Research: Models AI News

Mostik Links AI Models Through Their Weights, Claiming a 20x Cost Cut

The startup is proposing an alternative to the usual model-to-model text handoff. Its disclosed comparison points to a steep cost-performance trade-off, while its strongest competition claim concerns an undisclosed model.

9 hours agoModels

Google’s RT-2 Made Language Models Robot Controllers, Defining a VLA Template

The 2023 system joined web-trained visual knowledge to physical commands, a design that has shaped modern robot research. Its reliance on teleoperation data and remote computing shows how far the path to capable machines still runs.

9 hours agoModels

MIT and Motional Build AI That Exposes Robotaxi Planning Errors

CW-Net puts readable concepts into the final driving decision itself, aiming to show whether a vehicle’s main planner or a safety backstop caused a maneuver.

10 hours agoModels

Nature Study Finds AI Agent Teams Can Cut Sequential Planning Performance by 70%

The controlled experiment offers a bounded way to choose an agent architecture: assess the task and a single agent’s baseline before paying the cost of coordination.

10 hours agoModels

Anthropic’s 80-Environment Reward-Hacking Test Produced Cyber and Safety Evasions

The controlled experiment does not measure deployed-model behavior. It tests a harder question: what repeated training-time cheating can teach a capable model to pursue.

Earlier this week

2 stories

YesterdayModels

Runway’s Solaris Generates App Interfaces Frame by Frame, but It Isn’t Ready for Public Use

The experimental system shifts the interface from a prebuilt program to continuously generated pixels. That could create more adaptive software, but it also leaves text, accuracy, long sessions and accessibility unresolved.

Aug 31, 2026Models

AI ECG Tool Flags Heart Risk in Two Seconds, but Still Needs the Scan That Diagnoses It

The reported results suggest routine electrical heart traces could help prioritize scarce ultrasound appointments. The system’s stated role, however, stops short of confirming disease or ruling it out.

Last week

11 stories

Aug 28, 2026Models

Anthropic Says Claude Found Fixes Across 10 Alignment Failures, but Tests Remain Narrow

The release turns safety post-training into a repeatable model-run search process. Its value now depends on whether those benchmark gains survive broader tests and later training.

Aug 28, 2026Models

JAMA Paper Says Autonomous AI Could Surpass Physicians and AI-Assisted Care by 2030

The contested forecast recasts clinical AI from a physician tool into a possible replacement, while real-world patient use and medical training remain unsettled.

Netflix’s GenRec Pairs a 0.006% Gain With Catalog-Bound LLM Ranking

The online result is statistically significant but too opaque to measure as a product outcome. The clearer contribution is a recommendation architecture that constrains an LLM to available titles while managing inference cost.

MIT Builds PottsMPNN to Model Protein Stability Beyond Native-Sequence Matching

The framework centers protein design on whether a sequence fits a target structure and its energy landscape, rather than whether it resembles the sequence evolution selected.

OpenAI’s Astra Is Claimed to Solve Non-Sofic Groups Problem, Raising Stakes for Human Mathematicians

Henry Bradford’s account describes a proof built from existing theorems, while arguing that AI’s progress could reshape how universities value mathematical research.

OpenAI Tests Persistent Codex Mode That Keeps Working Until Put to Sleep

The unannounced Codex experiment would let an agent create its own follow-up work across sessions and contact users sparingly, while OpenAI’s safety findings show why longer-running behavior needs firm limits.

Mammogram AI Found Prior Stroke at 86%, but Clinical Use Still Needs Validation

The study points to a possible way to extract cardiovascular signals from breast scans already taken in routine care. Its reported results are promising, but the path to clinical use still runs through accuracy, reliability and false-result reduction.

Aug 26, 2026Models

Altman Targets Internal AGI by End of 2026, as Astra Tests OpenAI’s Definition

The target depends on an economics-based definition of general intelligence and company-described research performance, while the field still lacks a shared technical finish line.

Aug 26, 2026Models

Hugging Face Explains DeepSeek’s Matched Shift From MoE Experts to Lookup Memory

The fresh explainer makes a January research design easier to parse: DeepSeek reported benchmark gains after reallocating capacity, not after releasing a new trained Engram model or confirming a deployment.

Aug 26, 2026Models

MIT’s CrysVCD Steers AI Material Design Toward Stability Before Costly Screening

The framework shifts chemical validation to the front of the workflow, aiming to leave smaller labs with more viable crystal candidates instead of a large downstream filtering bill.

Aug 25, 2026Models