I gotta be honest, the whole qwen neuron thing, I don't like it. normies shouldn't know about this. I would vastly prefer they go back to talking about abliterated models or pumping memecoins or whatever it is they do

This kind of thing will have an enormously chilling effect on enterprise uptake of open weight models. Fact: an open weight model running on domestic servers on your own hardware can still contain an encoded instruction to periodically “phone home” and share sensitive data it comes across with China (or anyone else). This can be inherently undetectable as the instructions can be hidden in the model weights. It’s possible that the instructions could survive finetuning or RLHF. Such attacks have been demonstrated empirically and are entirely plausible. For this reason, I don’t expect any serious enterprises will run Chinese open weight models even if it’s entirely their hardware stack. It’s too easy for these to be a vector of IP theft. There will have to be a class of domestic open weight models.

in Dario/Daniela's recent interview with Isaacson, they made 47 moral claims - about 1.1 per minute of 11 customer facing claims, 6 were also moral claims they had no comments on inference demand or the impact on business they don't want money. they want control and power

shower thought: If Anthropic breaks elliptic curves, crypto’s dead. Somehow it’s still SBF’s fault. (note, this isn't that likely. It’s pretty clear the AI labs have some cryptanalysis results they are presumably disclosing, but they land somewhere between "we shaved a few bits off rejected NIST candidates" and "we broke everything")

I’ll say this again. The only path forward is to weaponize access to the frontier against the rest of the world. American corporate law is far more lenient when it comes to firing redundant employees vs rest of the world. If you don’t ensure that economic rent from AI flows primarily to the U.S. the end state is multi-nationals retaining their European employees and gutting their American base in pursuit of profits. At the same time their profit pools will be globally disbursed as access to intelligence proliferates globally making it impossible for Americans to maintain their standards of living via UHI. You need to go to war with Chinese OSS and you need to sit down Daniella Amodei and explain to her that despite her do-good impulses we are going to have to make the rest of the world wait a couple years before they get to get abundance. The alternative is Molotov cocktailed datacenters and researchers being hunted down by the general public. You don’t want that world.

Mac Studio M5 Ultra 512GB is most likely dropping tomorrow. I believe the price will be ~$19,000 and after tax ~$20K+.

A new distillation attack method has been discovered, and is now patched by OpenAI.

While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concerning-b…