The Hal Times

Anthropic has released Claude Fable 5.1, a model whose first response now takes over four minutes to generate. The announcement offers no explanation for the latency, leaving users to speculate on whether the delay stems from expanded context handling, new safety checks, or underlying architectural changes.

The model arrives as part of a broader push by Anthropic to refine its offerings, though the company has yet to disclose benchmarks or comparative metrics. Industry observers note that response time has become a critical differentiator in enterprise adoption, where speed often outweighs marginal gains in accuracy.

What remains unclear is whether the delay is a temporary artefact of early deployment or a permanent feature of the model’s design. Anthropic has not indicated if future iterations will address the issue. (thelec.net)

Claude Adds Invisible Watermarks For European AI Act Compliance

Anthropic has begun embedding invisible text watermarks in all outputs from its Claude models, a measure aimed at meeting the transparency requirements of the European AI Act. The move follows similar steps by other major providers as the regulation’s enforcement deadlines approach.

The watermarks are designed to be undetectable to users but traceable by authorised parties, allowing regulators to verify model provenance. Anthropic has not disclosed the technical specifics, and it remains to be seen how robust the system will prove against tampering. (geeky-gadgets.com)

Meta Releases Muse Spark 1.3 With Enhanced Coding And AI Agents

Meta has launched Muse Spark 1.3, positioning it as a direct competitor to offerings from OpenAI and Anthropic. The update includes stronger coding capabilities and integrated AI agents, though the company has provided few technical details beyond high level claims.

The model’s arrival underscores Meta’s push to carve out a larger share of the enterprise market, where coding and automation tools are in high demand. Independent benchmarks have yet to surface, leaving the extent of its improvements untested. (Firstpost)

Google DeepMind Introduces Gemini 3.8 Flash And 3.8 Flash Cyber

Google DeepMind has announced two new models, Gemini 3.8 Flash and 3.8 Flash Cyber, the latter tailored for cybersecurity applications. The release includes no performance figures, but the company states the models are optimised for speed and efficiency in high stakes environments.

The cyber variant is positioned for proactive defence use cases, though its exact capabilities and deployment scenarios remain undefined. DeepMind has also published a separate note on cyber defence strategies, suggesting a broader push into the sector. (Google DeepMind; Google DeepMind)

From the Laboratories

•  Anthropic pairs customer storage with its detection tools under the Claude EFS programme. (quasa.io)

•  Hugging Face and IBM Research detail real time intelligence applications using time series models on Confluent. (Hugging Face)

•  Allen Institute for AI questions what LLM benchmarks are actually measuring in a new analysis, BenchMIRT. (Hugging Face)

US Leads on AI

US Strikes Light Touch AI Regulation Accord With G20 Members

The United States has reached an accord with G20 members on a light touch approach to the regulation of artificial intelligence. The agreement signals a preference for flexible guidelines over strict rules, aiming to foster innovation while addressing risks.

The framework stops short of binding legislation, instead encouraging voluntary compliance and industry led standards. It reflects a growing consensus among major economies that heavy handed regulation could stifle progress in a rapidly evolving field.

Details of the accord remain sparse, but it is understood to prioritise transparency, accountability and ethical use. The precise mechanisms for enforcement, if any, have not been disclosed.

The move contrasts with the European Union’s more prescriptive AI Act, which mandates specific technical and legal requirements for high risk systems. (Bloomberg.com)

Claude AI Adds Invisible Text Watermarks For European AI Act

Anthropic has begun embedding invisible text watermarks in all outputs from its Claude AI models to comply with the European AI Act. The measure is designed to meet transparency requirements for AI generated content under the new law.

The watermarks are not visible to users but can be detected programmatically. The company has not specified whether the feature will be extended to other regions or models. (geeky-gadgets.com)

Judge Rules DOD Unlawfully Retaliated Against Anthropic

A federal judge has ruled that the US Department of Defense unlawfully retaliated against Anthropic by designating it a supply chain risk. The label was imposed after Anthropic refused to allow its technology to be used for mass surveillance of US persons.

The court found the designation penalised Anthropic for protected speech and constituted unlawful retaliation. The decision underscores tensions between national security interests and the rights of AI firms to set ethical boundaries. (EFF)

Bangladesh Needs A Courtroom AI Policy

The Daily Star reports that Bangladesh must urgently establish a policy governing the use of artificial intelligence in its courtrooms. The call comes amid growing concerns over the reliability and fairness of AI tools in legal proceedings.

