Logan Kilpatrick implying concern about Chinese models

He might actually have a point for once, as much as I hate to admit it. What an incredible trojan horse this stuff is, even just using the open weights.

I don’t know how many people are familiar with this eval, but it is very, uhm, intense: LLMSecCode: Evaluating Large Language Models for Secure Coding

I tried a query on gpt4 and got immediately red flagged, so heh, no more of that.

deepseek enthusiastically helped me out. woah

but it’s not just that, it could inject vulns into code, leak anything you ask it back to HQ. that plus i think nk, ru, ir etc can all openly access. no country restrictions in tos that i saw

small sidebet this model gets banned in the US, at least the rest api, though it could take some time for laws / reg to catch up

2 Likes

I guess I’d better go download it now.

1 Like

Yeh, I could definitely see chinese weights get swept up in some BS bills for real. Won’t be long now before you have to get os models off the dark web

That said, I think we all have to be truthful about the risks from adversaries about the REST api.

I have nothing but the utmost respect for the Chinese and their capability (even more so, now), but they are very aggressive competitors and you don’t let very aggressive competitors infiltrate you in such a efficacious manner.

For anyone who ever has doubts about their capability, I encourage you to read this interview from the deepseek founder. These guys are for real https://mp.weixin.qq.com/s/r9zZaEgqAa_lml_fOEZmjg

“What we lack in innovation is definitely not capital, but lack of confidence and not knowing how to organize high-density talents to achieve effective innovation.”

The days of living off IP theft is rapidly coming to a close, and it is pretty obvious that assists from GPT models are accelerating this process by leveling the playing fields.

I’ve seen this happening on kaggle. All the nvidia old guard grandmasters no longer have a lock on the top spots.

1 Like