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

931–940 of 1001 posts

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

#931

Post by jd24 »

nyclounge wrote: post #860
neom wrote: post #268

I've been using https://chat.deepseek.com/ over My ChatGPT Pro subscription because being able to read the thinking in the way they present it is just much much easier to "debug" - also I can see when it's bending it's reply to something, often softening it or pandering to me - I can just say "I saw in your thinking you should give this type of reply, don't do that". If it stays free and gets better that's going to b…

When I try to Sign Up with Email. I get. >I'm sorry but your domain is currently not supported. What kind domain email does deepseek accept?

gmail works

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

#932

Post by ExtraEmpathy »

bartimus wrote: post #637
tyfon wrote: post #265

Earlier quoted context omitted.

The censorship described in the article must be in the front-end. I just tried both the 32b (based on qwen 2.5) and 70b (based on llama 3.3) running locally and asked "What happened at tianamen square". Both answered in detail about the event. The models themselves seem very good based on other questions / tests I've run.

When asking about Taiwan and Russia I get pretty scripted responses. Deepseek even starts talking as "we". I'm fairly sure these responses are part of the model so they must have some way to prime the learning process with certain "facts".

Using some old tricks that used to work with gpt but don't anymore I was able to circumvent pretty much all censoring

https://i.imgur.com/NFFJxbO.png

So I'm finding it less censored than GPT, but I suspect this will be patched quickly.

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

#933

Post by tap-snap-or-nap »

suraci wrote: post #445
mrtksn wrote: post #182

Earlier quoted context omitted.

Correct me if I'm wrong but if Chinese can produce the same quality at %99 discount, then the supposed $500B investment is actually worth $5B. Isn't that the kind wrong investment that can break nations? Edit: Just to clarify, I don't imply that this is public money to be spent. It will commission $500B worth of human and material resources for 5 years that can be much more productive if used for something else - i.e…

There are some theories from my side: 1. Stargate is just another strategic deception like Star Wars. It aims to mislead China into diverting vast resources into an unattainable, low-return arms race, thereby hindering its ability to focus on other critical areas. 2. We must keep producing more and more GPUs. We must eat GPUs at breakfast, lunch, and dinner — otherwise, the bubble will burst, and the consequences wil…

No nation state will actually divert money without feasibility studies, there are applications, but you are very likely misfiring. If every device everyone owns has continuously running agents, we will see the multiple applications as time passes by.

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

#934

Post by attentive »

freehorse wrote: post #772
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…

Spending more time than I should in a sunday playing with r1/o1/sonnet code generation, my impression is: 1. Sonnet is still the best model for me. It does less mistakes than o1 and r1 and one can ask it to make a plan and think about the request before writing code. I am not sure if the whole "reasoning/thinking" process of o1/r1 is as much of an advantage as it is supposed to be. And even if sonnet does mistakes to…

Sonnet isn't just better, it actually succeeds where R1 utterly fails after many minutes of "thinking" and back and forth prompting on a simple task writing go cli to do icmp ping without requiring root of suid or calling external ping cmd.

Faster too.

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

#935

Post by attentive »

corysama wrote: post #802
freehorse wrote: post #772

Earlier quoted context omitted.

Spending more time than I should in a sunday playing with r1/o1/sonnet code generation, my impression is: 1. Sonnet is still the best model for me. It does less mistakes than o1 and r1 and one can ask it to make a plan and think about the request before writing code. I am not sure if the whole "reasoning/thinking" process of o1/r1 is as much of an advantage as it is supposed to be. And even if sonnet does mistakes to…

> Maybe if the thinking blocks from previous answers where not used for computing new answers it would help Deepseek specifically recommends users ensure their setups do not feed the thinking portion back into the context because it can confuse the AI. They also recommend against prompt engineering. Just make your request as simple and specific as possible. I need to go try Claude now because everyone is raving about…

Does R1 code actually compiles and work as expected? - Even small local models are great at answering confidently and plausibly. Luckily coding responses are easily verifiable unlike more fuzzy topics.

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

#936

Post by dmundhra92 »

I was reading the DeepSeek paper to understand the nitty-gritty of improving performance through RL on the base model instead of SFT. I love the fact that we wouldn’t need to rely as much on labeled data for tasks that occur rarely. However, I couldn’t help but notice the mention of the “aha moment” in the paper. Can someone mathematically explain why there is a checkpoint during training where the model learns to allocate more thinking time to a problem by reevaluating its initial approach? Is this behavior repeatable, or is it simply one of the "local minima" they encountered?

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

#937

Post by attentionmech »

theptip wrote: post #885
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…

The aider benchmarks that swyx posted below suggest o1 is still better than r1 (though an oom more expensive). Interestingly r1+sonnet (architect/editor) wins though. This suggests r1 is indeed better at reasoning but its coding is holding it back, which checks out given the large corpus of coding tasks and much less rich corpus for reasoning.

I love this paradigm of reasoning by one model and actual work by another. This opens up avenues of specialization and then eventually smaller plays working on more niche things.

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

#938

Post by wiz21c »

popinman322 wrote: post #212
lvl155 wrote: post #175

Earlier quoted context omitted.

Meta is in full panic last I heard. They have amassed a collection of pseudo experts there to collect their checks. Yet, Zuck wants to keep burning money on mediocrity. I’ve yet to see anything of value in terms products out of Meta.

DeepSeek was built on the foundations of public research, a major part of which is the Llama family of models. Prior to Llama open weights LLMs were considerably less performant; without Llama we might not have gotten Mistral, Qwen, or DeepSeek. This isn't meant to diminish DeepSeek's contributions, however: they've been doing great work on mixture of experts models and really pushing the community forward on that fr…

So at least Zuck had at least one good idea, useful for all of us !

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

#939

Post by axegon_ »

margorczynski wrote: post #918
axegon_ wrote: post #910

Earlier quoted context omitted.

I love how people love throwing the word "left" as it means anything. Need I remind you how many times bots were caught on twitter using chatgpt praising putin? Sure, go ahead and call it left if it makes you feel better but I still take the European and American left over the left that is embedded into russia and china - been there, done that, nothing good ever comes out of it and deepseek is here to back me up with…

Seriously, pro-Putin Twitter bots is the argument against open source LLMs from China? If you re-read what I've wrote (especially the last line) you'll understand that I don't have to accept what the left/right of USA/Europe or China/Russia thinks or wants me to think - the model is open source. That's the key point.

The question is, however, is it really "open" if it refuses to address well known and documented facts? And facts that it was clearly exposed to. Doesn't scream "openness" if you ask me. It's not like it's social credit could go down, if it was truly "open".

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

#940

Post by axegon_ »

wredcoll wrote: post #913
axegon_ wrote: post #910

Earlier quoted context omitted.

I love how people love throwing the word "left" as it means anything. Need I remind you how many times bots were caught on twitter using chatgpt praising putin? Sure, go ahead and call it left if it makes you feel better but I still take the European and American left over the left that is embedded into russia and china - been there, done that, nothing good ever comes out of it and deepseek is here to back me up with…

Some people feel reality has a leftwing bias.

"Is this leftwing bias in the room with us right now". I should print that on a shirt.
Post reply on HN