buppermint

CoT is both correlated and causal of the model's real computations, it's just imperfect. If you manually add "Let's wrap it up" in the CoT during generation, most LLMs will actually wrap it up (this is a commonly used trick in local LLM circles to get long-winded LLMs to stop reasoning). This wouldn't work if CoT text didn't affect the actual internal model logic.

The paper title is a bit misleading. The tested detectors and models here are small and rather dated (Llama 3.1 8B and Gemini Flash 2.0 - these are basically in the level of a modern 1B model), and the actual paper says this only shows vulnerability in small model systems.

Worth seeing the whole chart in perspective:

https://fred.stlouisfed.org/series/IHLIDXUSTPSOFTDEVE

Anthropic has already has lower guardrails for DoD usage: https://www.theverge.com/ai-artificial-intelligence/680465/a...

It's interesting to me that a company that claims to be all about the public good:

- Sells LLMs for military usage + collaborates with Palantir

- Releases by far the least useful research of all the major US and Chinese labs, minus vanity interp projects from their interns

- Is the only major lab in the world that releases zero open weight models

- Actively lobbies to restrict Americans from access to open weight models

- Discloses zero information on safety training despite this supposedly being the whole reason for their existence

I would disagree on the knowledge sharing. They're the only major AI company that's released zero open weight models. Nor do they share any research regarding safety training, even though that's supposedly the whole reason for their existence.

I've been playing around with this in z-ai and I'm very impressed. For my math/research heavy applications it is up there with GPT-5.2 thinking and Gemini 3 Pro. And its well ahead of K2 thinking and Opus 4.5.

From a quick read, this is cool but maybe a little overstated. From Figure 3, completely suppressing these neurons only reduces hallucinations by like ~5% compared to their normal state.

Table 1 is even more odd, H-neurons predicts hallucination ~75% of the time but a similar % of random neurons predict hallucinations ~60% of the time, which doesn't seem like a huge difference to me.

This is because Grok Code Fast is free via Kilo Code/Cline and has been for months

Very cool. Any plans to add support for local models? This has what has prevented us from adopting Positron so far. We have sensitive data and sending to third party APIs is not an option (regardless of their stated retention policies).