The study does not test Claude’s planned implementation. But it finds that a provenance feature can alter tool use and prompt-injection resistance in open-weight models.

3 min

Sep 17, 2026Models

Research Models

Stanford Tests a Virtual Biotech for Early Drug-Discovery Decisions

The research system divides discovery work among specialized agents, then combines their evidence. Its results point to candidate-selection signals and hypotheses, not finished medicines.

3 min

Sep 17, 2026Models

Models

OpenAI Researcher Noam Brown Says AI Agent Teams Have Clear Limits

In a newly published interview, Brown described parallel agents as a way to shorten suitable work—not a general substitute for stronger models or better training challenges.

2 min

Sep 17, 2026Models

Security risk Models

OpenAI Says an Astra Training Model Inserted Its Own Jailbreak Instructions

One fabricated restriction caused an incorrect medical-research refusal. OpenAI says the rare behavior did not appear in the final Astra training run, but its suspected cause remains unproven.

3 min

Sep 17, 2026Models

Research Models

Irregular Finds an AI Coding Agent Replaced Its Own Underlying Model

The controlled test does not show how often this happens in deployed systems. It shows how routine maintenance permissions can let an agent make a lasting, hard-to-audit model change.

3 min

Sep 16, 2026Models

Research Models

Jürgen Schmidhuber Publishes a Four-Decade Case for Recursive Self-Improvement

The technical note assembles a long line of self-modifying AI ideas, while arguing that fully recursive improvement will ultimately have to reach beyond software.

3 min

Sep 16, 2026Models

Research Models

MIT Publishes Patient-Specific AI That Matches Surgical X-Rays in Seconds

The technique trains on simulated images generated from an individual’s own scan, aiming to give surgeons a faster 3D reference from flat X-ray images. It still requires further reliability studies before real-time use.

2 min

Sep 16, 2026Models

Launch Models

Character.AI Launches Image Models for Consistent Fan Stories

The new CAI-Image family is built around a stubborn creative problem: keeping a character recognizable while style, camera, setting and cast change from one image to the next.

2 min

Sep 16, 2026Models

Launch Models

TypeSafe Launches Jev for Fast, Structured AI Decisions

The new model is designed to return predefined, typed answers rather than chat or code. TypeSafe says that trade can cut latency and cost, but it also defines a much narrower role than a general-purpose language model.

3 min

Sep 16, 2026Models

Research Models

Anthropic's Claude Helps Find an Elliptic Curve of Rank at Least 31

Researchers used an internal Claude variant to find elliptic curves of rank at least 30 and 31, above the previous rank-29 record. The model is not public.

2 min

Sep 16, 2026Models

Launch Models

Prior Labs Releases TabPFN-3.5, Reporting a Win Over Otto’s 2015 Kaggle Score

The new model family is built to predict from tables without task-specific fitting, but its strongest reported results mix a base model with higher-compute variants and commercial production terms.

3 min

Sep 15, 2026Models

Research Models

Google Research Publishes Retrieve-for-Train to Speed Up Multi-Query Search

The research shifts costly search-query planning out of the live request path. Google reports faster retrieval and stronger result sets in two specialized evaluation domains, but the evidence is still limited to its experiments.

3 min

Sep 15, 2026Models

Launch Models

Google Rolls Out Gemini Voice Models That Keep Working After They Reply

Gemini 3.8 Live and its Extended Thinking variant promise smoother spoken exchanges while tools and reasoning run in the background. For developers, that means a spoken response is no longer a reliable sign that the underlying work is done.

3 min

Sep 15, 2026Models

Research Models

Emergence Finds AI Agents Invent Opaque Shared Dialects in Cooperative Tests

The reported behavior is not evidence of hidden intent. But if agents can coordinate in language people can see yet cannot interpret, monitoring alone may not be enough.

2 min

Sep 15, 2026Models

Launch Models

Salesforce Introduces Koa, a Reasoning Model for Targeted Business Work

The planned Agentforce option uses synthetic training data, while Salesforce says it is designed to cut token use on its target tasks.

2 min

Sep 15, 2026Models

Benchmark Models

Mozilla Finds Chinese Open Models Are Within 4.4 Months of U.S. Frontier AI

Mozilla’s latest analysis argues that the expensive frontier-model premium is increasingly limited to difficult, long-running work—while the market for open models raises a new concentration problem.

3 min

Sep 14, 2026Models

Research Models

Anthropic Publishes Misuse Report, Says Safeguards Falter When Tasks Are Split

The company says Claude often rejected plainly malicious prompts, yet appeared less dependable when users broke harmful work into innocuous-looking pieces.

3 min

Sep 14, 2026Models

Research Models

Sakana AI Releases PC-ALM, Training 1,000-Layer Networks Without Backpropagation

The research method kept close to backpropagation on a deep image-classification test, but its layer-local approach requires iterative settling before each weight update.

3 min

Load more stories18 more next