actually i was in on the joke (there's too much long-term coherence), but it took me enough time, scrutiny and thought that i enjoyed it as a demonstration of hard it is now to tell. though i don't super enjoy this genre of slop posting more generally 🙂
@karpathy
-
Alignment as Computer Security Problem with Neural Networks
By
–
When people say Alignment I just hear Computer Security (now with neural nets) and it makes more sense.
-
Questioning the Fundamental Definition of AI Agents
By
–
Yeah I don’t think “agents” is used in this way but … it feels a bit wrong. What’s some actual fundamental distinguishing property wrt what already exists? “It happens to use an LLM somewhere” is I think the current usage but imo it’s kind of lame.
-
Digital Agents and Intelligent Systems Already Surround Us Daily
By
–
Hmm disagree. Mac OS is a highly intelligent agent with lots of background tasks. Gmail is. X is. Businesses run many on your behalf, eg anytime you swipe a credit card. There’s lots of highly sophisticated, highly intelligent digital entities we use/dispatch all the time.
-
AI’s Unique Role: Democratizing Commands to the Technosphere
By
–
It’s a cool analogy but traditional software already satisfies it – there are “worker drones” delivering, filtering and ranking posts/emails etc etc. So imo the analogy is subtle. Eg maybe AI allows a lot more people to issue new diverse commands to the technosphere. Or it’s a
-
Manual Backpropagation Wasn’t Actually a Major Problem
By
–
Yeah except it wasn’t actually that bad. I mean it was a bit annoying, code bloating and you’d have to run grad check, but it also didn’t feel like a major impediment after a bit of practice. I don’t recall spending significant portion of my time deriving/writing backward pass.
-
LLM domain training should produce manuals not weight updates
By
–
yes exactly, "training" an LLM on a target domain should output a manual not a weight diff.
-
LLM Self-Awareness in Task Adaptation and Problem-Solving
By
–
This is not the core issue. The core issue is that the LLM has to autonomously and in general way figure out that it is natively not well adapted to do this task in its head, that it doesn’t succeed in doing so, and that it should do this and that instead to solve it.
-
Missing Major LLM Learning Paradigm Beyond Pretraining and Finetuning
By
–
We're missing (at least one) major paradigm for LLM learning. Not sure what to call it, possibly it has a name – system prompt learning? Pretraining is for knowledge.
Finetuning (SL/RL) is for habitual behavior. Both of these involve a change in parameters but a lot of human