Artificial intelligence
Reporting and analysis on artificial intelligence
Journal topics
AI designs enzymes and proposes mathematical proofs
The October 11 roundup brings scientific results together with video, search and robotics models. We distinguish laboratory tests, calculations and work still awaiting verification.
Worlds with memory, new voices and AI models: September 27
How a generator preserves a scene, a model selects an action and a robot learns soccer. New design, speech and automation tools, with the conditions that qualify their demonstrations.
Gemini 4, persistent agents and the coded Marmont letter
Models work for longer, speech recognition shrinks to megabytes, and AI helps decipher an old letter. The new tools, reported results and demonstrated limitations in detail.
AI news of the week: Qwen, Jev, world models and local AI
One week brought systems for changing the camera in existing video, translating live speech, maintaining persistent virtual worlds, making fast decisions from predefined options and controlling different robots. We separate demonstrations from specifications, note what can be downloaded and identify claims that still come only from developers.
DeepSeek V4.1 Flash: reading long prompts with less memory
A long conversation needs computing power and room for information already processed. Why the model divides reading and writing between two blocks and sometimes recomputes data instead of storing it.
YuE2: a song starts with a score
A Beethoven melody becomes hard rock, a pop song turns to jazz, and a laptop generates a composition in just over two minutes. The demonstrations, results and setup behind those examples.
New AI models: DeepSeek, a DNA atlas and a problem in fluid motion
Nine billion predicted DNA changes, ten thousand mathematical agents, editing recorded speech and a robot taking out the rubbish. The demonstrations, claims and designs behind the new systems.
New AI models: software work, local DeepSeek, and interactive video
New language models, interactive worlds, image editors, weather prediction, and a fly nervous-system map: a detailed account of September releases and the results presented.
GPT-6 Astra at work: four hours for a game, seven minutes for a mistake
An Unreal Engine game, cursor drawing, a piano solo, and a reconstructed building: detailed trials combining successful outputs with visual mistakes.
GPT-6 Astra: a million-token context with a long-request surcharge
Astra accepts over a million tokens, but higher rates start above 272,000. Caching makes repeated reads cheaper while the request stays in the long-context price bracket.
Five and a half hours of barely audible speech. Why Whisper large-v2 beat newer models
Six recordings, dozens of trials, and a sentence that appeared after 86 seconds of silence. We sought a modern open model for local transcription, but an older Whisper delivered the best result.
GLM-5.3 moved closer to the frontier. It still does not lead in coding
Z.ai’s release table shows a large gain over GLM-5.2, but it does not prove outright leadership. We examine what was measured, where vendor evidence ends, and how to test the model on a real repository.
The model that filled the queue: why Kimi K3 matters beyond its records
Demand forced Moonshot AI to pause new subscriptions. The more consequential event came later: the company released the weights of a model that had moved close to the strongest closed systems.