Hacker News Reader

Comment by reissbaker | Hacker News Reader

It's a 7B "unified model" LLM/VLM (not a diffusion model!) that out-benchmarks Dall-E 3 and Stable Diffusion Medium. It's released under the DeepSeek License, which is pretty-open license that allows commercial use but restricts military use, along with a few other content-based restrictions.

View on HN

> restricts military use

I'm sure the powers-that-be will absolutely pay attention to that clause.

Yeah, they should! Not that the missile then makes a 180° turn to "return to sender" because it noticed that the target is a Chinese military base.

The code is open sourced

but what about training data?

The weights and data pipeline are open sourced and described explicitly in the paper they published. The non-reasoning data isn't nearly as interesting as the reasoning data though

See Sleeper Agents (https://arxiv.org/abs/2401.05566).

Who in their right mind is going to blindly take the code output by a large language model and toss it on a cruise missile? Sleeper agents are trivially circumvented by even a modicum of human oversight.

There's no meaningful inspection of LLM code, because the real code is the model weights.

You could say the same for the GPL, yet it's wording is enough to curb adoption from corporations.

Large organisations like the military have enough checks and balances to avoid these kind of licences with a 10ft pole.

How are these licenses enforceable?

Lawsuits, but it's mainly just CYA for DeepSeek; I doubt they truly are going to attempt to enforce much. I only mentioned it because it's technically not FOSS due to the content restrictions (but it's one of the most-open licenses in the industry; i.e. more open than Llama licenses, which restrict Meta's largest competitors from using Llama at all).

I've always wondered why nobody has tried to scale image-generation models to modern LLM sizes, such as 200-500B parameters instead of 1-7B...

Reply on news.ycombinator.com