DeepSeek: Inference-Time Scaling for Generalist Reward Modeling
arxiv.org
arxiv.org
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!
Is DeepSeeks openness in part to reduce the big American tech companies?
Why do you imply malice in OSS companies? Or for profit companies opensourcing their models and sourcecode?
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.
Okay then. Fine by me.
> Gitlab
Perfect example. They have OSS offerings. They are not an OSS _company_.
This also serves to exclude the hundreds of VC-backed "totally open source 100% not going to enshittify this when our investors come asking for returns". Which, again, I'm fine with.
The business model of the purist OSS company is not one that's been found to be terribly successful. Nevertheless, it _is_ one which has a sort of moral high ground at least. I would prefer to leave definitions as is so as to keep that distinction (of having the moral high ground) crystal clear.
Does that make sense?
What's wrong with China? They're wonderful in the OSS ecosystem.
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.
But believing a man could achieve such a feat alone is inspiring to be frank.
It's very difficult to be truly unbiased and neutral and it's not my goal, I just think it's a common thought, that needs to be challenged. To associate products/results of scientists, quants, engineers and companies they are employed with an entire Nation is inherently simplistic.
In that case, why did the CIA/NSA develop TOR and made it OSS? If the governments in the UK/France/Turkey are so brutally against encryption, why does the USA release safe encryption products?
If the world were absolute, we would absolutely be doomed and I hope to be part of a world, where freedom of thought, responsibility of each, constructive cooperation and a mesh of companies can work and produce value from and with each other permissionlessly. A world where Copyright/Patents are not needed anymore, because a stronger framework supports the individual contributor and also companies. Leftist, Right and Centrists views how an economy should look like are flawed, because they introduce idealogies to a mathematical non-linear partially closed but mostly open system.
Every idealistic concept shouldn't be believed, but explored. To hate one system over another one is also flawed, because it doesn't produce data and forces hypothesis testing without consequentially following conclusions. Economy is too complex for a man to design. It shouldn't be put into a canvas of restricted operations, but circuits would need to be developed locally. If we empower small communities and allow changes to be made quicker with less bureaucracy, this seemingly grand introduction of chaos leads to emergence of a larger stability of the whole. We are soo far away from that man..
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%)
I think what we are all really excited about having finally AI at home and being unchained and freed from a central SaaS controlling all the AI is ever going to tell you.
So, 6-7y ago google had these AI Chats internally and never intended to release it, a friendly googler told me.
Then ChatGPT came along and locked you into their SaaS. That was fantastic in the beginning, but the more you used the AI, the more you felt helpless, swound by anyone who may have access to an AI at OpenAI that is unfiltered and uses the full power of the model. Then came the jailbreak and accounts being banned for using it.
Then came the freedom by LLAMA and DeepSeek and waves of otheres. It rolled into your laptop real quick and this freedom is priceless! Something we should be really thankful for that it happened and support more OSS.
Google and Facebook would never share their trove of data with us ever and very few people have enough storage and compute to even attempt to replicate them. But their Data Dominance doesn't protect them anymore. Once the models became intelligent enough to slurp up large chunks of the web, they became a better search, a better teacher and a better experience than sponsored ads, with ads with internal google/bing products listed up, then SEO websites and somewhere hidden what we really were looking for. Or often.. just being deleted for copyright and other reasons.
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.
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!
Extremely, extremely good. That was in fact the real point of the deepseek paper - it was extremely cheap to turn a frontier(ish?) model into a reasoning model. There is nothing suspicious about this timeline from an ML Ops point of view.
In fact DeepSeek themselves in a sort of victory lap released six OTHER models from other providers finetuned with reasoning as part of the initial drop.
DS3: 5M training run Grok3: 400M training run
for 2% difference in the benchmarks.
Thinking about it more, I think a big part of it is that correct answers are not a limiting factor for me it feels like. Claude is good enough and it is more what to do with all these correct answers is my problem. I am also naturally biased to a model if paying for it.
Gemma is by far the best at giving advice and planning ones days and life priorities. Not sure how to benchmark that.
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.