Decels don't exist. Most people in EA I talk to are basically e/acc, but they aren't dumb and don't want to die accelerating. Technological progress is pretty good overall. We are just getting to a power level where we should be cautious.
@wesg52
@tegmark
Could someone explain to me where the jump from “Llama memorized a bunch of lat longs” to “Llama has an understanding of space” occurs?
@ylecun
@tegmark
Why does it matter who has more scientists on their side? This kind of thing should be settled with technical arguments and research collaboration.
@ylecun
"This model didn't kill anyone so let's 10x the power and release the next one. The pattern must hold!" This is just blatantly bad logic. While I don't think that lamma 2 or probably 3 will be catastrophic for the world, your logic sure will be.
@trobuling
@wesg52
@tegmark
That is a good point. I guess they're analyzing the network to find the neurons that correspond to an understanding of space. I don't understand yet why having neurons that "encode spatial and temporal coordinates" is equivalent to "an internal world model"
Today was my last day as a software engineer at
@SupplyPike
. I'll be in the MATS program this summer working with
@bshlgrs
! I'm super excited to start working on AI safety fulltime!!
@ProfNoahGian
I think EA looks at AI similar to nukes. We don't want to give everyone on earth a nuke. I think it is hard to argue that current AI is nuke level, but maybe next year? What part doesn't resonate with you: should we give everyone nukes or does AI != nukes?
A common argument against worrying about AI safety is "there is no data / empirical evidence that a superintelligence will wipe us out, so we shouldn't worry". This is true. But there isn't any data to the counter either. No one has observed what an ASI would do because it
@lucidcoderio
Ever? I can see 10 years but forever is a long time. What special quality do you think humans have that computers will never be able to replicate?
@littIeramblings
I don't think it is productive to blame the CEOs, they are all in a complicated prisoners delima racing towards AGI. If one CEO says "I'm going to stop, this is bad", the shareholders will just replace them. I think the system of incentives is what needs to change.
i ran the "needles" eval originally for claude 2.1. some observations:
1. claude 3 smashes a more robust version that randomizes over different documents, needles, etc
2. claude 3 is tired of your shit, asking us to let it get back to real work
@RuxandraTeslo
Yes, there are many researchers working on this exact thing because people keep asking about this. I think this podcast is a good introduction to the ideas if you want to learn more
@Aspie96
Yes, great point! I think the PauseAI position would say, "Pause until we understand AI, then accelerate." It is a trade-off of stopping one area of research (building uncontrollable AI) and working on another (AI safety).
Bar idea:
Have a depressurized room that has half the oxygen as normal. You get drunk twice as fast. Once you leave you are back to normal oxygen and can drive home.
I think the only truly scary part about AI is the alignment problem i.e. AI we can't control. Misuse by people with controllable AI isn't as scary to me. Tell me why I'm wrong
After listening to
@ylecun
on Lex Friedman I think my biggest disagreement with him is how confident he is. He thinks he knows exactly how the future will play out with little uncertainty.
I just talked to
@clattner_llvm
for a while at
#TCDisrupt2023
. First time ever talking to someone that was on Lex Friedman! I'm super excited about the work he is doing on Mojo. It seems like the correct way to make a new language. Meet developers where they are at.
Why did we land on file systems being hierarchical? Makes information hard to sort when a tree structure is forced. Is there any good "tagging" file system? I want to be able to see all of my documents and give them tags like "notes", "2023", or "taxes" and filter by them.
Another two safety researchers leave: Ilya Sutskever (co-founder & Chief Scientist) and Jan Leike have quit OpenAI.
They co-led the Superalignment team, which was set up to try to ensure that AI systems much smarter than us could be controlled.
Not exactly confidence-building.
@tegmark
Is it possible to get past people caring about which AI scientists believe what? I hate that AI scientists are talking about what others believe instead of laying out technical arguments more. We need more Munk debates!
@lexfridman
@sama
Hey Lex. I think a good question would be something like "do you think it is moral for a couple individuals to build AGI without the rest of humanity's permission?"
@tegmark
Amazing! Great work peaking into the LLM's mind. I didn't imagine we'd be able to see this. Did you ever see a case where the LLM knew something was true but still stated it was false? It would be fantastic if we could detect if an LLM was knowingly lying and account for it.
@lux
@ylecun
Yeah, I basically agree with this. I'm not worried as much as bad actors with an LLM. I'm worried about good actors not being able to control their LLM.
Say hello to GPT-4o, our new flagship model which can reason across audio, vision, and text in real time:
Text and image input rolling out today in API and ChatGPT with voice and video in the coming weeks.
RazorHack Cyber Challenge was a 3-day event that brought together cybersecurity enthusiasts of all levels to participate in a capture the flag competition. Check out some snapshots from the event!
Photos courtesy of Kate Kettler
#UArkEECS
@Converse555
@trobuling
@wesg52
@tegmark
When I think of a my world model, I think about how I can visualize my room and place objects, move around it, and can perform mini experiments in it. Maybe this is equivalent, but I don't quite see it.
@ylecun
I've had a really hard time understanding your logic in your tweets Yann. Do you have a writeup anywhere where you discuss each of these points in detail and address current safety arguments?
Started moving my apps over to
#bunjs
and it is much faster! Excited to get rid of the old clunky tools like nodemon dotenv express ws and jest. Downside I'm running into is the monorepo support isn't quite there yet to replace yarn as my package manager.
@bunjavascript
@AnthropicAI
Seems like sparsity is a great way to have more understandable models which is useful and exciting! One question, does this lead to "better" models? Or does the company that makes a "polysemantic" model outcompete the one that doesn't?