[hackinthebochs]

People like his takes because he gives an authoritative gloss to what they already believe. But his points are usually lacking in argument or rigor. Anyone that essentially expects the public to trust them when it comes to the outcome of AI/AGI should be view with suspicion.

[qualifiedai]

On the contrary, he gives good arguments about why open is safer and closed is more dangerous whereas other side gives, imho, convoluted arguments and asks for them to be proven wrong (as opposed to trying to prove themselves right).

[hackinthebochs]

His arguments in defense of barreling forward with AI are terrible. They have zero chance to convince someone who doesn't share his intuitions/interests. For example: https://twitter.com/ylecun/status/1718764953534939162

How easily smart people convince themselves of what they want to be true with zero self-awareness makes me much more fearful of what's to come.

>other side gives, imho, convoluted arguments and asks for them to be proven wrong (as opposed to trying to prove themselves right).

The question is what should our default stance be until proven otherwise? I submit it is not to continue building the potentially world-ending technology.

[qualifiedai]

The default in science is that the side arguing a point has the burden of proving it correct. Not asking the other side to prove them wrong.

[hackinthebochs]

Burden of proof is a tragically misapplied concept. The question is what position should be the default given some uncertainty? This isn't decided in a vacuum. The default should be decided based on the facts and utility estimates of various outcomes where relevant. In the case where the issue is of purely academic interest, the burden of proof is on those making the claim. When the issue has real world consequences, then utility estimates of various outcomes are the overriding concern.

When you see a gun on the table, what do you do? You assume its loaded until proven otherwise. For some reason, those who imagine AI will usher in some tech-utopia not only assume the gun is empty, but that pulling the trigger will bring forth endless prosperity. It's rather insane actually.

[zmgsabst]

You’re assuming your conclusion by saying AI is like a gun.

That’s not a rational argument for why we should be concerned — you merely asserted you were.

Why is catering to your feelings the default position?

[flagrant_taco]

The default position should be based on risks and unknowns, I don't think the GP was making any recommendation based on personal feelings.

The fact is we don't know how current ML actually does what it does, we don't know what we'll have next month, and we wouldn't know how to recognize an AI or AGI if we developed one. The risks and unknowns are high, the default position should be to not develop the technology unless and until someone proves without reasonable doubt why we can and should do it, and how we'll do it safely.

[naasking]

> The default in science is that the side arguing a point has the burden of proving it correct

Right, and Yann is arguing the point that AI and LLMs are not or will not be dangerous. Where's his proof? As the parent posters have said, he has none.

[qualifiedai]

Show me LLM that has reached a level of automony and self sufficiency of a cat (Yann's example) or even of a simple beattle.

LLMs are tools in amplifying individual human intelligence, 100% of automony and will come from their user (human).

Also, even as tools they have fundamental limitations which stem from their autoregressive nature

[harshreality]

Before the era of quantum physics, the idea of a weapon like fission or fusion weapons would've been inconceivable without resorting to magical thinking.

Your argument, and Yann's, is that AGI, or what you call AGI, is a kind of quasi-intelligent golem that, despite being generally intelligent, doesn't have human-level intelligence. Your claim that it will never be human-equivalent, much less trans-human/ASI, is built into your worldview. It's not a conclusion. It's an assumption on your part.

People like you and Yann can believe that if you want, but you have no evidence, because nobody knows what's required for human-level intelligence. Nobody knows whether some kind of system involving neural nets could develop human-level intelligence or beyond. It could involve different architecture or training methods. There's no assumption by AI doomers that AGI will be achieved by a LLM with more parameters or more or better training data.

[flagrant_taco]

How would you define or recognize a level of intelligence or autonomy that is sufficient to raise concerns? Would you be able to recognize it before its too late?

An early GPT-4 test ended up with GPT successfully solving capchas by tricking a TaskRabbit worker into doing it for them [1]. When asked by the worker if it was a robot, GPT decided to lie to the worker and claim it had a visual impairment that made it difficult to solve the puzzle. That sounds like a level of autonomy and social engineering skills that could be concerning to a reasonable person.

[1] https://www.businessinsider.com/gpt4-openai-chatgpt-taskrabb...

[naasking]

> Show me LLM that has reached a level of automony and self sufficiency

Autonomy and self-sufficiency are not the only ways a system can be dangerous.

But even if this claim were true, ChaosGPT proves that some humans will almost immediately set about using such a non-autonomous tool to create a dangerous autonomous agent. This is my problem with LeCunn, nearly all of his points are trivially refuted by real world observations, yet he keeps repeating them as if they simply must be true.

> Also, even as tools they have fundamental limitations which stem from their autoregressive nature

That's yet another speculative point that LeCunn constantly asserts. Scaling laws have not shown any indication of even approaching a limit.

