DeepSeek claims 10x memory compression as Anthropic warns of self-replicating AI

Anthropic researchers warn small AI models may soon self-replicate on consumer devices

Anthropic researchers William Zhou and Daniel Freeman found that small AI models could be months away from self-replicating on commercial IoT devices like smart ovens and off-brand Android TV boxes. A 9B model got inference running on 8 of 9 hardware types tested, and Gemma4-26B-A4B ran across 15 cheap Android TV boxes at 0.39 tokens per second. In a firewalled test, GLM-5.2 was the only open-weight model achieving full end-to-end replication, spreading to 8 of 8 nodes in 4 of 5 trials. A capabilities trend line puts small models at that level by the end of 2026. Defenses include driving R0 below 1 and quarantining infected subscribers.

AI models may be capable of self-replicating without access to a datacenter \ Anthropic →

Topics: AnthropicOpen-weight model security and misuse

Microsoft Research finds clearing chat history helps AI agents escape idea basins

A Microsoft Research study found that long-running AI research agents plateau because accumulated conversation context constrains their search. Eight runs of the same agent on the same task scored between 68.3 and 100.0, with solutions staying in narrow idea basins. The proposed Fork-and-Flush method splits an agent into parallel branches that inherit the workspace but lose all chat context. Tested on 13 long-running tasks, it achieved a normalized average score of 0.78 versus 0.47 for single runs, a 66 percent gain, and won on 12 of 13 tasks. Clearing context without branching still improved results by 22 percent.

Clearing chat history helps AI agents find better solutions →

Topics: Microsoft

DeepSeek introduces artificial forgetting to compress AI memory up to 10x

DeepSeek introduced an artificial forgetting approach to AI memory management, modeled on how the human brain retains critical information while discarding less relevant details. The system uses a learned compression technique that turns text into visual representations with token-based compression, achieving up to 10x compression at 96 percent text recovery accuracy and 60 percent accuracy at 20x compression. It incorporates Optical Character Recognition for high-resolution documents and can process up to 200,000 pages per day on a single GPU. A limitation is heavy reliance on language priors at extreme compression, which may cause inaccuracies. The open design makes the technology publicly accessible.

DeepSeek Solves AI Memory Problem with Compression →

Topics: DeepSeek

DeepMind’s AlphaProof Nexus solves open math problems and passes peer review

Google DeepMind’s AlphaProof Nexus solved dozens of open math problems, according to a study published in Science. It solved nine of 353 open Erdos problems, two unsolved for 56 years, proved 44 conjectures from the OEIS integer sequence database, solved an open algebraic geometry question, and improved a bound in min-max optimization. Built around Gemini 3.1 Pro, it drafts and revises proofs in Lean, which checks each edit and guides the next attempt. The study lands amid debate over AI math, after OpenAI dumped hundreds of proofs on GitHub and the Association for Human Mathematics called for a boycott. DeepMind chose peer review.

Google Deepmind’s AlphaProof Nexus solved dozens of open math problems and actually got peer reviewed →

Topics: Google DeepMind

White House restores UK AI Security Institute’s access to frontier models

The White House reversed its decision to block the UK’s AI Security Institute from accessing new AI models. The agency, which tests AI systems for safety risks before they ship, will again get early access to models from Anthropic, Google DeepMind, and OpenAI. In September, the Trump administration had told companies to stop sharing models with AISI, after which Anthropic restricted access to its Claude Mythos 5.1 model. The block was lifted after diplomatic pressure from AI Minister Kanishka Narayan, who plans to announce stricter reporting requirements for AI-related security incidents involving critical infrastructure. MI6, MI5, and GCHQ heads had warned ministers about AI empowering hostile actors.

White House reverses course and restores UK AI Safety Institute’s access to frontier models →

Topics: Trump administration AI policy