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

561–570 of 1001 posts

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

#561

Post by SparkyMcUnicorn »

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.

Kagi is using fireworks.ai according to the docs, which is the 685B model.

Kagi and Aider benchmarks definitely put R1 in the lead, but I have yet to try them out.

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

#562

Post by buyucu »

nejsjsjsbsb wrote: post #393

This might tempt me to get a graphics card and run local. What do I need minimum to run it?

I'm running the 32b version on a home server (Ryzen 7) using llama.cpp, which uses Vulkan for hardware acceleration on the iGPU.

For hobbyist inference, getting a iGPU with lots of system ram is probably better than getting a dedicated Nvidia gpu.

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

#563

Post by SparkyMcUnicorn »

littlestymaar wrote: post #548
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.

> but I tried Deepseek R1 via Kagi assistant Do you know which version it uses? Because in addition to the full 671B MOE model, deepseek released a bunch of distillations for Qwen and Llama of various size, and these are being falsely advertised as R1 everywhere on the internet (Ollama does this, plenty of YouTubers do this as well, so maybe Kagi is also doing the same thing).

They're using it via fireworks.ai, which is the 685B model.

https://fireworks.ai/models/fireworks/deepseek-r1

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

#564

Post by bboygravity »

littlestymaar wrote: post #548
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.

> but I tried Deepseek R1 via Kagi assistant Do you know which version it uses? Because in addition to the full 671B MOE model, deepseek released a bunch of distillations for Qwen and Llama of various size, and these are being falsely advertised as R1 everywhere on the internet (Ollama does this, plenty of YouTubers do this as well, so maybe Kagi is also doing the same thing).

Ah interesting to know that. I don't know which version Kagi uses, but it has to be the wrong version as it's really not good.

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

#565

Post by bazmattaz »

siltcakes wrote: post #140
jvanderbot wrote: post #11

Earlier quoted context omitted.

How likely is this? Just a cursory probing of deepseek yields all kinds of censoring of topics. Isn't it just as likely Chinese sponsors of this have incentivized and sponsored an undercutting of prices so that a more favorable LLM is preferred on the market? Think about it, this is something they are willing to do with other industries. And, if LLMs are going to be engineering accelerators as the world believes, the…

I trust China a lot more than Meta and my own early tests do indeed show that Deepseek is far less censored than Llama.

Interesting. What topics are censored on Llama?

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

#566

Post by XorNot »

roenxi wrote: post #480
baq wrote: post #311

Earlier quoted context omitted.

Side note: I’ve read enough sci-fi to know that letting rich people live much longer than not rich is a recipe for a dystopian disaster. The world needs incompetent heirs to waste most of their inheritance, otherwise the civilization collapses to some kind of feudal nightmare.

Reasoning from science fiction isn't a particularly strong approach. And every possible future is distopian - even the present is distopian in a practical sense. We have billions of people who live well below any standard I woudl consider acceptable.

Reasoning from science fiction is just stupid. A story first and foremost has to have conflict: if it doesn't there is no story, and thus all the stories have one.

Science fiction also follows the anxieties of the time it is written in, as well as the conventions of the subgenre it's representing: i.e Star Trek doesn't have drones or remote surveillance really. Though it does accidentally have LLMs (via the concept of holodeck characters).

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

#567

Post by singularity2001 »

magicalhippo wrote: post #188
buyucu wrote: post #131

I'm impressed by not only how good deepseek r1 is, but also how good the smaller distillations are. qwen-based 7b distillation of deepseek r1 is a great model too. the 32b distillation just became the default model for my home server.

I just tries the distilled 8b Llama variant, and it had very poor prompt adherence. It also reasoned its way to an incorrect answer, to a question plain Llama 3.1 8b got fairly correct. So far not impressed, but will play with the qwen ones tomorrow.

not adhering to system prompts is even officially mentioned as one of the caveats of the distilled models

I wonder if this has to do with their censorship agenda but other report that it can be easily circumvented

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

#568

Post by singularity2001 »

piecerough wrote: post #223
andix wrote: post #218

I was completely surprised that the reasoning comes from within the model. When using gpt-o1 I thought it's actually some optimized multi-prompt chain, hidden behind an API endpoint. Something like: collect some thoughts about this input; review the thoughts you created; create more thoughts if needed or provide a final answer; ...

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.

One big question will be whether chain of thought within the embedding space will work better than in the token space.

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

#569

Post by riffraff »

just-ok wrote: post #552
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…

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.

But it took the deepseek team a few weeks to replicate something at least close to o1.

If people can replicate 90% of your product in 6 weeks you have competition.

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

#570

Post by Mond_ »

just-ok wrote: post #552
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…

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.

Post reply on HN