From the archive
AI research tests new methods for training, evaluating and understanding artificial intelligence. Find news on machine learning papers, reasoning, model behaviour and experiments that challenge existing assumptions. Read what researchers measured, how they tested it and where the limitations lie, with links to the original work and discussion of what the findings could mean beyond a benchmark.
Lead story
Scale Labs and Reflection have released SWE-Bench Pro V2, a tougher and more tightly controlled benchmark for AI software-engineering agents. It covers 642 tasks across 11 repositories, including a new 51-task Hard split, and is designed to make claims about coding ability harder to inflate with leaky tests,…
Read the report & discussion ↗Watch Desk
Radical Numerics is building large AI models for biology while warning that the same systems that expand scientific capability could also widen the biosecurity threat. The company’s co-founder and chief executive, Eric Nguyen, made the case in an interview with Latent Space, arguing that defensive capabilities must keep…
Watch Desk
Eye Security says researchers used a 12-step AI-agent method to uncover a CVSS 10.0 vulnerability chain in ASUS Control Center. The result is a useful security story for a less glamorous reason: the agents accelerated reconnaissance and code analysis, but human researchers still had to guide them through…
Berkeley BAIR Watch
A University of California, Berkeley and Princeton team has developed TANGO, an AI navigation system that lets a humanoid robot plan movement for all 29 of its joints. Tested on a Unitree G1, it reduced collisions in cluttered environments and completed narrow passages, low obstacles and floor obstacles…
Watch Desk
Five UC Berkeley computer science graduate students have been named Siebel Scholars for 2027, receiving $35,000 each to pursue research spanning AI, robotics, security and machine-learning hardware. The useful detail is not simply that Berkeley has won another academic laurel. The awards are buying time for work aimed…
Cohere Watch
Bell Cyber and Cohere say they have deployed a specialised AI model inside Bell’s security operations in Canada, where it will help analysts investigate threats while sensitive data stays on Canadian-hosted infrastructure. The important shift is practical: this is a production cybersecurity tool, not another model waiting patiently…
Watch Desk
Researchers have reported 104 public findings across 65 candidates in China’s Next-generation Commercial Cryptographic Algorithms Program, including practical breaks in five designs. The early results suggest that the programme’s new algorithms need considerably more scrutiny before anyone treats them as ready-made security foundations.
Watch Desk
Alexey Borsky has released mini-AGI, an open-source byte-level language model designed to train continually on an ordinary 8GB VRAM GPU. Its interesting promise is not another oversized chatbot, but a personal model that can keep learning from new data without forgetting everything it knew yesterday.
Watch Desk
Researchers and Chutes have released a public dataset containing more than 6.12 billion production LLM requests, offering an unusually detailed look at how real-world AI traffic behaves. The most useful finding is strikingly practical: 99% of repeat requests arrive within 15 minutes, meaning serving systems may not need…
AWS AI Watch
AWS has added skill-focused evaluation tools to Strands Evals and Amazon Bedrock AgentCore Evaluations. The important change is that teams can now inspect not only an agent’s final answer, but also whether it selected the correct procedure and followed that procedure properly.
Watch Desk
Nutanix has acquired French AI compute-orchestration company Ryax, aiming to improve GPU utilisation and help enterprise customers manage agentic AI workflows. The deal matters because it targets the unglamorous bottleneck between an impressive AI experiment and a system that can actually run reliably in a business.
Watch Desk
Rumman Chowdhury has launched the Independent AI Evaluation Foundation with $10m in philanthropic backing, aiming to build independent safety evaluations and shared infrastructure for investigating AI failures. The important shift is institutional: AI labs would no longer be the only people examining what went wrong inside their own…
Watch Desk
Help Scout has adapted its design process for AI by making its principles explicit, rebuilding its design system so language models can interpret it, and creating internal tools for prototyping and research. The practical result, according to designer Buzz Usborne, is faster execution and more production code without…
Watch Desk
Researchers at Tsinghua University have developed Cache-to-Cache, a method for AI models to exchange compressed internal representations directly rather than passing information through text. The reported result is faster model-to-model communication and accuracy gains in tests, pointing to a potentially important change in how multi-model systems are wired…
Watch Desk
Cisco Talos has released CAIRN, an open-source framework designed to identify and analyse malware and hacking tools that use AI chatbots. The practical point is that security researchers are getting a framework for tracking AI metadata and behavioural fingerprints as malicious software begins borrowing the same tools used…
Watch Desk
Anthropic co-founder and chief science officer Jared Kaplan has warned that letting frontier AI systems train themselves could trigger an “intelligence explosion”, with a potentially decisive moment arriving between 2027 and 2030. The point is not that catastrophe is underway, but that one of the people building these…
DeepSeek Watch
DeepSeek V4.1 Flash is dramatically cheaper than Qwen3.8 Max on published token prices, and OrcaRouter’s telemetry puts it far ahead on time to first token. The catch is that one model is built for quick, interactive traffic while the other may suit slower, heavier jobs, so the cheapest…
Watch Desk
The Netherlands’ new strategic-innovation agency NADI is launching a €40m challenge with Germany’s SPRIND to accelerate chip design using artificial intelligence. The important point is not another lab demo, but public money being aimed at shortening one of computing’s slowest and most expensive development cycles.
