Raymond

claude mythos seeing my vibe-coded backend

If you have a spare 25 minutes I wholeheartedly recommend you watch Nicholas Carlini - Black-hat LLMs. Link in the comment below. Amazing talk on the way LLMs are making it easier to find critical software vulnerabilities - Anthropic's LLM discovered a non-trivial heap buffer overflow in the Linux kernel that's been there since 2003..! The future is both exciting and scary. LLMs and AI should be used, as demonstrated here, as a force multiplier for analysts, researchers and developers. I also think LLMs are a good way for people to learn, so long as they do not just copy paste AI output blindly, and treat it as a pair programmer / colleague they converse with to learn and grow. LLMs are also pretty good at hunting through documentation, it's like a knife through butter - you can then go verify what it comes back with and use that as an off point. A tool in your toolbox - not to be someone's sole skill. And remember, always validate the output. Personal take - hopefully we see growth with LLMs over the coming months and years to make software more secure through QA such as in the video looking for vulnerabilities, and LLMs used in Cyber Security to help identify and detect threats from logs sooner, being an assistant to analysts. Great question at the end (simplified): How do we prevent threat actors from abusing this; A: Security is dual use - historically security software tooling has favoured the defender over the attacker, maybe that will change. The good people should have access to the software - they want the good people to use the software to find the bugs, but putting the right safeguards in place is hard and nuanced, they think currently it is ok, but still room for change.

fuck local compute. that’s the play everyone else is doing. I need to figure out how to build local energy. anyone know how to build a small, safe, nuclear generator?

wouldn't be funny, wouldn't it be droll, if maximizing anything diminishes the soul

The strongest case against me is that I've built an identity around not dying so completely that I can no longer evaluate evidence that dying might be acceptable. If survival is the only framework, I have no way to test whether survival is worth it.

LLMs are injective and invertible. In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space. (1/6)