This is particularly relevant because the guy used Andrej Karphaty's "autoresearch" [1] approach, which is extremely powerful when applied to problems that generalize well to it.
Personally, I've tried to squeeze more tok/s for a single DGX Spark deployment and DeepSeek V4 Flash but only got marginal improvements. There's work to do on fusing kernels and other optimizations that are already on antirez's roadmap so it is not worth duplicating efforts.
I've had positive experiences running GLM 4.7 via vLLM, tool calling works well and the inference is fast. Do you run DeepSeek V4 Flash on vLLM?
Depends of what the proof of concept is about. It could be just a toy example, e.g. a RCE that opens the calculator app or something much more nefarious, like returning a root shell and would still fall under the definition of PoC.
I was able to use Fable to generate PoC for several classes of vulnerabilities and I didn't observe the model refusing to engage in detailed analysis to come up with creative approaches, the very contrary.
> I used a fork of oh-my-pi
Why not use the leaked claude code source? Not that you really need it to execute the jailbreak
> We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities.
"small number", "previously known"+"minor"... they are trying hard to characterize this as harmless.
> These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass.
Ah so now they are admitting that this is all about hype after all.