[esafak]

LLMs are building blocks on the path to AGI. Can you even ask your cat any question you can ask chatGPT?

People dismiss LLMs because they are not embodied, and lack continuous training. That is to come.

[sdenton4]

/even of a simple beattle./

So, more Ringo than Lenon?

[bbor]

A nice thought but not very helpful. What’s the null hypothesis here?

[CSMastermind]

The null hypothesis is that technological advancement will be a net positive for humanity, as all previous technological advancements have been.

[SantalBlush]

Calling a personal assumption a null hypothesis doesn't make it more objective.

[concordDance]

I've read a fair bit of LeCunn and frankly have been very unimpressed. He is, as far as I can tell, the only big AI name who doesn't think that if we somehow managed to make an actual superintelligent AI it wouldn't be dangerous.

[lucubratory]

And his reasoning for it is absurd, a combination of evolutionary psychology (not the most scientifically rigorous field to begin with and not even his field) and the belief that no one would choose to make an unsafe AI system, while simultaneously arguing for open sourcing everything so that anyone on the planet can do so if they want.

[jlpom]

His reasoning is that good guy’s AI will counteract bad guy’s ones or a an AI that has developed his own objective against humanity if that happens, which is I think a reasonable argument.

[_heimdall]

He effectively is arguing that the "good guys" will develop a superhero AI that will protect us simply because it feels that is the right thing to do. I'm not sure how any logical basis can be used to back that up, or where any meaningful example of similar behavior that didn't ultimately lose to the "bad guys" can be found in human history.

[pixl97]

In the 20th century this same thinking lead to MADD and massive stocks of nuclear weapons that still present an existential threat to humanity. Depending on the scaling potential of intelligence this just adds further risks at this level, not less or balanced risks.

[avsteele]

Please point me to an example of his good arguments.

I only see his posts on Twitter but haven't been impressed.

[orochimaaru]

This was a view that was initially taken by the government for encryption as well. But everyone can agree open sourcing the algorithms and libraries has been the best move.

It is the same with AI/AGI. Anything closed source and having regulatory oversight is useless, decreases innovation, increases bureaucracy and will only serve those who wish to build a “moat” to further their hold on the technology.

[hackinthebochs]

Sure, open source worked well for encryption. It wouldn't work well for bioweapons or nuclear weapons technology. The question is which category does AGI best fit? We shouldn't elide analysis for obtuse comparisons to the past.

[vkou]

> Anything closed source and having regulatory oversight is useless

Source isn't even the problem, unless you're a billionaire, you can't afford to train the model.

This isn't the wheel, the printing press, the PC, or the public internet, which created opportunities for everyone.

This is pay-to-play that is only affordable to the nation-state the mega-corporation, and the latter might let you play around as a digital sharecropper on their platform, until they cut you off, because they can make more money by having first-party ownership of whatever you built.

[bangkoksbest]

The model is the instance because the instance is what is being used at inference time. The training simply produces some set of weights.

Open source would mean someone could see and run the model code locally/independently to create an instance. For an LLM, this is much less insurmountable of an issue in terms of resources needed.

[_heimdall]

We should probably be calling for an open model or open data approach if that's what we think is best. Calling it open source leaves plenty of gray area in the definition.

For example, I run a few instances of open source software though my instances' databases are private.

[pixl97]

That is questionable.

We are still answering questions on LLM/AI scaling. Ya, you might have your cottage industry AI giving out answers on the trickle of information you feed it, but will that even be comparable to one being fed megawatts of power with terabytes of data per second flowing into its databanks?

[mycologos]

At the same time, it seems like some antidote is needed to the breathless, quasi-mystical hype that cryptic OpenAI claims seem designed to stoke. Demanding precise and substantive criticisms of something about which almost no technical details have been provided seems like an unfair bar.

[gexla]

But the post links to Twitter (X.) That's the wrong place for argument or rigor. But do you really need that? Seems to me like his post is essentially saying "Pay attention to all this stuff which people have been working on for a long time." He's just casting a vote. The argument and rigor is already there for further reading. That's interesting for me, because I don't know much about the topic. If you already know all this stuff, then maybe it's not interesting.

[peyton]

I mean I don’t think predicting the future is something that typically involves rigor. The outcome is pretty clear: whatever makes a ton of money. Probably a trusted friend in your pocket that sometimes helps you buy stuff. The most negative predictions are silly because they don’t involve making a ton of money for anybody.

[makeitdouble]

The point about money is important, but we should also keep in mind most outcomes will make a ton of money for someone somewhere.

Hell, there's wars killing tens of thousands of people going on right now, and a ton of money is changing hands making a juicy business for whole industries.

[ethanbond]

Your theory for predicting the future is “people don’t lose money?”