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
OpenAI used DevDay 2026 to expand its developer tools for agents, including computer use in the Agents API and cloud-based Codex environments. The practical shift is towards agents that can operate software and run coding tasks remotely, not just answer a prompt and wait politely for the next…
Read the report & discussion ↗Baseten Watch
Baseten says an AI-built inference engine for Qwen-3.6-35B-A3B ran up to 90% faster than vLLM on single-stream decoding in its tests on one NVIDIA B200. The company’s experiments also included a separate image-segmentation model, suggesting the approach may reach beyond one language model, though neither system is serving…
Meta AI Watch
Meta says mathematicians working with its Muse Spark models have produced six papers, five of them answering previously open research questions. The work ranges from a sharp threshold for fitting random points to an ellipsoid to a counterexample that overturns a conjecture about groups.
vLLM Watch
Engineers have integrated Helion into vLLM’s linear backend, reporting higher LLM inference performance with less kernel implementation complexity. On NVIDIA Hopper GPUs, the tuned backend outperformed the default CUTLASS and DeepGEMM options in the workloads described, with throughput gains above 10% in some cases.
Watch Desk
Canada has named a 13-member National Council on Artificial Intelligence to advise the federal government on AI strategy, safety and the technology’s role in public services. The appointments put prominent researchers and investors around the same table, but raise a practical question: will this council do more than…
Epoch AI Watch
AI chips shipped through 2027 could eventually support tens to hundreds of millions of concurrent agents running frontier models, according to a new Epoch AI analysis. Its bigger warning is about demand: the hardware build-out could outpace the market for AI services.
OpenClaw Watch
OpenClaw says every skill and plugin uploaded to ClawHub now gets an additional security review from Tencent’s AI-Infra-Guard, alongside NVIDIA’s SkillSpector. The change puts automated checks into the marketplace’s upload review, where a risky plugin can otherwise become someone’s very avoidable afternoon.
OpenAI Watch
OpenAI has launched GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens, with cached input priced at $0.10. That makes the cost of using a frontier model, especially for repeated or agentic workloads, the story to watch.
Watch Desk
An AI system called Ataraxos beat top-ranked Stratego player Pim Niemeijer 15 games to one, with four draws, in a 20-game series. The result, reported by Ars Technica, is notable not just for the win: the researchers say they trained the system for less than $8,000 using 16…
AWS AI Watch
AWS says fine-tuning a Qwen3.6-27B search agent with multi-turn reinforcement learning improved its results on three of four held-out benchmarks, while sharply reducing failed tasks on BrowseComp-Plus. The useful detail is that the training rewarded the whole search journey, not just one answer at a time.
Watch Desk
Researchers say a code-execution approach helped robot agents complete more simulated tasks while using fewer language-model resources. In 700 task instances, PyRUA-Lean raised reported success from 63.1% to 71.7% under equal LLM-call budgets, a result that could matter for building more efficient robot agents.
Watch Desk
Context Language Models treat a model’s working context as an editable file, which it can change as a task unfolds. The researchers say their approach outperformed existing context-management strategies across several tasks while using fewer floating-point operations.
Cohere Watch
Cohere has released Embed 5, a pair of embedding models for finding information in enterprise documents, images and other data. Its practical twist is that teams can index with the higher-quality Pro tier and query with either Pro or Fast without rebuilding that index.
Allen Institute for AI Watch
Ai2 has open-sourced AstaBrief, an 8-billion-parameter model for turning a research question and retrieved literature into a cited report. Ai2 says its Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for the Claude-powered Thinking mode it tested.
Zhipu AI / Z.ai GLM Watch
Z.ai has introduced GLM-5.2 as an open-weights model aimed at demanding AI tasks, with a new “High” thinking level and a separate coding harness called Z Code. The report describes a package pitched at both capable model users and teams wanting to run or adapt models themselves.
Watch Desk
Researchers inside major AI companies are helping shape policy and challenge executive decisions, according to Axios. Its account points to a practical shift: the people building frontier AI systems may have more sway over what their companies do, and say, than their job titles suggest.
Watch Desk
Cloudflare has released two open-weight decision models, Clef and Clef-flash, which handle structured choices and can also process images and video, according to a report carried by news.lavx.hu citing The Register. The launch gives developers another option beyond hosted decision APIs, although running the models locally comes with…
Watch Desk
arXiv says it will limit each submitter to two preprint submissions a month, as the repository faces a record volume of papers. The change puts a practical limit on a major route for sharing research, while raising a hard question about how scholarly publishing should adapt to AI-assisted…
