70-90% of Anthropic Code Now Written by Claude as Recursive Self-Improvement Becomes 'Present Phenomenon'

Some 70 to 90 percent of the code behind future Anthropic models is now written by Claude itself. Chief science officer Jared Kaplan believes fully automated AI research is less than a year away. Alignment lead Evan Hubinger: "Recursive self-improvement is not a future phenomenon. It is a present phenomenon."

Anthropic is shipping 1M-token context windows for Opus 4.6 and Sonnet 4.6, stretching what the model can process to book length.

Sam Altman is betting today's frontier models can discover the next architecture after transformers.

Percepta hard-coded a WebAssembly interpreter into transformer weights, executing arbitrary C code as tokens. WebAssembly is a format that lets code run in browsers, and embedding an interpreter directly in model weights means neural nets can now function as general-purpose computers.

DeepMind used AlphaEvolve to establish new lower bounds for five classical Ramsey numbers. Ramsey theory deals with finding order in large structures, and these problems are so hard that mathematician Erdos once joked about their difficulty.

Physical Superintelligence PBC launched Get Physics Done, an open-source system that scopes physics problems, runs derivations, and verifies results against nature's constraints.

Terry Tao launched a Mathematics Distillation Challenge to compress reasoning into compact cheatsheets that boost LLM performance.

An AMA survey found 81 percent of physicians now use AI, more than double the 2023 rate.

New US Senate guidelines permit aides to use Gemini, ChatGPT, and Copilot for official work.

Meta is reportedly planning layoffs of 20 percent or more to offset AI costs.

Anthropic is in talks with a consortium including Blackstone and Hellman & Friedman to form a joint venture selling Claude to portfolio companies.

Uber co-founder Travis Kalanick relaunched his ghost kitchen company as "Atoms," pivoting to "gainfully employed robots" for food, mining, and transport.

Elon Musk revealed his "Terafab Project" launches this week.

QatarEnergy has not restarted helium production at Ras Laffan nine days after Iranian drone strikes knocked 30 percent of global supply offline. This threatens chip fabs that depend on ultrapure helium for cooling and leak detection.

The AI-driven memory shortage is so severe that RAM kits now ship with one fake stick alongside one real one.

STMicroelectronics unveiled plans to deploy over 100 humanoid robots for repetitive tasks in its older chip fabs to avoid plant closures.

Phantom MK-1 humanoid soldiers brandishing firearms have been delivered to Ukraine for evaluation.

Airbus is preparing two uncrewed combat aircraft from Kratos for first flight, targeting a German Air Force drone wingman by 2029.

China approved the world's first commercial invasive brain-computer interface that restores hand movement, reaching market before Neuralink.

German researchers achieved the first functional recovery of an adult mouse hippocampus after cryopreservation by vitrification, showing memory can survive being frozen.

A Sydney entrepreneur used ChatGPT to design a personalized mRNA cancer vaccine for his rescue dog, shrinking the tumor considerably.

The White House's planned UAP disclosure is reportedly "massive" according to former defense official Christopher Mellon, including satellite imagery of craft that look like nothing humans have built.

On March 22, the State of the World Forum is planning to hold the first AI x NHI Convergence Summit, reflecting a growing sense that artificial and non-human intelligences may be converging.

Steven Spielberg says his film "Disclosure Day" stems from a "sneaking suspicion that we are not alone here on Earth right now."