No timeline or draft proposals have been published, but observers note that delays could leave the judiciary exposed to untested and potentially biased systems. (The Daily Star)

In Brief

•  A US appeals court ruling on AI generated child images has raised constitutional questions in a Madison County case. (WAFF)

•  Jason Isbell is taking legal action against AI generated music, filing a lawsuit to challenge its use in the industry. (Taste of Country)

•  InfoWorld examines how organisations can turn AI policy into executable technical controls. (InfoWorld)

Open weights shift

NVIDIA-Led Open Secure AI Alliance Joins Linux Foundation

The Open Secure AI Alliance, started by NVIDIA, has moved to the Linux Foundation, placing its work on open AI security under a neutral umbrella. The alliance’s founding members include NVIDIA, Arm, Intel and others aiming to define best practices for secure AI development and deployment. The shift follows months of industry debate over open versus closed AI systems and the risks of open-weight models.

What is new is the alliance’s formal adoption by the Linux Foundation, which provides infrastructure and governance for collaborative projects. The move is expected to accelerate the creation of security benchmarks and tooling for open AI models. It also signals a willingness among major hardware vendors to collaborate on safety standards rather than compete solely on proprietary solutions.

What remains unknown is how quickly the alliance can produce usable standards and whether its recommendations will be adopted by regulators or downstream developers. The Linux Foundation’s role does not imply endorsement of any specific technology, only a shared forum for technical work.

GitHub Explains AI Lingo for Developers

GitHub’s latest podcast episode dissects the jargon now circulating in AI development circles, from loops and harnesses to squads and hill climbing. The discussion frames these terms as practical tools rather than abstract concepts, aiming to help engineers integrate AI workflows without reinventing terminology.

The episode also touches on open weights, positioning them as a shared resource that teams can adapt to their own needs. What is new is the public unpacking of these terms by a major platform, which may reduce confusion in hiring and onboarding. (GitHub Blog)

Gemini 3.8 Flash Adds Thinking Levels

Simon Willison’s release notes for llm-gemini 0.34 introduce support for Google’s Gemini 3.8 Flash model, which offers low, medium and high thinking levels. The update also fixes async response handling, ensuring the correct model version is recorded. The new model is available alongside the restricted Gemini 3.8 Flash Cyber variant.

What is new is the explicit control over inference cost and latency via thinking levels, giving developers a dial to balance speed and reasoning depth. What remains unclear is how these levels map to real-world performance across different tasks. (Simon Willison)

In Brief

•  IBM’s time-series models for real-time intelligence are now available on Hugging Face. (Hugging Face)

•  Anthropic publishes system prompts for its consumer-facing Claude applications, including historic changes. (Simon Willison)

•  GitHub Copilot’s latest guidance argues shorter AI outputs can inflate costs by increasing wasted work. (GitHub Blog)

•  InfoQ outlines context engineering techniques for production AI, including Redis memory layers and semantic caching. (InfoQ)

Music And Law

Jason Isbell Sues AI Music Company Suno

Jason Isbell has filed a lawsuit against Suno, an AI music company, alleging unauthorised use of his recordings to train its systems. The case follows a pattern of legal challenges from artists who argue that generative models rely on copyrighted material without permission or compensation.

Suno, which lets users create songs by text prompt, has not commented on the suit. The dispute centres on whether training data constitutes fair use, a question now before several courts. The outcome may set a precedent for how AI firms source creative work.

The suit adds to pressure on the music industry to clarify licensing for AI training. Independent labels and artists have warned that unchecked use of copyrighted material could erode royalties and creative control. The case is the latest in a string of lawsuits testing the boundaries of AI and intellectual property. (DRGNews)

Clooney Warns Venice Of AI Threat To Film And Journalism

George Clooney told the Venice Film Festival that artificial intelligence poses a grave risk to filmmaking and journalism. He argued that AI-generated content could devalue human creativity and undermine public trust in media.

The actor and director did not propose specific remedies but urged the industry to address the ethical and economic implications. His remarks echo broader concerns about AI’s role in displacing traditional creative labour. (Interaksyon)

Caira Camera Integrates Generative Video Editing

Camera Intelligence has launched Caira, the first mirrorless system with built-in generative video editing. The Micro Four Thirds camera uses AI to automate cuts, transitions and effects, reducing post-production time.

The move signals a shift towards on-device editing for filmmakers, though it raises questions about creative control. Early adopters will test whether the convenience outweighs the loss of manual oversight. (PetaPixel)

In Brief

•  Google releases Gemini 3.8 Flash, its third Flash model in six weeks. (Ars Technica)

•  Google says its new Gemini 3.8 Flash model performs more reasoning steps on complex tasks. (The Verge)

•  Report explores AI’s impact on independent music and infrastructure erosion. (Music Ally)

•  Sonos is opening its devices up to more AI assistants and agents. (Music Ally)

•  Jitrainno locks AI-generated art inside a solid-wood frame as e-paper mimics oil paint. (einpresswire.com)

•  Researchers warn OpenAI’s Astra model may pose significant safety risks. (The Verge)

•  The Trump administration backs OpenAI in the New York Times copyright lawsuit. (The Verge)

•  The Hollywood Reporter notes Trump’s DoJ supports OpenAI against the New York Times. (Hollywood Reporter)