Category intelligence

AI News Briefing — March 25, 2026

26 current items analyzed and ranked.

Executive synthesis

AI News Summary

OpenAI made headlines by announcing the shutdown of Sora, its video generation tool, just 15 months after launch—a rare strategic retreat amid fierce competition from ByteDance and Google. Meanwhile, Anthropic is locked in a federal court battle with the Pentagon after refusing to let Claude be used for autonomous weapons and mass surveillance, with a judge questioning the DoD's motivations.

Anthropic launched computer-use capabilities in Claude Code and Cowork, allowing AI agents to control users' desktops directly
Arm announced it is manufacturing its own AI chips, with Meta, OpenAI, and Cerebras as first customers
Elon Musk revealed a chip megaproject spanning Tesla, SpaceX, and xAI
Meta AI published two significant research efforts: Hyperagents for recursive self-improvement and LeWorldModel (LeWM) solving JEPA collapse in world models
TinyLoRA from Meta FAIR/Cornell/CMU achieved 91.8% on GSM8K with just 13 trainable parameters
Luma Labs released Uni-1, an autoregressive image model that reasons about intent before generating
AI-generated CSAM rose 14% in 2025, with a 260x increase in video content, per the Internet Watch Foundation

Key Themes

AI Safety & Government Policy · 6Agentic AI & Computer Use · 6AI Hardware & Infrastructure · 3Frontier AI Research · 4Major Company Strategy · 4

Primary evidence

Top Ranked Signals

News Ars Technica - All content Mar 24

OpenAI announces plans to shut down its Sora video generator

By Kyle Orland

88 score

AI Analysis

OpenAI announced it is shutting down Sora, its AI video generation tool, just 15 months after its high-profile launch. The abrupt move signals a strategic retreat as competitors like ByteDance's Seedance and Google's Veo have overtaken Sora in quality and adoption.

OpenAI is preparing to shut down Sora, the video generation app that drew widespread attention when it launched in late 2024. OpenAI announced the move in a social media post Tuesday just after a Wall Street Journal story broke the news. The company said it will have more to share soon on "timelines for the app and API and details on preserving your work." "To everyone who created with Sora, shared it, and built community around it: thank you," OpenAI wrote. "What you made with Sora mattered, an

Major Company StrategyVideo GenerationProduct Shutdowns

News AI (artificial intelligence) | The Guardian Mar 24

OpenAI shutters AI video generator Sora in abrupt announcement

By Dara Kerr

86 score

AI Analysis

The Guardian's coverage of OpenAI shuttering Sora emphasizes the abruptness—coming just six months after a splashy standalone app launch with social feed features. OpenAI said it would share details on preserving users' work.

Tech firm ‘says goodbye’ to Sora, made publicly available in 2024, just six months after its launch of a stand-alone appIn an abrupt announcement on Tuesday, OpenAI said it was “saying goodbye” to its AI video generator Sora. The move comes just six months after the company’s splashy launch of a stand-alone app where people could make and share hyper-realistic AI videos in a scrolling social feed.“To everyone who created with Sora, shared it, and built community around it: thank you,” the compan

Major Company StrategyVideo GenerationProduct Shutdowns

News AI (artificial intelligence) | The Guardian Mar 24

Anthropic and Pentagon face off in court over ban on company’s AI model

By Nick Robins-Early

85 score

AI Analysis

Anthropic is fighting the Pentagon in federal court after the Trump administration ordered all US agencies to stop using Claude, following Anthropic's refusal to allow its AI for domestic mass surveillance and fully autonomous lethal weapons. A judge questioned the DoD's motivations for labeling Anthropic a supply-chain risk.

After Anthropic refused to let its AI to be used in autonomous weapons systems, Trump ordered US agencies to quit using itSign up for the Breaking News US email to get newsletter alerts in your inboxAnthropic faced against the Department of Defense in a federal court on Tuesday afternoon, as the artificial intelligence company seeks a temporary pause on the government’s decision to bar the US military and any contractors from using its technology. The two sides have been locked in an escalating

AI SafetyGovernment PolicyDefense AIAnthropic

News Feed: Artificial Intelligence Latest Mar 24

Pentagon’s ‘Attempt to Cripple’ Anthropic Is Troubling, Judge Says

By Paresh Dave

83 score

AI Analysis

A federal judge expressed concern about the Pentagon's actions against Anthropic, questioning the DoD's motivations for designating the Claude developer a supply-chain risk. The case could set major precedent for AI companies' ability to set ethical boundaries.

During a hearing Tuesday, a district court judge questioned the Department of Defense’s motivations for labeling the Claude AI developer a supply-chain risk.

AI SafetyGovernment PolicyDefense AIAnthropic

News Ars Technica - All content Mar 24

Claude Code can now take over your computer to complete tasks

By Kyle Orland

82 score

AI Analysis

First spotted on Social yesterday, Anthropic announced Claude Code and Claude Cowork can now directly control users' computer desktops—pointing, clicking, scrolling, and navigating applications to complete tasks. The system prioritizes API connectors when available but falls back to screen-level interaction when needed.

