One of the most brilliant AI minds on the planet, and he's focused on education. How to make all the innovation of the last decade accessible so the next generation can build what we don't know how to do today.

No magical thinking here. No empty blather about how AI is going to make us obsolete with the details all handwaved away. Karpathy sees that, for now, better humans are the only way forward.

Also, speculation as to why AI coders are "mortally terrified of exceptions": it's the same thing OpenAI recently wrote about, trying to get an answer at all costs to boost some accuracy metric. An exception is a signal of uncertainty indicating that you need to learn more about your problem. But that doesn't get you points. Only a "correct answer" gets you points.

Frontier AI research seems to have yet to operationalize a concept of progress without a final correct answer or victory condition. That's why AI is still so bad at Pokemon. To complete open-ended long-running tasks like Pokemon, you need to be motivated to get interesting things to happen, have some minimal sense of what kind of thing is interesting, and have the ability to adjust your sense of what is interesting as you learn more.

▲

steveBK123 16 hours ago | parent | next [-]

It's nice seeing commentary from someone who is both knowledgable in AI and NOT trying to pump the AI bag.

Right now the median actor in the space loudly proclaims AGI is right around the corner, while rolling out pornbots/ads/in-chat-shopping, which generally seems at odds with a real belief that AGI is close (TAM of AGI must be exponentially larger than the former).

▲

theptip 14 hours ago | parent | next [-]

Zvi made this point the other day, and then this counterpoint which I agree with more - if you think AGI is soon but you need to keep up the exponential datacenter growth for 2-3 years (or whatever “around the corner” means for the company in question) then a land-grab on consumer ARR is a faster way to short-term revenue (and therefore higher valuations at your next round).

OAI is also doing F100 and USG work; it takes longer to book the revenue though.

By selling porn and shopping you are in some sense weakening your position with regulators which you'll need when AGI starts displacing jobs - but you can also imagine thinking that this is a second order problem and winning the race is way more urgent.

	▲	AstroBen 27 minutes ago \| parent [-]
		Well if they think AGI is soon with no evidence then that's not worth a whole lot If they think AGI is soon AND they can back it up, why can't they convince investors and have them make up the difference for the next 2-3 years? Seems like a much better strategy than the pornbots and a weaker position with regulators I mean if anything this makes it seem like they're not confident they'll get to AGI that soon.. they're not optimizing for that eventual outcome

▲

throwaway314155 11 hours ago | parent | prev [-]

What exactly is a "pornbot"?

▲

thadk 15 hours ago | parent | prev | next [-]

Imagine a PhD mortally terrified of exceptions!

Now I see why Karpathy was talking of RL up-weights as if they were a destructive straw-drawn line of a drug for an LLM's training.

▲

TZubiri 15 hours ago | parent | prev [-]

>An exception is a signal of uncertainty indicating that you need to learn more about your problem

No, that would be a warning. Ab exceprion is a signal something failed and it was impossible to continue

▲

nextaccountic 11 hours ago | parent [-]

Many exceptions are recoverable. This sometimes depends on the context, and on how well polished the software is

	▲	TZubiri 9 hours ago \| parent [-]
		Yes. Note how I didn't say impossible to recover, just impossible to continue. The execution couldn't continue in one path due to an error it needed to be caught in another path. The difference with standard conditional mechanisms like if loops is mostly semantical. Exceptions are unforeseen errors, (technically they are sets of errors, which can have size 1, but the syntax is designed for catching groups of errors, if you want to react to a single error case you could also just use a condition with a return value and it ceases being an exception. )