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

TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

anfalmushtaq.com

Post reply on HN
Print view View original

31–40 of 91 posts

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#31

Post by Salgat »

HumanOstrich wrote: post #15
dzogchen wrote: post #2

For a model to be ‘fully’ open source you need more than the model itself and a way to run it. You also need the data and the program that can be used to train it. See The Open Source AI Definition from OSI: https://opensource.org/ai

That is incorrect -- you do not have to provide the full training data to meet the requirements. I recommend reading the actual Open Source AI Definition[1] and the FAQ[2]. There's also the whitepaper[3] that goes into much more detail about the state of affairs. [1]: https://opensource.org/ai/open-source-ai-definition [2]: https://hackmd.io/@opensourceinitiative/osaid-faq#What-is-th... [3]: https://opensource.org/wp…

FYI their open source ai definition released with a lot of controversy, unsurprisingly because it had heavy contribution from corporations with their own interests. It's best to ignore it for now until the wider community has decided on an appropriate open source definition.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#32

Post by moffkalast »

szundi wrote: post #3
dzogchen wrote: post #2

For a model to be ‘fully’ open source you need more than the model itself and a way to run it. You also need the data and the program that can be used to train it. See The Open Source AI Definition from OSI: https://opensource.org/ai

Yes. I cannot comprehend this to this day. A model weights data + runner is how different from a closed source executable? Why do everyone call these open source?

Because marketing (open source is a buzzword after all), and the media just repeats what they read in press releases verbatim. But most people working with the models themselves call them open-weight, except for some occasional exception like OLMo that publishes the dataset and training scripts and is actually open source.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#33

Post by moffkalast »

atq2119 wrote: post #27
a2128 wrote: post #6

Earlier quoted context omitted.

Is it reasonable to expect companies to redistribute 100TB of copyrighted content they used for their LLM, just on the off-chance someone has a few million laying around and wants to reproduce the model from scratch?

Perhaps that's not reasonable to expect, but Meta apparently kind of did it anyway, if not in a way that helps reproduce their LLM: https://arstechnica.com/tech-policy/2025/02/meta-torrented-o...

Actually they did, the entire 15T tokens that were supposedly used for training the llama-3 base models are up on HF as a dataset: https://huggingface.co/datasets/HuggingFaceFW/fineweb

It's just not literally labelled so because of obvious reasons.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#34

Post by WA »

albert_e wrote: post #12

OT: What is a good place to discuss the original video -- once it has dropped out of the HN front-page? I am going through the video myself -- roughly halfway through -- and have a fw things to bring up. Here they are now that we have a fresh opportunity to discuss: 1 - MATH and LLMs I am curious why many of the examples Andrej chose to pose to the LLM were "computational" questions -- for instance "what is 2+2" or s…

Regarding 1 - MATH:

Somewhere in the video he says that LLMs have expert (only slightly fuzzy) knowledge about a lot of topics, but fail with simple math questions. Many non-technical people anthropomorphize LLMs and don't know that they can't think or calculate like a real calculator. LLMs compute tokens and you can improve the performance, if you don't put too much computation into a single result token.

I think it's an excellent example to show the capabilities and limits of LLMs. For softer topics, you can argue a lot more about what's considered to be right or wrong. With Math, you have a single correct answer that can be evaluated and people assume that computers are good at computer things, such as calculating numbers, even though LLMs actually aren't good at this.

The takeaway is: Prompting and "computational complexity per token" matter and if you understand how it works for math, you probably understand how it works for softer things like answers about law or whatever.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#35

Post by amelius »

sigmoid10 wrote: post #9
szundi wrote: post #3

Earlier quoted context omitted.

Yes. I cannot comprehend this to this day. A model weights data + runner is how different from a closed source executable? Why do everyone call these open source?

