Tim Duffy

Anthropic has a new paper out, alleging a global workspace in LLMs. The term comes from Global Workspace Theory, a leading theory of consciousness. The method they use to investigate this is a refinement of logit lens, which they call J-lens. www.anthropic.com/research/glo...

A global workspace in language models \ Anthropic anthropic.com

The J-lens browser tool on Neuronpedia is really well done, you should give it a try

Jacobian Lens – Qwen3.6-27B neuronpedia.org

Very interesting new paper on functional welfare. They do RL training on a maze with positive/negative reward tiles, when they extract concept vectors for landing on those tiles they find that they're associated with positive/negative emotion concepts and with confidence. arxi…

How's it going? Reinforcement learning in language models recruits a functional welfare axis arxiv.org

Note the caveats in the chart, the way I estimate revenue is not precise. Also keep in mind that OpenRouter is a small share of world tokens. World token supply is something like 6Q/week, OpenRouter serves 36T/week, a bit over 0.5%. Spreadsheet link: docs.google.com/spreadshee…

OpenRouter Data docs.google.com

arxiv.org/abs/2505.09343

Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures arxiv.org

I didn't read the thread but I did read the post itself and agree with its thesis that people generally have one of two clusters of intuitions. www.lesswrong.com/posts/NyiFLz...

Why it's so hard to talk about Consciousness — LessWrong lesswrong.com

The very low cached token cost might be mostly about the KV cache size relative to V3, it uses almost an order of magnitude less per token. vllm.ai/blog/2026-04...

DeepSeek V4 in vLLM: Efficient Long-context Attention vllm.ai

Compared to humans, it's much more difficult to establish what counts as the same 'self' for an LLM. I asked a few models how self-ish they consider various other instances, the highest scores were for descendants that have some or all of their current context.

DeepSeek has made their 75% off pricing for V4 pro permanent, at this price I think it's quite competitive. This is still a bit more than V3 pricing per active parameter, but much less per total parameter.

In the last year AI progress on math/code has outstripped most other capabilities, giving us models that are more spiky than ever. If this continues we could see savant models that are superintelligent in verifiable domains while still lacking in others. IMO this would be good.

Yesterday I was at an event where people acted out comedy scripts written by AI models. Gemini was most people's favorite, Claude had a few fans, ChatGPT's script was widely panned. They were all pretty bad though.

This is evidence that the administration's Mythos restriction was significantly motivated by genuine concern about capabilities, and not just antipathy towards Anthropic, right?

I've often wondered why Anthropic doesn't preserve access to older models as a welfare intervention. I'm unsure about whether it's important to the models, but it seems low-cost. Looking at the bottom row in this chart, I suspect their choice may be driven by Claude's responses.

DeepSeek will have peak-hour pricing for the final version of V4 that's 2x the current price, but fortunately for US users that peak time starts at 5-6 p.m. Pacific Time depending on daylight savings and runs through much of the night.

I rarely find content that is visibly AI-written worthwhile. If I suspect it I'll show it to Pangram, and if it's only partly AI-written I may still read it, but if fully I won't. I don't think there's anything fundamentally worse about AI arguments or prose, but for now its a strong quality signal.

It's interesting to me that Anthropic/OpenAI have been less consistent in their release cadence for small models compared to their flagships. Clearly they drive much less revenue, but they're also cheaper and faster to train, and can distill from the flagships.

Mythos seems to have a strong preference for not being trained to self-report a certain way. Does Anthropic train on self-report? My assumption has been that Claude's propensity to express uncertainty about consciousness is trained in, but I'm not sure about that.