DeepSeek: Inference-Time Scaling for Generalist Reward Modeling
22 comments
·April 4, 2025ALLTaken
wood_spirit
When OpenAI surged ahead Meta ended up giving away its incredibly expensive to make llama model to reduce the OpenAI valuations.
Is DeepSeeks openness in part to reduce the big American tech companies?
ALLTaken
Correlation isn't causation, I hate to say this, but here's really applicable. Facebook aka Meta has always been very opensource. Let's not talk about the license though. :)
Why do you imply malice in OSS companies? Or for profit companies opensourcing their models and sourcecode?
mwigdahl
Personally I don't impute any malice whatsoever -- these are soulless corporate entities -- but a for-profit company with fiduciary duty to shareholders releasing expensive, in-house-developed intellectual property for free certainly deserves some scrutiny.
I tend to believe this is a "commoditize your complement" strategy on Meta's part, myself. No idea what Deepseek's motivation is, but it wouldn't surprise me if it was a similar strategy.
throwaway314155
Meta is decidedly not an "OSS company" no matter how much they put out.
phoronixrly
If only totalitarian nation states used their subjects' money to undermine the dominance of US-based software vendors by releasing open-source alternatives created with slave labour... Oh wait, it can't work because software patents are here to the rescue again ... Wait, open source is communism? Always has been. /s
Febra33
> Let's not confuse the company with the country
What's wrong with China? They're wonderful in the OSS ecosystem.
echelon
It varies on a company to company basis. BOOX, for instance, are notorious GPL violators.
There's also significant alpha in releasing open weights models. You get to slow down the market leaders to make sure they don't have runaway success. It reduces moats, slows funding, creates a wealth of competition, reduces margin. It's a really smart move if you want to make sure there's a future where you can compete with Google, OpenAI, etc. There's even a chance it makes those companies bleed a little. The value chain moves to differently shaped companies (tools, infra) leaving space for consumer and product to not necessarily be won by the "labs" companies.
refulgentis
I love open source and the general vibe of good vibes you're bringing, but...this isn't SOTA, or close, even on the papers own terms. (i.e. excluding models released the last 6 months, including their own, which is a strange, yet understandable, choice given the results they report)
Quickest way to show this:
- Table 2, top of page 7
- Gemma 2 27B, 0 interventions, has 94.1/56.6/60.2
- Gemma 2 27B, with all their interventions, has 86/64/69.
- Gemma 2 27B, with all their interventions, sampled 32 times, is at 90.4/67.2/70.3.
- Gemma 2 27B came out in...June 2024. :/
Quick heuristics employed here:
- What models did they compare against? (this isn't strictly an issue, the big screaming tell is "What models did they compare against compared to their last N papers?"
- How quickly does the paper have to move towards N samples, and how big does N get before they're happy enough to conclude? (32). How much does that improve performance on their chosen metric? (1.8%)
resters
DeepSeek R1 is by far the best at writing prose of any model, including Grok-3, GPT-4o, o1-pro, o3, claude, etc.
Paste in a snippet from a book and ask the model to continue the story in the style of the snippet. It's surprising how bad most of the models are.
Grok-3 comes in a close second, likely because it is actually DeepSeek R1 with a few mods behind the scenes.
vessenes
why do you think that grok 3 is deepseek, out of curiosity?
resters
I replied to the child of your comment
azinman2
Yes that’s a pretty giant accusation, especially given they’re buying boatloads of GPUs and have previous versions as well (it’s not like they’re starting with 3).
resters
1) Grok-2 was akin to GPT-3.5
2) Grok-3 comes out a month after DeepSeek R1 was open sourced. I think Grok-3 is DeepSeek R1 with some added params and about a month of training on the giant cluster, possibly a bit of in-house secret sauce added to the model or training methodology.
What are the chances that XAI just happened to have a thinking model close to as good as revolutionary DeepSeek but happened to launch it 30 days later?
It was both smart and pragmatic for XAI to simply use the best available open source stuff and layer their own stuff on top of it. Imagine they doubled the parameter count and trained it for 30 days, that would not even use half of the GPU power!
mentalgear
Happy to see deekseek using the correct (and much more idiomatic) term "inference-time scaling", instead of the grotesque construction of "test-time compute" that openAI came up with.
bilsbie
Any idea why I lost interest in deep seek? I used it and grok3 a whole bunch when they first came out but now I’ve fallen back to Claude for everything.
manmal
For coding, I‘m finding Claude‘s responses most to the point and on-task. While many other models try to extrapolate or lecture or patronize. DeepSeek is pretty good though. Maybe it’s the high latency (probably due to prompt processing)?
ftbsqcfjm
Interesting work on open-ending language models to foster imagination and narrative generation. The idea of role-playing as different characters is novel. I wonder how well it would generalize to non-fantasy domains and if the lack of grounding could lead to hallucinations. Excited to see where this research goes!
NitpickLawyer
> The idea of role-playing as different characters is novel.
It is not. I remember Karpathy being really excited about the "1 million gpt personas" dataset and highlighted it as a way to avoid reward hacking in RLAIF. That was 3-6 months ago I believe.
Of course paper / code / weights beats idea, and it's exciting to see how far this can go.
Not jus being impressed that every paper coming out is SOTA, but also leads the way in being Open-Source in the pure definition of OSS, even with permissible licensing.
Let's not confuse the company with the country by over-fitting a narrative. Popular media is reenforcing hatred or anything that sponsors them, especially to weaker groups. Less repercussions and more clicks/money to be made I guess.
While Politicians may hate each other, Scientists love to work with other aspiring Scientists who have similar ambitions and the only competition is in achieving measurable success and the reward it means to the greater public.
Without any bias, but it's genuinely admirable when companies release their sources to enable faster scientific progress cycles. It's ironic that this company is dedicated to finance, yet shares their progress, while non-profits and companies dedicated purely to AI are locking all knowledge about their findings from access.
Are there other companies like DeepSeek that you know of that commonly release great papers? I am following Mistral already, but I'd love to enrich my sources of publications that I consume. Highly appreciated!