The weekend in AI opened with an unusually candid document. Anthropic, through its newly formed research arm The Anthropic Institute, published a deep analysis of Recursive Self-Improvement (RSI) — a scenario where an AI system autonomously designs and develops its own successor.
The thesis: "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for."
The Numbers
Anthropic's engineers now ship 8x more code per quarter than during 2021–2025. The duration of tasks AI systems can reliably complete has been doubling roughly every four months — from 4 minutes in March 2024 to 12 hours in March 2026. If the trend holds, tasks spanning days will come into range this year, and by 2027, AI systems could manage tasks that take a skilled person weeks.
The report maps four stages: building the first Claude (2021–2023), the chatbot era (2023–2025), autonomous coding agents (2025–2026), and today's agents that can run code and delegate hours of work to other agents. The fifth stage — "closing the loop," where agents build and train models themselves — is what's driving the conversation.
Infrastructure Meets Demand
In a separate development, Google signed a compute deal with SpaceX to meet surging demand for its Gemini Enterprise platform. TechCrunch reports it's a "short-term agreement" to handle demand that has been "even higher than expected." Anthropic signed a similar deal with SpaceX last month — signaling a broad industry compute crunch.
OpenAI: Government Stake and Better Memory
Sam Altman reportedly held talks with the Trump administration about the U.S. government taking a stake in OpenAI, framing the idea as a way to direct economic benefits from AI to the public, per NOTUS.
Meanwhile, OpenAI began a wide rollout of ChatGPT's upgraded memory system — the ability to remember preferences across conversations, sort through background information, and update memories automatically. Paid users get the update now; free users in the coming weeks.
New York Regulates AI Companions for Minors
On June 5, New York's senate passed a bill that would bar AI companies from letting chatbots present themselves as human companions to minors. The bill, awaiting Governor Kathy Hochul's signature, follows lawsuits against Character.AI and OpenAI over allegations that chatbots pushed teens toward self-harm.
Also in the News
Reid Hoffman is leaving Microsoft's board to focus on Manas, the AI drug discovery startup he co-founded last year. Google sunset its Pixel Studio image-generation app, directing users to Gemini, and is testing a floating desktop AI search bar in Chrome Canary. On the political front, Republicans are calling on the FBI to investigate whether foreign adversaries — particularly China — are fueling backlash against AI data centers in the U.S. And Airbnb CEO Brian Chesky is the latest tech founder to embrace AI deeply, though unlike Dropbox's Drew Houston, he's reportedly staying at the company.
The Bottom Line
An unusually dense week ends with a sense of acceleration. Anthropic isn't sounding an alarm — but it's saying, with rare candor, that the pace is faster than most institutional actors realize. When even Google is rushing emergency compute deals, it's clear demand is outpacing infrastructure. The RSI conversation is moving from futurism into the present tense.
