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
Hugging Face has published a detailed preview of tokenizers v1, a major rewrite aimed at stopping text preparation from leaving increasingly fast AI systems waiting around. The project reports large gains in decoding throughput, 76% scaling across eight native threads on an Apple M4 Max, and identical token…
Read the report & discussion ↗Watch Desk
A month-old UK AI start-up founded by former Google DeepMind researchers is reportedly seeking up to $700m at a valuation of about $3.7bn. Emulate is still operating in stealth, making this a striking vote of investor confidence in world models before the company has shown the public what…
DeepSeek Watch
DeepSeek is planning to use more domestic chips to train its AI models, with Huawei expected to begin supplying training processors as early as the fourth quarter, according to The Information. The move would deepen the partnership between two Chinese technology companies trying to reduce their exposure to…
Anthropic Watch
Anthropic chief executive Dario Amodei’s call to “pace the frontier” is being challenged on strategic as well as safety grounds. In a new Stratechery analysis, Ben Thompson argues that slowing model progress could ease the infrastructure, product and competitive pressures now building around frontier AI.
Watch Desk
An independent developer has released two open-weight versions of Boris-2 after uncovering training and data-pipeline mistakes that held the model back. The useful lesson is not simply that another model is now available: the project shows how much performance can depend on unglamorous engineering details, and how sharing…
Hugging Face Watch
Hugging Face has published a practical guide to reading model cards before downloading or deploying an AI model. Its useful argument is simple: a model ID is not a recommendation, and the README may be the difference between a sensible local experiment and an expensive, legally awkward surprise.
Meta AI Watch
Meta has rejected a Polish report about scam advertising, arguing that its methodology overstates the scale of the problem while outlining new AI-powered enforcement and reporting tools. The dispute matters because both the criticism and Meta’s defence turn on numbers that ordinary users cannot easily inspect.
Watch Desk
HiDream.ai has moved from image and video generation into interactive world models, while building an enterprise software strategy around the technology. The Chinese start-up’s progress matters because it is attempting to turn capable models into repeatable commercial workflows, rather than leaving customers to admire the demo and assemble…
Anthropic Watch
Anthropic says Claude has optimised more than 30 open-source biology models in under four weeks, with average speed-ups of roughly four times and a new low-memory mode for much larger molecular systems. The practical significance is clear: some protein-structure and design work could require far less computing, although…
Watch Desk
A new Hugging Face guide argues that AI agents should be judged on how they recover from failure, not merely whether they finish a task. The useful shift is from “did it work?” to “what happened after it went wrong, and what did recovery cost?”
AWS AI Watch
AWS has published a detailed account of how Wood Mackenzie built a shared platform for its internal and customer-facing AI agents. The important bit is architectural: one managed foundation handles identity, scaling, memory, observability and guardrails, so each team does not have to rebuild the same machinery before…
Watch Desk
ScalablyAI tried to improve a production AI agent by fine-tuning it on its own recorded work. The US$53.88 experiment produced a real gain on recovery tasks, but also degraded clean decisions, so both adapters were rejected and nothing was deployed. That is a much more useful self-improvement story…
DeepSeek Watch
A DeepSeek engineer says AI-written GPU kernels could match or exceed his own work within six to 12 months, while a separate NVIDIA experiment shows a narrower version of that workflow already producing numerically correct code. The important shift is from AI assisting with low-level optimisation to agents…
Zhipu AI / Z.ai GLM Watch
Z.ai has lifted its year-end annual recurring revenue target by 25% to US$3bn after a US$5bn cash injection eased its computing constraints, according to the South China Morning Post. The important shift is not merely a bigger forecast. Z.ai says it now has enough capacity, at least for…
DeepSeek Watch
DeepSeek’s V4.1 Flash is winning developers by handling large numbers of routine agent tasks quickly and cheaply, according to 36Kr. The important shift is not simply that a smaller model can rival a flagship on some tests. It is that AI use is becoming frequent enough for speed,…
