The week in AI

xAI's announcement was the week's biggest infrastructure news. Anthropic put Claude Cowork, an agent for non-coding desktop work, into research preview on 12 January. On the same day DeepSeek published Engram, a paper that adds a lookup memory to large language models. Cursor described hundreds of coding agents building a web browser together.

On the talent side, Barret Zoph and Luke Metz left Thinking Machines Lab and returned to OpenAI on 14 January.

Colossus 2 is xAI's second training cluster in the Memphis area. The power figures come from Musk's own post as quoted by the press, so they are company-reported. They are 1 gigawatt now and 1.5 gigawatts planned for April 2026.

On 31 December 2025, xAI said it was buying a third building to bring the site to nearly 2 gigawatts. It also said Colossus 2 would hold more than 555,000 Nvidia GPUs, at a cost of about $18 billion.

Press reports put Colossus 1 and Colossus 2 together at more than one million H100-equivalents. No independent party has measured the power draw or the GPU count.

Cowork brings the agent loop from Claude Code into the Claude desktop app. It runs locally inside an isolated virtual machine and can read and edit the user's files. It can also call outside tools through the Model Context Protocol (MCP), the open standard for connecting models to tools and data.

According to the same release notes, Anthropic opened Cowork to Pro subscribers on 16 January. The atlas has only partly confirmed the record behind these dates.

On 13 January Anthropic also expanded Anthropic Labs, its incubator for experimental products. Mike Krieger moved into Labs to build alongside Ben Mann, and Ami Vora now leads Product.

A mixture-of-experts (MoE) model saves compute by sending each token through only a few of its expert subnetworks. Engram adds a second kind of sparsity. Common N-grams, short runs of tokens, are hashed into a large table of stored embeddings, and the model fetches the matching vector in constant time instead of rebuilding that knowledge layer by layer.

The authors argue this leaves the transformer's depth free for reasoning. The memory tables can sit in host memory instead of on the GPU. They report a U-shaped scaling law for splitting parameters between experts and memory, which means the best results come from a mix of both and not from putting everything into one.

The authors report these results against an MoE baseline with the same parameters and FLOPs. MMLU went up 3.4 points, BBH 5.0 and HumanEval 3.0. On a multi-query needle-in-a-haystack (NIAH) long-context test the score rose from 84.2 to 97.0. DeepSeek released the code on GitHub, and no independent evaluator has reproduced the results.

Cursor organised the agents in a hierarchy. Planner agents split the work, worker agents write the code, and judge agents check the results. This setup let hundreds of agents work on one codebase for days.

The write-up says a middle amount of structure worked better than a very loose or a very rigid setup. It also says prompts mattered more than the details of the agent harness. Of the models Cursor tried, GPT-5.2 did best on long, sustained runs. All of these figures and findings come from Cursor.

• Higgsfield extended its Series A by $80 million to $130 million at a $1.3 billion valuation on 15 January, according to TechCrunch, with Accel, Menlo, GFT and AI Capital Partners investing and a reported $200 million annual revenue run rate.
• OpenAI launched ChatGPT Go worldwide at $8 a month on 16 January and said it will test ads in the free and Go tiers in the US.
• StepFun released Step3-VL-10B on 13 January, an open 10B vision-language model that StepFun says rivals models 10 to 20 times its size, such as GLM-4.6V and Qwen3-VL-Thinking.
• Meituan updated its 560B thinking model as LongCat-Flash-Thinking-2601 on 14 January, adding reinforcement learning across many environments, tolerance for noisy tools and a parallel Heavy Thinking mode.
• ByteDance Seed released Stable-DiffCoder on 15 January, an open 8B diffusion code model trained by block-diffusion continued pretraining on the Seed-Coder pipeline, which ByteDance says beats comparable autoregressive and diffusion code models.

• Barret Zoph, co-founder and chief technology officer of Thinking Machines Lab, returned to OpenAI on 14 January with Luke Metz and Sam Schoenholz, and Fidji Simo of OpenAI said the move had been in the works for several weeks.
• Luke Metz, a co-founder of Thinking Machines Lab, returned to OpenAI on 14 January, less than a year after the lab launched in February 2025, and Soumith Chintala became the lab's chief technology officer.
• Ahmad Al-Dahle, Meta's former head of generative AI and later co-head of AI products, left Meta on 14 January to become chief technology officer of Airbnb.