Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Andrej Karpathy – It will take a decade to work through the issues with agents

dwarkesh.com

Post reply on HN
Print view View original

721–730 of 1001 posts

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#721

Post by senko »

ekjhgkejhgk wrote: post #134
Imnimo wrote: post #68

>What takes the long amount of time and the way to think about it is that it’s a march of nines. Every single nine is a constant amount of work. Every single nine is the same amount of work. When you get a demo and something works 90% of the time, that’s just the first nine. Then you need the second nine, a third nine, a fourth nine, a fifth nine. While I was at Tesla for five years or so, we went through maybe three…

The interview which I've watched recently with Rich Sutton left me with the impression that AGI is not just a matter of adding more 9s. The interviewer had an idea that he took for granted: that to understand language you have to have a model of the world. LLMs seem to udnerstand language therefore they've trained a model of the world. Sutton rejected the premise immediately. He might be right in being skeptical here…

The thing is, achieving say, 99.99999% reliable AI would be spectacularly useful even if it's a dead end from the AGI perspective.

People routinely conflate the "useful LLMs" and "AGI", likely because AGI has been so hyped up, but you don't need AGI to have useful AI.

It's like saying the Internet is dead end because it didn't lead to telepathy. It didn't, but it sure as hell is useful.

It's beneficial to have both discussions: whether and how to achieve AGI and how to grapple with it, and how to improve a reliability, performance and cost of LLMs for more prosaic use cases.

It's just that they are separate discussions.

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#722

Post by simianwords »

ssivark wrote: post #613
robotresearcher wrote: post #277

Earlier quoted context omitted.

In my view 'understand' is a folk psychology term that does not have a technical meaning. Like 'intelligent', 'beautiful', and 'interesting'. It usefully labels a basket of behaviors we see in others, and that is all it does. In this view, if a machine performs a task as well as a human, it understands it exactly as much as a human. There's no problem of how to do understanding, only how to do tasks. The 'problem' me…

> In this view, if a machine performs a task as well as a human, it understands it exactly as much as a human. There's no problem of how to do understanding, only how to do tasks. Yes, but you also gloss over what a "task" is or what a "benchmark" is (which has to do with the meaning of generalization). Suppose an AI or human answers 7 questions correctly out of 10 on an ICPC problem set, what are we able infer from…

How do you quantify generality? If we have a benchmark that can quantify it and that benchmark reliably tells us that the LLM is within human levels of generalisation then the llm is not distinguishable from a human.

While it’s a good point that we need to benchmark generalisation ability, you have in fact agreed that it is not important to understand underlying mechanics.

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#723

Post by Mikhail_Edoshin »

the_gipsy wrote: post #441
godelski wrote: post #429

Earlier quoted context omitted.

> There's a view that suggests that language is intelligence. I think you find the limits when you dig in. What are you calling language? Can you really say that Eliza doesn't meet your criteria? What about a more advanced version? I mean we've been passing the Turing Test for decades now. > That language is a requirement for understanding. But this contradicts your earlier statement. If language is a requirement the…

> It is the condition where a person does not have an internal monologue. These people are just nutjobs that misinterpreted what internal monologue means, and have trouble doing basic introspection. I know there are a myriad of similar conditions, aphantasia, synaesthesia, etc. But someone without internal monologue simply could not function in our society, or at least not pass as someone without obvious mental dimin…

There is us a book written by a woman who suffered a stroke. She lost the ability to speak and understand language. Yet she remained conscious. It took her ten years to fully recover. The book is called "A stroke of insight".

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#724

Post by johnisgood »

latentsea wrote: post #547
bravura wrote: post #528

Earlier quoted context omitted.

How large is a lion? Learning the size of objects using pure text analysis requires significant gymnastics. Vision demonstrates physical size more easily. Multimodal learning is important. Full stop. Purely textual learning is not sample efficient for world modeling and the optimization can get stuck in local optima that are easily escaped through multimodal evidence. ("How large are lions? inducing distributions ove…

> How large is a lion? Twice of half of its size.

Can you be more specific about "size" here? (Do not tell me the definition of size though).

You are not wrong though, just very incomplete.

Your response is a food for thought, IMO.

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#726

Post by beowulfey »

stouset wrote: post #652
ziofill wrote: post #545

Earlier quoted context omitted.

I agree with this. A metaphor I like is that the reason why humans say the night sky is beautiful is because they see that it is, whereas an LLM says it because it’s been said enough times in its training data.

To play devil’s advocate, you have never seen the night sky. Photoreceptors in your eye have been excited in the presence of photons. Those photoreceptors have relayed this information across a nerve to neurons in your brain which receive this encoded information and splay it out to an array of other neurons. Each cell in this chain can rightfully claim to be a living organism in and of itself. “You” haven’t directly…

while true, that doesnt change the fact that every one of those independent units of transmission are within a single system (being trained on raw inputs), whereas the language model is derived from structured external data from outside the system. it's "skipping ahead" through a few layers of modeling, so to speak.

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#727

Post by jstummbillig »

willyxdjazz wrote: post #715

Maybe I'm being too simplistic, but I think we're mixing two distinct debates. Today we have an extraordinary invention—comparable to the wheel in its time. That invention is: predictive inference over all human knowledge. Period. I don't like calling it "Artificial Intelligence" because it's not intelligence; it's a prediction system that can project responses by illuminating patterns across all human knowledge enca…

> I don't like calling it "Artificial Intelligence" because it's not intelligence

A pattern I noticed in a AI[sic] discussions: Handwavily declaring what intelligence is not, while not explaining what is.

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#728

Post by meowface »

dang wrote: post #634
meowface wrote: post #126

Earlier quoted context omitted.

Eh, Dwarkesh has to market the podcasts somehow. I think it's fine for him to use hooks like this and for HN threads to respond to the hooks. 99% of HN threads only ever reply to the headline and that's not changing anytime soon. This will likely cause many people (including myself) to watch the full podcast when we otherwise might not have. The criticism that people are only replying to a tiny portion of the argumen…

99%? I have to stick up for HN here!

Ok, maybe not 99%. Probably at least 50% of comments in 70% of threads, though...

Re: Andrej Karpathy – It will take a decade to work through the issues with agents

#730

Post by password54321 »

meindnoch wrote: post #320

People keep talking about AGI as if it's some mystical leap beyond human capability. But let's be honest; software development at a modern startup is already the upper bound of applied intelligence. You're juggling shifting product specs, ambiguous user feedback, legacy code written by interns, and five competing JS frameworks, all while shipping on a Friday. Models can now do that. They can reason about asynchronous…

Computers being good/fast at automating/calculating things that people find difficult is not a new phenomenon. By your standards we have had general intelligence decades ago.
Post reply on HN