Open source becomes really complicated once assets with unclear license are involved in any way. Lots of people for example would say that Jedi Knight 2 is open source because Raven Software released the source code and tools needed to build the game. But that alone doesn't mean you can run it, because you still need to get a hold of all the assets (models, textures, sounds) which may or may not still be property of…

OK, but that leaves the tools used to train the model (aka the build scripts). These could be open sourced.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#36

Post by amelius »

est wrote: post #20

I have read many articles about LLMs, and understand how it works in general, but one thing always bothers me: why other models did't work as good as SOTA ones? What's the history and reason behind the current model architecture?

I guess nobody really knows why. Everybody just goes with what works, and tries only small variations. It's a bit like alchemy.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#37

Post by m11a »

albert_e wrote: post #12

OT: What is a good place to discuss the original video -- once it has dropped out of the HN front-page? I am going through the video myself -- roughly halfway through -- and have a fw things to bring up. Here they are now that we have a fresh opportunity to discuss: 1 - MATH and LLMs I am curious why many of the examples Andrej chose to pose to the LLM were "computational" questions -- for instance "what is 2+2" or s…

> I am curious to know more about the limitations / perils of using LLMs to train/evaluate other LLMs.

At the extreme, there is this paper on ‘inbred LLMs’: https://www.nature.com/articles/s41586-024-07566-y

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#38

Post by rob_c »

est wrote: post #20

I have read many articles about LLMs, and understand how it works in general, but one thing always bothers me: why other models did't work as good as SOTA ones? What's the history and reason behind the current model architecture?

Simply put the stability of attention based models over non attention based ones.

Google dropped MHA self attention which was a major idea that they showed to work. OpenAI saw and built an empire on FeedForward attention models which are (compared to most alternatives) super stable at generation. DeepSeek showed evidence it's possible to further push these models and use effectively compression in the model design to pass around sufficient information for training. (Hence the Latent) They also did a lot of other cool stuff, but the main "core of the model" difference is this part...

Other than that, the biggest hurdle has been hardware. There's probably no way you could get kit from 2010 without even aes acceleration to evaluate most full-fat mhlffa models let alone train them. There's been a happy convergence of matrix acceleration on GPUs for gaming, graphical and high fidelity stimulation work. This combined with matrix based ML maths combined with high throughput memory advances means we can do what we're doing with llms now.

So inevitable outcome or happy convergence? That's for historians to decide imo. I think it's a bit of both.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#39

Post by rob_c »

amelius wrote: post #36
est wrote: post #20

I have read many articles about LLMs, and understand how it works in general, but one thing always bothers me: why other models did't work as good as SOTA ones? What's the history and reason behind the current model architecture?

I guess nobody really knows why. Everybody just goes with what works, and tries only small variations. It's a bit like alchemy.

Simply put no other model has the same number of effective skip connections or passes as much information through the model from input to output.

Earlier models had huge bottlenecks in terms of information limits and precision. (Auto encoders Vs uNets for example) And LSTM are still semi unstable.

Why the attention design as posited by Google works so well is part the skip forward and part "now we have enough information and processing power to try this".

It's well motivated but from a first principles up do we expect this to work well, it's a bit less well understood still. And if you're good at that you'll likely get a job offer very quickly.

Re: TL;DR of Deep Dive into LLMs Like ChatGPT by Andrej Karpathy

#40

Post by TeMPOraL »

a2128 wrote: post #6
dzogchen wrote: post #2

For a model to be ‘fully’ open source you need more than the model itself and a way to run it. You also need the data and the program that can be used to train it. See The Open Source AI Definition from OSI: https://opensource.org/ai

Is it reasonable to expect companies to redistribute 100TB of copyrighted content they used for their LLM, just on the off-chance someone has a few million laying around and wants to reproduce the model from scratch?

Redistribute? No. Itemize and link to? Yes.

With LLMs, the list doesn't even have to be kept up to date, nor the links alive (though publishing content hashes would go a long way here). It's not like you can get an identical copy of a model built anyway, there's too much randomness at every stage in the process. But, as long as the details of cleanup and training are also open, a list of training material used would suffice - people would fetch parts of it, substitute other parts with equivalents that are open/unlicensed/available, add new sources of their own, and the resulting model should have similar characteristics to the OG one who we could, now, call "open source".

Post reply on HN