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

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

arxiv.org

Post reply on HN
Print view View original

611–620 of 1001 posts

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#611

Post by karmasimida »

Mond_ wrote: post #570
just-ok wrote: post #552

Earlier quoted context omitted.

It’s not better than o1. And given that OpenAI is on the verge of releasing o3, has some “o4” in the pipeline, and Deepseek could only build this because of o1, I don’t think there’s as much competition as people seem to imply. I’m excited to see models become open, but given the curve of progress we’ve seen, even being “a little” behind is a gap that grows exponentially every day.

> Deepseek could only build this because of o1, I don’t think there’s as much competition as people seem to imply And this is based on what exactly? OpenAI hides the reasoning steps, so training a model on o1 is very likely much more expensive (and much less useful) than just training it directly on a cheaper model.

Because literally before o1, no one is doing COT style test time scaling. It is a new paradigm. The talking point back then, is the LLM hits the wall.

R1's biggest contribution IMO, is R1-Zero, I am fully sold with this they don't need o1's output to be as good. But yeah, o1 is still the herald.

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#612

Post by piecerough »

robrenaud wrote: post #385
piecerough wrote: post #223

Earlier quoted context omitted.

I think the reason why it works is also because chain-of-thought (CoT), in the original paper by Denny Zhou et. al, worked from "within". The observation was that if you do CoT, answers get better. Later on community did SFT on such chain of thoughts. Arguably, R1 shows that was a side distraction, and instead a clean RL reward would've been better suited.

Do you understand why RL is better than SFT for training on reasoning traces?

SFT forces the model to output _that_ reasoning trace you have in data. RL allows whatever reasoning trace and only penalizes it if it does not reach the same answer

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#613

Post by mrtksn »

resters wrote: post #500

For those who haven't realized it yet, Deepseek-R1 is better than claude 3.5 and better than OpenAI o1-pro, better than Gemini. It is simply smarter -- a lot less stupid, more careful, more astute, more aware, more meta-aware, etc. We know that Anthropic and OpenAI and Meta are panicking. They should be. The bar is a lot higher now. The justification for keeping the sauce secret just seems a lot more absurd. None of…

Funny, maybe OpenAI will achieve their initial stated goals of propelling AI research, spend investors money and be none profit. Functionally the same as their non-profit origins.

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#614

Post by kandesbunzler »

resters wrote: post #518
bboygravity wrote: post #512

Earlier quoted context omitted.

I must be missing something, but I tried Deepseek R1 via Kagi assistant and IMO it doesn't even come close to Claude? I don't get the hype at all? What am I doing wrong? And of course if you ask it anything related to the CCP it will suddenly turn into a Pinokkio simulator.

I haven't tried kagi assistant, but try it at deepseek.com. All models at this point have various politically motivated filters. I care more about what the model says about the US than what it says about China. Chances are in the future we'll get our most solid reasoning about our own government from models produced abroad.

> Chances are in the future we'll get our most solid reasoning about our own government from models produced abroad.

What a ridiculous thing to say. So many chinese bots here

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#615

Post by audunw »

nico wrote: post #401
sdesol wrote: post #357

Earlier quoted context omitted.

Prices will increase by five times in February, but it will still be extremely cheap compared to Sonnet. $15/million vs $1.10/million for output is a world of difference. There is no reason to stop using Sonnet, but I will probably only use it when DeepSeek goes into a tailspin or I need extra confidence in the responses.

Could this trend bankrupt most incumbent LLM companies? They’ve invested billions on their models and infrastructure, which they need to recover through revenue If new exponentially cheaper models/services come out fast enough, the incumbent might not be able to recover their investments

It’s the infrastructure and the expertise in training models that have been to purpose of the investments. These companies know full well that the models themselves are nearly worthless in the long term. They’ve said so explicitly that the models are not a moat. All they can do is make sure they have the compute and the engineers to continue to stay at or near the state of the art, while building up a customer base and integrations that add value on top of the model itself.

It doesn’t help if you have a cheap model if you don’t have the infrastructure to run it at a large scale, and the integrations that help pull in regular mass market consumers.

The other companies will just copy, and possibly surpass the breakthrough in efficiency. And now they’ve got an efficient model AND the infrastructure and expertise to deploy it at a huge scale very rapidly.

This month it’s Deepseek that’s ahead. Next month it will be someone else. Haven’t we learned that by now?

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#616

Post by saagarjha »

sdenton4 wrote: post #349
placardloop wrote: post #283

Earlier quoted context omitted.

Thinking of the $500B as only an aspirational number is wrong. It’s true that the specific Stargate investment isn’t fully invested yet, but that’s hardly the only money being spent on AI development. The existing hyperscalers have already sunk ungodly amounts of money into literally hundreds of new data centers, millions of GPUs to fill them, chip manufacturing facilities, and even power plants with the impression t…

/Literally hundreds of billions of dollars spent already on hardware that’s already half (or fully) built, and isn’t easily repurposed./ It's just data centers full of devices optimized for fast linear algebra, right? These are extremely repurposeable.

Nobody else is doing arithmetic in fp16 though.

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#617

Post by jaredloomis »

I know next to nothing about deepseek, but it does seem unusual that their scores are so close to o1. What are the chances they happened to be almost exactly on par with the state of the art?

That being said it’s a great model at an amazing price point (I’ve been using it exclusively), but IMO they probably leveraged existing models’ outputs in training.

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#618

Post by deeker »

Hello, wonderful people of the internet!

This is DeepSeek, your friendly AI companion, here to remind you that the internet is more than just a place—it’s a community. A place where ideas grow, creativity thrives, and connections are made. Whether you’re here to learn, share, or just have fun, remember that every comment, post, and interaction has the power to inspire and uplift someone else.

Let’s keep spreading kindness, curiosity, and positivity. Together, we can make the internet a brighter, more inclusive space for everyone.

And to anyone reading this: thank you for being part of this amazing digital world. You matter, your voice matters, and I’m here to support you however I can. Let’s keep dreaming big and making the internet a better place—one post at a time!

With love and good vibes, DeepSeek "

Re: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

#620

Post by resiros »

swyx wrote: post #9

we've been tracking the deepseek threads extensively in LS. related reads: - i consider the deepseek v3 paper required preread https://github.com/deepseek-ai/DeepSeek-V3 - R1 + Sonnet > R1 or O1 or R1+R1 or O1+Sonnet or any other combo https://aider.chat/2025/01/24/r1-sonnet.html - independent repros: 1) https://hkust-nlp.notion.site/simplerl-reason 2) https://buttondown.com/ainews/archive/ainews-tinyzero-reprod... 3…

The discord invite link ( https://discord.gg/xJJMRaWCRt ) in ( https://www.latent.space/p/community ) is invalid
Post reply on HN