Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.
This is a linkpost for https://drive.google.com/file/d/1wzHohvoSgKGZvzOWqZybjm4M4veKR6t3/view
Mentioned in
Yann LeCun on AGI and AI Safety
6th Aug 2023
29trevor
24Thomas Kwa
19Chris_Leong
1titotal
2RobertM
6Steven Byrnes
5cubefox
5Roman Leventov
4Ilio
1Leo P.
3Chris_Leong
3Leo P.
1mdtheory
New Comment
13 comments, sorted by Click to highlight new comments since: Today at 5:50 PM
It's important to note that Lecun and Andreessen (Facebook/Meta board member) are well-established to be currently conducting an infowar against AI safety- they're pretty committed to accelerating AI and making sure their company, Facebook/Meta, is the company that gets maximum advantage in that race; at all costs, and by any means. As another example, releasing SOTA open-source AI models in order to advance open-source AI and leverage their dominance there, even though open-source AI allows companies in China and other countries to develop better engineering expertise to compete with American AI companies.
Currently, Facebook seems to be competing against OpenAI, Deepmind, and Anthropic for influence over AI policy in DC. Since Facebook/Meta's closed-source systems are lagging, their strategy seems to be using AI safety as a sacrificial lamb in order to appeal to pro-american-innovation norms that have been very dominant in DC for a while. Their edge there is being the company that credibly committed to steering clear of all that confusing AI safety hogwash.
Obviously, there's much more to it than what I've said here, like investor confidence and other complex and controversial factors which I'm not currently willing to talk about in a public comment. There's seriously a lot going on with Facebook/Meta and AI, you could spend years researching that rabbit hole and you'd never stop finding things worth finding.
But if anyone decides to write a list of detailed explanations of why Lecun and Andreessen's arguments are obvious horseshit, you should expect them to follow up by throwing substantial time and money into generating more horseshit counterarguments tailored around your arguments; specifically to look good to policymakers in DC and other influential people who can't or won't go into details about the problem itself.
The nice thing is that Lecun and Andreessen seem to be unwilling or unable to lie about superintelligence being feasible at all, they have to admit it's a real situation, so they can't just do the usual thing where they appeal to common sense and say it's all a sci-fi grift. AI safety and Facebook/Meta are debating on even ground here, the vast numbers of people who can't/won't entertain the idea of vastly-smarter-than-human AI aren't going to be participants in either side.
It really shouldn't surprise people to see high-level figures from Facebook/Meta, of all places, being really well-versed at information warfare; but many people are still approaching this like an honest debate, which it stopped being a long time ago.
3
1How do we know this? If it is "well-established", then by whom and what is their evidence?
3This seems like an epistemically dangerous way of describing the situation that "These people think that AI x-risk arguments are incorrect, and are willing to argue for that position". I have never seen anyone claim that andressen and Lecunn do not truly believe their arguments. I also legitimately think that x-risk arguments are incorrect, am I conducting an "infowar"? Adopting this viewpoint seems like it would blind you to legitimate arguments from the other side.
That's not to say you can't point out errors in argumentations, or point out how the Lecunn and andressen have financial incentives that may be blinding their judgments. But I think this comment crosses the line into counterproductive "Us vs them" tribalism.
1This seems like an epistemically dangerous way of describing the situation that "These people think that AI x-risk arguments are incorrect, and are willing to argue for that position".
I don't think the comment you're responding to is doing this; I think it's straightforwardly accusing LeCun and Andreesen of conducting an infowar against AI safety. It also doesn't claim that they don't believe their own arguments.
Now, the "deliberate infowar in service of accelerationism" framing seems mostly wrong to me (at least with respect to LeCun; I wouldn't be surprised if there was a bit of that going on elsewhere), but sometimes that is a thing that happens and we need to be able to discuss whether that's happening in any given instance. re: your point about tribalism, this does carry risks of various kinds of motivated cognition, but the correct answer is not to cordon off a section of reality and declare it off-limits for discussion.
I find it interesting how he says that there is no such thing as AGI, but acknowledges that machines will "eventually surpass human intelligence in all domains where humans are intelligent" as that would meet most people's definition of AGI.
The somewhat-reasonable-position-adjacent-to-what-Yann-believes would be: “I don’t like the term ‘AGI’. It gives the wrong idea. We should use a different term instead. I like ‘human-level AI’.”
I.e., it’s a purely terminological complaint. And it’s not a crazy one! Lots of reasonable people think that “AGI” was a poorly-chosen term, although I still think it’s possibly the least-bad option.
Yann’s actual rhetorical approach tends to be:
- Step 1: (re)-define the term “AGI” in his own idiosyncratic and completely insane way;
- Step 2: say there’s no such thing as “AGI” (as so defined), and that anyone who talks about AGI is a moron.
I talk about it in much more detail here.
Note that most of the talk is about several (in his opinion) promising research directions for ML in the coming years. Which, he hopes, would lead to planning and more general animal-like capability, or AGI, although he doesn't like that term. One upshot is that autoregressive language models will not scale to AGI. The slides in the screenshot above aren't really the topic of the talk, he in fact skipped the last two. I found the talk interesting, although I can't judge how realistic his proposals are.
1Plug: I thought about LeCun's framework for several weeks in the context of alignment, here is the result: Aligning an H-JEPA agent via training on the outputs of an LLM-based "exemplary actor". TLDR: nothing particularly convincing, LeCun didn't "solve" alignment.
I find it interesting how he says that there is no such thing as AGI, but (…) that would meet most people's definition of AGI.
In some frameworks, there’s no such things as Yann Lecun: just stories written by his brain… but that would meet most people’s definition of Yann Lecun.
In other words, that’s a figure of style saying « what many reify as (AGI/YL) is better thought as process ».
I find it interesting how he says that there is no such thing as AGI, but acknowledges that machines will "eventually surpass human intelligence in all domains where humans are intelligent" as that would meet most people's definition of AGI.
I don't see how saying that machines will "eventually surpass human intelligence in all domains where humans are intelligent" imply the G in AGI.
I’m surprised by his stance on AGI being impossible. I think humans are proof of general learning machines - we’re capable of learning anything not everything. So it’s definitely possible for something non-human or non-biological to exist with our level of general intelligence capabilities and also take advantage of it to the fullest extent. Also surprised META has taken the open source stance for AI models, which is a great thing and good to see Lecun championing that unlike other big AI companies. I agree with his point, and most AI optimists support this, that there’s no evidence to suggest that inherit in intelligence is the desire for self-preservation or dominance. But remains to be seen if highly intelligent systems at some threshold manifest consciousness, in which those inherent qualities stem from (humans as an example and as a result dominate earth as apex predator or Lecun’s example for us to create highly intelligent systems to be subservient to us).
Yann recently gave a presentation at MIT on Objective-Driven AI with his specific proposal being based upon a Joint Embedding Predictive Architecture.
He claims that his proposal will make AI safe and steerable, so I thought it was worthwhile copying the slides at the end which provide a very quick and accessible overview of his perspective:
Here's a link to the talk itself.
I find it interesting how he says that there is no such thing as AGI, but acknowledges that machines will "eventually surpass human intelligence in all domains where humans are intelligent" as that would meet most people's definition of AGI.
I also observe that he has framed his responses to safety on "How to solve the alignment problem?". I think this is important. It suggests that even people who think aligning AGI will be easy have started to think a bit more about this problem and I see this as a victory in and of itself.
You may also find it interesting to read Steven Byrnes' skeptical comments on this proposal.