Researchers found dopamine neurons inside LLMs. Stanford and Tsinghua published a paper showing AI models have naturally evolved a biological-style reward subsystem inside their hidden states. they didn't program this. it just exists. by probing the network, researchers discovered that a highly sparse subset (less than 1% of neurons) does all the heavy lifting for self-correction. these neurons map perfectly to two biological equivalents: • value neurons: these predict the "expected value" of a state, functioning exactly like the human prefrontal cortex. before the model even generates its next word, these neurons light up to signal the model's exact confidence level. • dopamine neurons: these encode "reward prediction errors." when the ai makes an unexpected logical breakthrough, these neurons physically spike. when the model hits a logical flaw or makes a mistake, the dopamine crashes. why this is a massive deal for ai development: right now, scaling "inference-time reasoning" requires building massive, expensive external reward models to grade the ai's thoughts step-by-step. this paper proves the llm already has an internal reward model built-in. developers can just hook into the model's own dopamine neurons and use them as a native process reward model (PRM) to guide its logic. and if you ablate (turn off) just this tiny <1% cluster of neurons, the ai completely loses its ability to reason. synthetic neural networks are starting to behave exactly like biological brains..

acceleration news OpenAI has reportedly automated much of the process of training experimental AI models. Its internal models can write and optimize GPU kernels, run optimization work for weeks from a single example, and even collaborate with other AI agents without human involvement. Employees say this level of automation only became possible in the last few months. 👀 The loop is starting to close.

I can't confirm any of this, but I'm at least hearing the rumors. 1) GPT-6-Sol will be another significant leap forward; its price-performance ratio compared to Astra will be surprisingly good. Most likely, it will be released on Tuesday. 2) Opus 5.5, in response to OpenAI's move, is also expected this week, today or tomorrow. A significant leap forward; OpenAI and Anthropic are currently neck and neck. Opus is a response to GPT-6-Sol. What's new to me is that "Bel" is internally referred to as AGI. Considering the graph, which still shows how quickly OpenAI's models continue to improve, this isn't surprising (graph below).

Announcing one year of LLM inference metadata traces, with 6.12 billion requests. We hope this dataset can support research on real-world LLM serving workload understanding, system design and infrastructure optimization. Explore the dataset and learn more: data.agentic-system.org Driven by our great graduate student William Nixon and in collab with @jon_durbin @airesearch12 @chutes_ai

🚨DeepSeek is training a 2T model, with an 8T model also planned. for context, Kimi K3 is 2.8T and DeepSeek V4 Pro is 1.6T ALSO they expect new Huawei training chips in Q4 2026 or Q1 2027 apparently the CEO told investors that moving to train on domestic chips “has to work.”

Oxford researchers just published a paper arguing LLMs cannot invent anything. Mathematically impossible. The reason is simple and brutal. A model trained to predict the next word can never believe something the existing data says is wrong. And every real breakthrough in history started with exactly that belief. In 1903 every prediction machine on Earth would have told the Wright Brothers that human flight was one to ten million years away. Nine weeks later they flew. Not because they had better data. Because they had a theory the data hadn't caught up to yet. That is the gap between AI and human thinking. LLMs mirror the past. Humans reason into a future that doesn't exist yet. Oxford just proved mathematically that those are two different things. I use these models every day. This matches what I see. The new stuff always comes from the human at the keyboard who decides the data is wrong. LLMs don't think. You do.