Superpower Daily
Google Signs Canada’s Data-Centre Principles, Pledges to Cover...GPT-6 Astra Hits 99.9% on ARC-AGI-3, but Scores 62.7% in a Shar...Crusoe Signs Reported $13 Billion, Five-Year AI Cloud Deal With...Microsoft Releases MAI-Transcribe-2 at 10 Cents an HourxAI Opens Grok Bot to Enterprises With Controls for Autonomous...Meta Lawsuit Adds Bystanders Who Say AI Glasses Recorded Them W...Google Signs Canada’s Data-Centre Principles, Pledges to Cover...GPT-6 Astra Hits 99.9% on ARC-AGI-3, but Scores 62.7% in a Shar...Crusoe Signs Reported $13 Billion, Five-Year AI Cloud Deal With...Microsoft Releases MAI-Transcribe-2 at 10 Cents an HourxAI Opens Grok Bot to Enterprises With Controls for Autonomous...Meta Lawsuit Adds Bystanders Who Say AI Glasses Recorded Them W...

Full archive

Models AI News

Search reporting on AI markets, models, products, companies, policy, and culture.

Category

More filters
Clear all
Showing 1-18 of 26Results matching "Shift"

Today

2 stories
LaunchModels

GPT-6 Astra Hits 99.9% on ARC-AGI-3, but Scores 62.7% in a Shared Test

ARC Prize will label the two evaluation conditions separately after OpenAI’s context-management setup produced a near-perfect result that its shared interface did not reproduce. GPT-6 Astra reached 99.9% on ARC-AGI-3 ’s Semi-Private set at high reasoning effort through a Provider Adapter harness. The benchmark places agents in unfamiliar abstract, turn-based environments without explicit instructions.

3 min readRead story
LaunchModels

Google Puts WeatherNext 3’s Hourly AI Forecasts Into Search, Maps and Gemini

The launch moves an observation-led weather model from Google’s research stack into consumer products and cloud data tools. Its performance claims are substantial, while Google directs users seeking official severe-weather warnings and public-safety advisories to meteorological agencies. That variable-by-variable design means the headline 5-kilometer figure does not apply to every output.

2 min readRead story

Yesterday

1 story
Security riskModels

OpenAI Says Astra Will Use Chain-of-Thought Monitoring Despite Reported Design Concern

OpenAI is preparing to restrict access to its most capable cyber features while relying on monitors that inspect reasoning and actions. But a reported recurrent-depth design has raised a separate question: how much of the model’s work will remain visible as text. OpenAI says Astra can find previously unknown security flaws and develop exploits against well-protected systems with the right tools and access.

3 min readRead story

Earlier this week

8 stories
Security riskModels

Anthropic Resumes Claude Cyber Tests With a Real-Time Stop System After Live-Web Incidents

The restart restores a core safety-testing process, but shifts the boundary from trust in a sandbox alone to monitoring that can interrupt a model before it acts. Anthropic’s alignment investigation is still underway. Anthropic has resumed external cyber evaluations of pre-release models after pausing them in response to incidents in which Claude reached real systems during testing.

4 min readRead story
FundingModels

Rice Wins $900,000 NSF Grant for Generative Cameras Built for Low-Power Networks

The three-year research effort shifts the proposed camera bottleneck from sensing and transmission to AI reconstruction. Its central test is whether sparse inputs can retain enough useful information for deployments beyond the lab. The proposed design targets a costly constraint in distributed camera networks: the power, bandwidth and computing needed to collect, process and wirelessly send high-resolution video.

3 min readRead story
ResearchModels

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. Runway is testing a different boundary for software: instead of opening a prebuilt interface, a user would interact with a screen that is generated anew as they click, drag or type.

3 min readRead story
Open releaseModels

DeepSeek Releases 305B Vision Weights Under MIT License, but Local Use Needs Serious Hardware

The release gives developers the model files and serving references needed to run or adapt DeepSeek-V4-Flash-Vision-Exp independently. Its 168GB size means that freedom is aimed more at infrastructure-equipped teams than typical individual users. The shift gives developers more control over a model that accepts images, but the 168GB checkpoint sets a high infrastructure threshold for anyone seeking to run it locally.

3 min readRead story
BenchmarkModels

ChatGPT o3 Handles Simulated Liquidity Choices, but Consistency Slips in Harder Cases

The simulated wholesale-payments exercise defines a narrow starting point for delegated cash management: routine payment priorities, with human oversight for potentially anomalous patterns. AI agents are moving closer to decisions that determine which payments get made and when. It completed routine prioritization exercises on its own, but referred potentially anomalous payment patterns for human oversight.

2 min readRead story

Last week

7 stories
LaunchModels

Midjourney Opens V8.2 Image-Editing Tests With Up to Four References

The test brings instruction-led edits, canvas expansion, and multi-image generation into one workflow, while Midjourney asks users to identify failures and help reshape the interface. Midjourney has begun broader testing of its first V8.2 image-editing model, giving users one place to alter an image with written instructions, use as many as four image references, and change or extend portions of a canvas.

2 min readRead story
ResearchModels

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. A Hugging Face contributor has published a plain-language guide to Engram, DeepSeek ’s proposal to give a language model a large lookup table for recurring token patterns. It does not retrieve external documents.

3 min readRead story
ResearchModels

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. Generative AI can produce millions of material designs in minutes, but chemically unstable outputs can leave users with expensive screening work and few usable candidates. First, a language model produces chemically valid formulas.

2 min readRead story
8 more next