Intro to Large Language Models [Video] | Hacker News Reader

Intro to Large Language Models [Video]

youtube.com

291 pointsgeorgehill32 comments

I initially thought this was his older video, but I see this video was published just an hour ago (from the time of this comment).

I was thinking of this other video he had published early this year: "Let's build GPT: from scratch, in code, spelled out" https://www.youtube.com/watch?v=kCc8FmEb1nY

Karpathy is generally well-reputed as a good tutor, especially for complex topics in AI / ML.

Really interesting analogy in the video is the discussion about the system thinking from book Thinking, Fast and Slow by Daniel Kahneman

System 1 thinking: Fast automatic thinking and rapid decisions. For example is when someone ask you 2 + 2, you don't think. You just reply quickly instantly. LLMs currently only have system 1 thinking.

System 2 thinking: Rational slow thinking to make complex decisions. For example when someone ask you 17 x 24 you think slowly and rationally to multiply. This kind of thinking is a major component we need for AGI. Current rumor from OpenAI about so called "Q*" algorithm could be something related to system 2 thinking (Just speculation at this point)

I was thinking about something similar.

System 1 thinking: patterns.

System 2 thinking: logic.

For example:

System 1 thinking: Does x sentence sound like a correct English sentence.

System 2 thinking: Verify x sentence is a correct English sentence by using grammar rules.

Someone fluent in English can form correct English sentences using only system 1 thinking, while someone that has just started learning English must think about grammar rules (using system 2 thinking) to do it.

I wonder why OpenAI doesn't try to get more feedback and training data from its users, though I do notice that sometimes it'll give me two answers and ask me to pick the better one.

For example I've noticed that a lot of the time when I ask ChatGPT a coding question it might get 90% of the answer. When I tell it what to fix and/or add, it usually gets the answer. I wonder if they're using these refined answers to fine-tune those original prompts.

I wonder how the LLM interacts with other software like the calculator or Python interpreter. It would be great if this were modular so that the LLM OS could be more like Unix than Windows which is what OpenAI seems to be trying to emulate.

Ultimately though it seems to me like AGI is fairly straightforward from here. Just train on more quality data - in particular enabling the machine to generate this training data, increase parameter size, and the LLM just gets better and better. Seems like we don't even need any new major breakthroughs to create something resembling AGI.

They should be capturing the changes that people make to the ChatGPT outputs. Many people will be copying the outputs to some other application and then make changes. If open AI would make it easier to modify the outputs right within ChatGPT, they could use that as feedback. Basically, fuse the end-user UI with the UI of the annotates.

I think they learned from Tay.AI and friends.

I have zero faith that the average ChatGPT user will make quality edits. If anything, this invites trolling and active dataset poisoning/manipulation the moment people figure out that's what they're doing.

Karpathy has an excellent zero-to-hero series on the topic in which he explains the very core of the neural networks, LLMs and the related concepts. With no background on the topic, I was able to get an idea what's all this about and even become dangerous: https://karpathy.ai/zero-to-hero.html

There's something enlightening in hands-on learning without using metaphors. He even opens the code of production grade tools to show you how exactly the concepts he explained and build together are actually implemented IRL.

This is a style of teaching that clicks with me. I don't learn well with metaphors and high abstractions and find it magical to remove the magic of amazing things and bring it down to easy to reason pieces which can create a complex structure with composition so you can just disregard the complexity as a separate thing of the core.

Great video (as usual). Andrej has a Feynman like way of explaining very complex topics in a succinct and digestible way.

An aside... incredibly, it looks like he recorded in one cut from his hotel room.

This was an incredibly informative talk, especially the ideal of giving LLMs the System 2 thinking capability. I think if LLMs can do system 2 thinking we are one more step closer to AGI. I’ve summarised the talk here - https://gist.github.com/anupj/f3a778dcb26972ba72c774634a80d7... - if you anyone wants to RAG the text for their custom GPT :)

This video was very informative and clear for a beginner like me who is curious about AI and ML. I'd like to learn more about how to finetune llama for different tasks and domains. Does anyone have any recommendations for resources that explain this concept in a simple way and gradually introduce the technical details and tools required?

There actually is a reward function for text that can be used to go beyond the human input data. It is plausibility:

If you question the response and check it against responses about related questions, how much does it align with those?

This is what we humans do, too. It’s also what we do in science. It’s things not “adding up” that tells us where we must improve.

Upon seeing the title I thought to myself "what's the point... karpathy already has an unbeatable series of videos on this" ... before seeing that this too was from him. I assume if one has been through zero-to-hero, this is going over stuff you already know, but will report back after watching.

Reporting back. As one may have guessed, it is light on technical details (no code for example) and serves as possibly an into to zero-to-hero. Some interesting insights into the near-future possibilities but nothing that can really be applied practically.

Summary and Transcript : https://www.wisdominanutshell.academy/andrej-karpathy/1hr-ta...

Andrej made the micrograd video after the code was released a couple years later. As someone who takes learning neural networks as a hobby, I really want to see him make videos about the llama.c project.

He said he planned to a few months ago, but was busy/waiting to add some more features/fixes to the code first.

I didn't realise llama.c was from him too.

https://github.com/karpathy/llama2.c

For the background for whatever Q* is, listen to 35:00-40:45.

Oh no, not another video on ... OMG it's from Andrej!

Perfect timing! I'm going through Andrew Ng's course, Machine Learning Guide podcast and I have an hour free this morning to watch this :)

Have they improved the sound quality on Ng's course? I must be sensitive to these things, because I couldn't bear to listen to it when I tried a few years ago.

No. I think I know what you're talking about - it's kinda harsh and raspy and loud at the same time. It doesn't bother me but someone in an adjacent room to me actually brought it up.

How good is this guy? It's just awe-inspiring.

It's Andre so it's must see TV.

Highly recommend!

Maybe the best speaker I've ever ran into, amazing. There should definitely be a voice model based on him

Didn't know there is scary jailbreak to manage, this is the risk we are going to face it.

Love Karpathy videos. It’s almost like he’s a stealth dev evangelist or something…

This video gave me an I know Kung Fu moment

I experienced that exact moment upon watching Andrej's video on backpropagation.

Here’s another awesome Karpathy lecture from Stanford:

https://youtu.be/XfpMkf4rD6E?si=1_EmuYDFfi7RNEhz

This video is the best for learning attention, specifically where he explains:

Think of attention like a directed graph of vectors passing messages to each other

Keys are what other tokens are communicating to you,

Queries are what you are interested in,

and Values are what you are projecting out yourself.

When you matrix multiply the queries x keys(transposed), you measure the interestingness or affinity between the two.

The majority representing non technical population, journalists and 50+ expert dependent bureaucrats (mostly law academics - never worked really) shitting their pants over alleged "ai" dangers indoctrinated by "ai" executives to push ahead regulation to secure their market position, or in the case of google because it makes their now shitty search business model obsolete in the long term, by creating entry barriers thus reducing competition.

Meanwhile a guy somewhere in Africa adjusting an answer probably stating that humans can do photosynthesis: bruh

Reply on news.ycombinator.com