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

901–910 of 1001 posts

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

#901

Post by ekjhgkejhgk »

theptip wrote: post #605
ekjhgkejhgk wrote: post #134

Earlier quoted context omitted.

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…

> LLMs seem to udnerstand language therefore they've trained a model of the world. This isn’t the claim, obviously. LLMs seem to understand a lot more than just language. If you’ve worked with one for hundreds of hours actually exercising frontier capabilities I don’t see how you could think otherwise.

> This isn’t the claim, obviously.

This is precisely the claim that leads a of lot people to believe that all you need to reach AGI is more compute.

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

#902

Post by qudat »

torginus wrote: post #898

Translated to the language of capital investment, this means that basically all the hardware bought for AI will be obsolete by then - am I reading this wrong or can we say that most of the data centers are basically worthless?

The danger is using your GPUs as leverage to build more data centers, which it seems some companies are doing. That’s going to hurt when the hardware value goes to zero and banks start collecting on debt

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

#903

Post by consumer451 »

There is so much to unpack here. Currently on my 3rd re-watch. Biggest take-aways:

1. This is the death knell for the the "AI" investment bubble. Agents that are useful for non-devs are 10 years away.

2. Andrej thinks that GPT5 pro is SOTA for code? Really? As a Sonnet normie.. can anyone please help me understand this?

edit:

3. You can't see any major tech developments on the GDP growth chart? Really? WTF? Have we all been smoking tech crack, this whole time? So GDP didn't grow extra from tech any single tech development, like the Internet? This broke my brain.

disclaimer: On the daily, I use LLM dev tools to add amazing LLM-enabled features to my pre-money SaaS. It's really cool and users love the features.

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

#904

Post by simonw »

ambicapter wrote: post #849
simonw wrote: post #522

It looks like Andrej's definition of "agent" here is an entity that can replace a human employee entirely - from the first few minutes of the conversation: When you’re talking about an agent, or what the labs have in mind and maybe what I have in mind as well, you should think of it almost like an employee or an intern that you would hire to work with you. For example, you work with some employees here. When would yo…

Do you have a comment? Most of what you've said here is a quote.

This is part of my wider hobby of collecting definitions of "agents" - you can see more in my collection here: https://simonwillison.net/tags/agent-definitions/

In this case the specific definition matters because the title of the HN submission is "it will take a decade to work through the issues with agents."

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

#905

Post by lysecret »

He’s the reason I got into ml in 2016 I owe him the world. Great interview I was a bit surprised how little use he got out of agents but it makes sense I wonder how he feels about having such an essential role in creating the whole vibe coding idea.

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

#906

Post by sheepscreek »

consumer451 wrote: post #903

There is so much to unpack here. Currently on my 3rd re-watch. Biggest take-aways: 1. This is the death knell for the the "AI" investment bubble. Agents that are useful for non-devs are 10 years away. 2. Andrej thinks that GPT5 pro is SOTA for code? Really? As a Sonnet normie.. can anyone please help me understand this? edit: 3. You can't see any major tech developments on the GDP growth chart? Really? WTF? Have we a…

In my case at least, it’s very good at following all instructions to the T. Claude 4.0 (haven’t used it much since 4.5 came out) would often miss some key things in my instructions. The output is very high quality as well. Many things (even complex coding tasks) work well in one shot.

For extremely complex multi-step problems though - it may need some help in breaking the tasks down to more manageable chunks. But will eventually ace it. As an example, I had good success with a project that involved:

- Rewriting all internals in a dotnet/C# application to use Apache Arrow types for data through the entire pipeline - Adapting the architecture to be streaming first instead of working through entire data in each stage - Designing and implementing a complex system that creates many different projections of the data based on everything that has read in the stream so far and create multiple outputs based on that, in parallel as the stream is being read in real-time - Recreating a prototype of the entire project in Rust

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

#907

Post by judahmeek »

__MatrixMan__ wrote: post #881
judahmeek wrote: post #572

Earlier quoted context omitted.

All intelligence is specific, as evidenced by the fact that a universal definition regarding the specifics of "common sense" doesn't exist.

Common is not the same as general. A general key would open every lock. Common keys... well they're quite familiar.

My point was that all intelligence is based on an individual's experiences, therefore an individual's intelligence is specific to those experiences.

Even when we "generalize" our intelligence, we can only extend it within the realm of human senses & concepts, so it's still intelligence specific to human concerns.

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

#908

Post by judahmeek »

walkabout wrote: post #831
judahmeek wrote: post #572

Earlier quoted context omitted.

All intelligence is specific, as evidenced by the fact that a universal definition regarding the specifics of "common sense" doesn't exist.

A universal definition of “chair” is pretty hard to pin down, too…

What are your sources for that claim?

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

#909

Post by godelski »

simianwords wrote: post #722
ssivark wrote: post #613

Earlier quoted context omitted.

> 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.

That's kinda their point

The difference though is they understand that you can't just benchmark your way into proofs. Just like you can't unit test your way into showing code is error free. Benchmarks and unit tests are great tools that provide a lot of help, but just because a hammer is useful doesn't make everything a nail.

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

#910

Post by rhetocj23 »

CaptainOfCoit wrote: post #732
sothatsit wrote: post #699

Earlier quoted context omitted.

I don't think he is saying agents are not useful at all, just that they are not anywhere near the capability of human software developers. Karpathy later says he used agents to write the Rust translation of algorithms he wrote in Python. He also explicitly says that agents can be useful for writing boilerplate or for code that can be very commonly found online. So I don't think he is saying they are not useful at all…

My biggest takeaway is that agents/LLMs in general are super helpful when paired together with a human who knows the inside and out of software development, who uses it side-by-side with their normal work. They start being less useful when you start treating them as "I can send them ill-specified stuff, ignore them for 10 minutes and merge their results", as things spiral out of control. Basically "vibe-coding" as a…

It does seem pretty clear that an individual who possess super high quality human capital, paired with something like an LLM (provided the LLM is good enough relative to the individual) can be a powerful combination.

The issues are:

1) There isnt enough supply of those individuals 2) Such an LLM of that kind doesnt exist (at least not in consistent nature) 3) The amount invested into what is going on will not yield returns commensurate to the required rate of return

Interestingly enough, I believe Andrej Karpathy is also focusing on education (levelling up the supply of human capital) - I came to the above conclusion about a month ago. And it 'feels' right to me.

Post reply on HN