Anthropic is joining the increasingly crowded field of companies with AI agents that can take direct control of your local computer desktop. The company has announced that Claude Code (and its more casual user-oriented Claude Cowork) can now "point, click, and navigate what’s on your screen" to "open files, use the browser, and run dev tools automatically" when necessary to complete tasks. When possible, Anthropic says Claude Code and Cowork will still prioritize using Connectors to directly acc

Agentic AIComputer UseAnthropicProduct Launch

News aibusiness Mar 24

Anthropic's Claude Can Now Take Control of Your Computer

By Graham Hope

80 score

AI Analysis

AI Business coverage of Anthropic's Claude computer control capability, noting it as part of a wave of personal agent applications across the industry.

The new capability is the latest release in a wave of personal agent applications.

Agentic AIComputer UseAnthropic

News Feed: Artificial Intelligence Latest Mar 24

Arm Is Now Making Its Own Chips

By Lauren Goode

78 score

AI Analysis

Arm, traditionally a chip design licensing firm, announced it is now manufacturing its own AI CPUs. Meta, OpenAI, Cerebras, and Cloudflare are among its first customers for the new hardware.

The chip design firm says Meta, OpenAI, Cerebras, and Cloudflare are among the first customers of its new artificial intelligence hardware.

AI HardwareSemiconductorsInfrastructure

News aibusiness Mar 24

Musk Reveals Chip Megaproject Spanning Tesla, SpaceX and XAI

By Scarlett Evans

75 score

AI Analysis

Elon Musk revealed a chip megaproject spanning Tesla, SpaceX, and xAI, reportedly in response to lagging production by existing chip manufacturers. The cross-company initiative signals a major vertical integration play in AI compute.

According to Musk, the project responds to lagging production by existing chip manufacturers.

AI HardwareInfrastructureElon Musk

News MarkTechPost Mar 24

Meta AI’s New Hyperagents Don’t Just Solve Tasks—They Rewrite the Rules of How They Learn

By Asif Razzaq

74 score

AI Analysis

As first reported in Research on Sunday, Meta AI researchers introduced Hyperagents, systems capable of recursive self-improvement that go beyond fixed meta-level mechanisms. Building on the Darwin Gödel Machine, Hyperagents can rewrite their own learning rules rather than just improving at tasks.

The dream of recursive self-improvement in AI—where a system doesn’t just get better at a task, but gets better at learning—has long been the ‘holy grail’ of the field. While theoretical models like the Gödel Machine have existed for decades, they remained largely impractical in real-world settings. That changed with the Darwin Gödel Machine (DGM), which proved that open-ended self-improvement was achievable in coding. However, DGM faced a significant hurdle: it relied on a fixed,

AI ResearchSelf-ImprovementMeta AIAgentic AI

News MarkTechPost Mar 24

Yann LeCun’s New LeWorldModel (LeWM) Research Targets JEPA Collapse in Pixel-Based Predictive World Modeling

By Asif Razzaq

73 score

AI Analysis

As first reported in Research on Sunday, Yann LeCun's team introduced LeWorldModel (LeWM), the first JEPA that trains stably end-to-end from raw pixels using only two losses, addressing the longstanding representation collapse problem. This advances LeCun's vision for world models as a path to more capable AI.

World Models (WMs) are a central framework for developing agents that reason and plan in a compact latent space. However, training these models directly from pixel data often leads to ‘representation collapse,’ where the model produces redundant embeddings to trivially satisfy prediction objectives. Current approaches attempt to prevent this by relying on complex heuristics: they utilize stop-gradient updates, exponential moving averages (EMA), and frozen pre-trained encoders. A team

AI ResearchWorld ModelsJEPAMeta AI

News MarkTechPost Mar 24

This AI Paper Introduces TinyLoRA, A 13-Parameter Fine-Tuning Method That Reaches 91.8 Percent GSM8K on Qwen2.5-7B

By Asif Razzaq

72 score

AI Analysis

Researchers from Meta FAIR, Cornell, and CMU introduced TinyLoRA, achieving 91.8% accuracy on GSM8K with just 13 trainable parameters (26 bytes) on Qwen2.5-7B. The method demonstrates LLMs can learn reasoning with extraordinarily few parameters through extreme weight sharing.

Researchers from FAIR at Meta, Cornell University, and Carnegie Mellon University have demonstrated that large language models (LLMs) can learn to reason using a remarkably small number of trained parameters. The research team introduces TinyLoRA, a parameterization that can scale down to a single trainable parameter under extreme sharing settings. Using this method on a Qwen2.5-7B-Instruct backbone, the research team achieved 91.8% accuracy on the GSM8K benchmark with only 13 parameters, totali

AI ResearchEfficient Fine-tuningLLM Reasoning

News MarkTechPost Mar 24

Luma Labs Launches Uni-1: The Autoregressive Transformer Model that Reasons through Intentions Before Generating Images

By Michal Sutter

70 score

AI Analysis

Luma Labs released Uni-1, a foundational image model using a decoder-only autoregressive transformer that reasons through user intentions before generating images. The architecture represents a shift from diffusion-based approaches to instruction-following generation.

In the field of generative AI media, the industry is transitioning from purely probabilistic pixel synthesis toward models capable of structural reasoning. Luma Labs has just released Uni-1, a foundational image model designed to address the ‘intent gap” inherent in standard diffusion pipelines. By implementing a reasoning phase prior to generation, Uni-1 shifts the workflow from prompt engineering’ to instruction following. The Architecture: Decoder-Only Autoregressive Tran

Image GenerationModel ReleaseArchitecture Innovation