TheSequence

Imagine bringing a robot into your kitchen and saying, “Help me clean up after dinner.” You have just compressed a remarkable amount of engineering into six words. The robot must distinguish leftovers from rubbish, discover where plates belong, and work out why a drawer refuses to close. Eventually, someone will hand it a wineglass. Everyone will suddenly become very interested in the quality of its training data. That kitchen captures the promise of a ChatGPT moment for robotics.

We are back with our interview series and a very special guest today! Anastasios Angelopoulos has been at the center of how the industry measures model quality in the wild. We discuss the origins of Chatbot Arena, what an Arena score actually captures, how preference and factuality should (and shouldn’t) be combined, Agent Arena’s performance–cost frontier, AutoEval, and what evaluation looks like once harnesses and tools enter the ranking. TheSequence is a reader-supported publication.

Everyone is talking about Astra and Anthropic’s latest releases, but three other developments from last week demand your attention: Meta’s Muse Spark 1.3, World Labs’ Atlas, and Google’s Gemini 3.8 Flash. Last week delivered more than another leaderboard reshuffle.

A model release arrives with an irresistible claim: a 7-billion-parameter student retains 95 percent of the performance of a 70-billion-parameter teacher. This sounds like one of the best trades in computing. Ten times smaller. Nearly as intelligent. Put it on a laptop, inside an agent loop, or behind an API with much better margins. But where did the missing 5 percent go? Perhaps the student keeps the teacher’s mathematics score but loses its ability to know when it is confused.

Our series about model distillation continues. We have a surprising mega interview. We dive into the Astra, Fable and Muse Spark releases to keep you up to date. We will discuss the possible “ChatGPT moments” for robotics. TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. The AI industry has developed a peculiar new benchmark: can you finish reading a model’s system card before its replacement ships?

Imagine that an AI lab spends several billion dollars assembling chips, power, researchers, and data. It trains the best model in the world. The benchmarks move. Developers migrate. The launch becomes an industry event. For a moment, the company looks like a medieval castle with very thick walls. Then the strange thing happens. Six months later, another lab reaches roughly the same capability. An open model offers most of it at a fraction of the price.

The past month delivered three releases worth reading closely, not because they move the same benchmark but because each is a different answer to the same question: how do you build a model that can work on its own for hours, and how do you make that affordable? Anthropic shipped Claude Fable 5.1 and Mythos 5.1, one set of weights sold under two safeguard regimes. Zhipu shipped GLM-5.3-Flash, a 320B model that activates 18B parameters and spent a week serving anonymous traffic on Chinese chips.

In July 2026, PrismML released Bonsai 27B, a model that makes the old relationship between parameter count and hardware look slightly absurd. A conventional 27-billion-parameter model in 16-bit precision needs roughly 54 gigabytes just for its weights. Bonsai ships in a ternary version around 5.9 gigabytes and a binary version around 3.9 gigabytes. The latter is designed to fit inside the memory budget of a high-end phone.

More distillation coming to you. We break down GLM and Qwen new models. We discuss some ideas about moats in the era of AI. We will discuss some of the new platforms in AI for science. TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. For the past three years, we have watched AI through a microscope pointed at the model. Which system reasons better? Which benchmark moved? Which lab discovered the next scaling trick?

Every subfield of machine learning has a moment where it stops being a collection of papers and starts being a stack. NLP had it when Hugging Face Transformers turned “reimplement BERT from the appendix” into a single from_pretrained call. Image generation had it with Diffusers. Robotics is having that moment right now, and the stack is called LeRobot. Here is the strange thing about robot learning a few years ago: the models were mostly fine. ACT worked. Diffusion Policy worked.