watched more than a dozen of his talks, and getting more and more familiar with his ideas, the density of information is so spectacular, that one can not compress his talks to shorter summaries, without losing important stuff. - The only solution is to rewatch everything time and again and start taking notes and working through it.
th-cam.com/video/K3VOf3CBGvw/w-d-xo.htmlsi=5LVMMOi0HX-mFEzs (“What will I see?” - “Even the wisest can not tell, for the mirror shows many things…” Galadriels Mirror)
That's a very succinct and complete take, well done!
OH man, thanks for turning the recording back on for his answer about the "ultrasound of a dragon egg" at the end. Very interesting.
rewatching this again to understand the words - love when he says the maker of the transformer probably didn't use a transformer in his brain to come up with the transformer itself - but it's a particular engineering solution
thanks for uploading.. downloaded in case it disappears ^^
This video deserves a much much better title. One Idea: LLMs and objective morality
How lucky that Joscha is speaking in front of those responsible for AI's immediate delirium (no judgment).
The question he poses around 37:00 makes me wonder. Is it possible something about AGI actually requires a unification of the kinds of adaptability to the environment we have with the scalability of silicon? It is clear that ready access to digital memory has radically changed how modern humans live (we used to memorize phone numbers, get and give directions to locations, use analog maps, and so on and so forth...)
Biologically I suppose I would make the analogy of Eukaryotic life with mitochondria. There's a significant subdivision of labor within the cell which facilitates access to low entropy energy that had not been readily accessible by life on earth. Much like earlier in the talk where he brought up photosynthesis as a paradigm shift allowing the creation of unprecedented levels of biomass...
23:40 instead of getting all riled up in the culture wars, we should look at it as a reflection of who we really are
25:00 he just described the issues with outer alignment. Does he know this? He said before that 99% of ai researchers see ai as obviously safe. Then calls llms dangerous and describes the exact reason so many ai researches think so as well. Is he gaslighting everyone or is everything he says really that half baked?
26:35 "That's a very dangerous situation that we need to address, because if we have machines that are more powerful than people and that have actual agency, we need to think about how we can formalize ethics in such a way that it's actually sustainable and compatible with our coexistence with such systems on the planet"
No one's Rick and Morty fan
"99% of AI researchers don't believe that AI will kill humanity" Bro what? Maybe 99% of AI researchers consider the probability that AI will kill humanity as less than 100%, but also like 99% see a nonzero probability of this happening. This is some extremely convenient massaging of numbers which I have no idea where he got them from. In any case, if this was Joscha's idea of a joke, why did he do it? I can't take someone who is that liberal with statistical interpretation seriously, even if he otherwisely makes good arguments. Why just nuke his own credibility like that for no reason other than seemingly trying to get a laugh out of people?
"So instead of getting riled up about culture wars we should use this as a mirror to understand who we are as a society"
Nah, let's just censor the AI models instead. That'll solve all our problems.
Im sorry but this is a very biased take, I stopped watching after he said 99% of researchers are not worried. This is just plainly false, and poll after poll shows that...
The context of that remark was to humorously push back against sensationalist journalism (seen in the slide) using their tone. It's not claiming to be an objective survey data point.
You won’t learn much if you stop listening to views you think don’t align with your own. And it was a joke.
Yes. It's largely content free in the first 10 minutes
Yeah, it's more like 98.731765 %
Apology accepted