The Day in AI

• Ricursive Intelligence released Ripple-65B Preview on 4 November, a 65B open-weight model that it says scores 80.52 on TerminalBench 2.1.
• Ricursive Intelligence says the model cost $32M in total research and development and needed about 70 times fewer RL rollouts than Reflection Beam.
• Ryan Williams of MIT is reportedly organizing the review of OpenAI's claimed refutation of the Exponential Time Hypothesis, first announced on 27 October.
• Microsoft Research has reportedly ended its AI Frontiers group.
• Shield AI is reportedly valued at $28.8B, up 127% since March, according to Benzinga.

Ricursive Intelligence put out Ripple-65B Preview on 4 November, and it's the lab's first public language model. It's a dense transformer pretrained from scratch on 12 trillion tokens, with context up to 1M tokens. The weights are under Apache 2.0, so commercial use is open. Ricursive calls its TerminalBench 2.1 score the top result among U.S. open-weight models. It puts Reflection Beam, a 501B model, just below at 80.1, and Meta's Muse Spark Open, an 816B model, at 76.4. These are the company's numbers, and no independent evaluator has reported yet.

The claimed efficiency matters more than the score. Ricursive estimates its total research and development cost at $32M, counting everything from the small ladder of test models through the final runs. It says it reached the score with 1.4 million reinforcement learning (RL) rollouts, about 70 times fewer than Reflection Beam used. A rollout is one attempt by the model at a task, which is then scored and used to update it.

The model is thin and deep, with a hidden width of 5120 and 80 layers. It uses gated attention, which lets each head scale down its own output, and value residuals, which carry the value vectors of early layers into later ones. Its global attention layers drop positional encodings entirely (NoPE). Ricursive hasn't said which of these choices accounts for how far the model gets with so little RL.

The claimed refutation of the Exponential Time Hypothesis is still under review. OpenAI said on 27 October that its models found sub-exponential algorithms for a hard class of logic puzzles. The hypothesis assumed these puzzles need time that grows exponentially with their size. According to The Information, Anthropic claimed its own refutation hours later.

Ryan Williams, a complexity theorist at MIT, is organizing the community review with Russell Impagliazzo and Ramamohan Paturi. OpenAI has reportedly been asked to publish the proofs and fund a conference. Williams said OpenAI checked and was reasonably confident no cryptographic systems are broken. The proofs haven't been checked independently. This follows the 6 November claim by Physical Superintelligence that its system, running an Anthropic model, proved the Aaronson-Ambainis conjecture.

• Microsoft Research has reportedly ended its AI Frontiers group, according to a post on the hunch.net blog.
• Anthropic is the subject of a METR report on automating AI development that is alarming some researchers, The Verge reports.
• Shield AI, the defense startup, is reportedly valued at $28.8B, up 127% since March, according to Benzinga.
• Dave Wehner, Meta's chief strategy officer, is leaving to become chief financial officer of the inference startup Fireworks, The Information reports.
• Alphabet reportedly acquired a stake in Qualcomm through its purchase of Modular, according to The Information.
• Vals AI reportedly found a free DeepSeek model within eight points of OpenAI's restricted cyber model on a network-hacking test.
• The New York Times reports that some Chinese companies approved to handle Nvidia's H200 chips have empty offices and ties to sanctioned firms.
• Censys found more than 800 publicly exposed copies of ARTEX, a Chinese open-source AI penetration testing tool linked to attacks since July, TechCrunch reports.