IBM opens a model for time-series forecasting
Granite PatchTST-FM-r2 can be tested on your own data. Alongside a single forecast, it provides a distribution of possible outcomes.
IBM Research / Hugging FaceRead story
NEUDYNE / SIGNAL
Science in context. Without language barriers.
Important developments in AI, human biology and emerging technology. We follow original sources and explain what the findings mean — including their limits, without the hype.
How we use sourcesArtificial intelligence
Science · technology · the future
When an AI test crosses the boundary: promised isolation is not enough
A safe assistant is defined not only by its answers, but also by what it is allowed to touch.
AnthropicRead story
Agents API: from a single answer to work that can continue
The interesting prospect is not another chatbot, but continuing an unfinished task without explaining everything again.
OpenAIRead story
GPT-Live-1 in the API: a voice assistant you can interrupt
A flowing conversation may be useful. It does not, by itself, mean the request was understood correctly.
OpenAIRead story
Images 2.5 brings more precise editing and a published system card
OpenAI has begun rolling out a new generation of its image model.
OpenAIRead story
OpenAI releases a proposed proof for the Navier–Stokes problem
The mathematical result awaits independent assessment; this is not a new model release.
OpenAIRead story
AlphaGenome Atlas opens predictions for billions of DNA changes
The research database helps choose variants to investigate, rather than diagnose disease.
Google DeepMindRead story
IRIS infers which signals shaped developing cells
The model links gene activity to signaling; evidence comes from cells and mouse embryos.
Nature MethodsRead story
NeoMME searches documents using both text and page images
Open models combine both inputs in one network. Performance evidence currently comes from their developers.
H company / arXivRead story
Gemini chooses where to look when analysing video
The new mode loads relevant parts of a recording according to the question. Savings need testing on the actual workload.
Google DeepMindRead story
Gemini Omni 1.1 adds control over continuity in generated video
Google expands its video creation tools. Higher resolution alone does not guarantee a faithful image.
Google DeepMindRead story
Claude searched for fixes to ten categories of model safety failures
A company experiment supports automating part of safety research, not a conclusion that AI safety has been solved.
AnthropicRead story
OpenAI measures research-agent work, but discovery speed does not follow directly
A new internal report separates the volume of automated work from scientific progress.
OpenAIRead story
Claude translated Fermat's proof into a form a computer can check
A system of agents produced the first complete Lean formalization of the famous theorem in eleven days. This is not a new mathematical proof, but an exceptionally large machine-checkable encoding of an existing route.
AnthropicRead story
GPT‑6 Astra adds capability—and critical cyber risk
OpenAI has begun a limited rollout of a new model for agentic work, coding and computer use. Its own evaluations show large gains, while the system card also reports lower reasoning monitorability and the company's highest cyber capability tier so far.
OpenAIRead story
WeatherNext 3 builds a global forecast every hour from live observations
Google's new model incorporates continuous satellite and station data and refines the spatial grid to as little as five kilometres. Higher resolution and live independent evaluation still do not turn a forecast into certainty or an official warning.
Google DeepMind and Google ResearchRead story
Claude Fable and Mythos 5.1 share a model but differ on risk access
Anthropic has released the generally available Fable 5.1 and the restricted Mythos 5.1, built on the same foundation but governed by different safeguards. Cheaper cache reads can reduce the cost of long-running agents, while base token prices are unchanged.
AnthropicRead story
Gemini 3.8 Flash targets agents while the Cyber variant stays restricted
Google has released a fast general model and a security variant for vetted defenders. Pricing remains introductory through year-end, but higher performance on difficult tasks may require more reasoning steps and tokens.
GoogleRead story
Open pathology models retain performance with far less compute
GigaPath‑Flash reduces the cost of analysing whole histology slides, while GigaTIME‑Flash predicts spatial protein maps. They are open research models, not tools approved for diagnosis or treatment decisions.
Microsoft ResearchRead story
GOLLuM teaches a language model to express uncertainty when choosing the next experiment
Combining language-model representations with a Gaussian process matched standard Bayesian optimisation with over 40% fewer trials in benchmarks. It was not an independently operated laboratory.
Nature Machine IntelligenceRead story
An AI map suggests human tissues do not age at the same pace
An analysis of 25,306 post-mortem samples from 970 donors found different periods of rapid structural change across tissues. It is a research map, not a biological-age test.
Nature AgingRead story
When experiments disagree, AI can show scientists where to look first
Bayesian machine learning linked discrepancies in historical measurements to details of experimental procedures. Experts then tested the suspects with simulations and new experiments, reducing the spread in californium-252 fission data by up to sixfold.
Nature CommunicationsRead story
The retina as an ageing window: AI estimated age within three years on average
A model trained on fundus images from 71,343 participants linked a retinal-age gap to disease and mortality. It is a population biomarker, however, not a personal diagnosis or proof that the retina causes ageing.
Nature CommunicationsRead story
A model found a stronger compact genome editor from only hundreds of measurements
EvoMax combined sparse experimental data, a protein language model and structural estimates. The resulting Fanzor editor worked better in human cells and was also tested against a humanized gene in mice.
Nature BiotechnologyRead story
ChemTSv3 changes how it searches molecular space on the fly
The open framework can switch among molecular strings, graphs and protein sequences during computer-aided design. It is a candidate-search tool, not evidence of an effective drug.
Communications ChemistryRead story
AI predicts whether a peptide will bind a protein — and where
VITAL combines sequence information with binding-interface geometry. Across benchmark datasets, it classified interactions, mapped contact sites and estimated binding strength in one framework.
Nature Machine IntelligenceRead story
AI agents speed up scientific software, but humans still own verification
Eight case studies report faster maintenance and modernisation of research tools. They are not an independent benchmark, and scientific validation remains the hard part.
OpenAIRead story
When a language model thinks longer, but not better
Research is trying to describe LLM overthinking structurally — locating repetition and breakdown rather than merely counting tokens.
Google DeepMindRead story
NEUDYNE / SCALEA
A closer look at Scalea
Council news, everyday life and nearby places to explore. Local context for everyone who calls the town home.
Explore Scalea ↗NEUDYNE / EDITORIAL
How we use sources
We publish original editorial summaries and commentary, not full translations of third-party articles. We always link to the original and credit the author or publisher. Full translations are published only under licence or where expressly permitted.
Transparent sourcing
Every article links to the original work. We distinguish experimental findings, author claims and our own outlook on practical applications.
One journal. Four languages.
Read in Czech, English, German or Italian. Discussions are shared across languages.