DeepSeek: Inference-Time Scaling for Generalist Reward Modeling

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.

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%)

> 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.

In this case there are very few truly "OSS companies" except for Red Hat and few other Linux distribution maintainers. Even companies centered around open source like Gitlab are usually generate most of their revenue of proprietary products or use liceses like BSL.

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!