The week in AI

AMI Labs was the largest money event of the week. It also reportedly took a founding team drawn mostly from Meta's AI research organisation. On 11 March Replit raised $400 million at a $9 billion valuation and released Replit Agent 4.

Anthropic shipped two developer changes. On 9 March it launched Code Review, which sends several agents to review each pull request. On 13 March it made the 1M-token context window generally available at standard pricing.

A world model is a system trained to predict how its environment will change, so an agent can plan by simulating the results of its actions before taking them. AMI says building these models is its whole purpose. The round is a seed round and the company has not described a product.

Yann LeCun left Meta to found AMI and is executive chairman. He remains a professor at NYU. Day-to-day leadership belongs to Alexandre LeBrun, the new CEO, who founded and ran the French medical-AI startup Nabla.

The rest of the leadership is reported but only partly confirmed. Pascale Fung, a senior director of AI research at Meta, is reported to have become chief research and innovation officer. Michael Rabbat, a director of research science at Meta, reportedly joined as VP of world models. Saining Xie, a research scientist at Google DeepMind, reportedly became chief science officer. Sifted reported in January 2026 that the first hires came from Meta and DeepMind.

Georgian led the round. Other investors include a16z, Coatue, Y Combinator, 1789 Capital and the Qatar Investment Authority. Replit reports more than 50 million users and says employees at 85% of the Fortune 500 use it. Its stated target is $1 billion in annual run-rate revenue by the end of 2026.

Replit Agent 4 came out the same day. Agent 3 was one long-running agent working through a task. Agent 4 adds a design canvas where the user designs alongside the agent, runs several agents in parallel, produces output in more than one format and lets teams work together. Replit says Agent 4 is 10 times faster than Agent 3, and that figure is Replit's own. Agent 4 runs only inside the Replit app.

Anthropic says code output per engineer at the company grew 200% in a year and human review couldn't keep up. Code Review assigns several agents to each pull request and spends more time per review to find bugs a quick read would miss. Anthropic says it already runs the system on nearly every internal pull request.

On 13 March the 1M-token context window came out of beta for Claude Opus 4.6 and Claude Sonnet 4.6. Requests over the old threshold used to cost a long-context premium and needed a beta header. Both are gone, so a 1M-token request now costs the standard per-token rate. Rate limits are the same at every context length, and a single request can include up to 600 images or PDF pages.

• Gemini Embedding 2 was released by Google DeepMind on 10 March and maps text, images, video, audio and PDFs into one shared embedding space, which Google says makes it the company's first multimodal embedding model.

• Yann LeCun left Meta to found AMI Labs on 10 March and is its executive chairman while staying a professor at NYU.
• Alexandre LeBrun left Nabla, the medical-AI startup he founded, to become CEO of AMI Labs on 10 March.
• Pascale Fung reportedly left Meta for AMI Labs on 10 March as chief research and innovation officer.
• Michael Rabbat reportedly left Meta for AMI Labs on 10 March as VP of world models.
• Saining Xie reportedly left Google DeepMind for AMI Labs on 10 March as chief science officer.
• Guodong Zhang, an xAI co-founder who ran Grok Imagine and Grok Code, left xAI on 12 March, and the Financial Times reported that Elon Musk had already removed him from his main duties, blaming him for problems with the coding product.