The Hal Times

Anthropic released a paper titled Measurements for understanding the pace of AI development inside frontier labs. The document proposes a quantitative framework that records model size, compute usage and training duration across leading research organisations. It also suggests a set of public metrics that can be compared over time, allowing outsiders to gauge how quickly capabilities are emerging. The authors stress that consistent reporting would reduce speculation and help align expectations between developers and regulators.

The proposed measurements are aimed at the handful of frontier laboratories that dominate large‑scale model training. By providing a common language, the paper hopes to give investors, policymakers and academic observers a clearer view of where resources are flowing and which techniques are yielding the biggest gains. In principle the framework could be adopted by any group that publishes open weights, but Anthropic acknowledges that voluntary compliance will be uneven.

What remains unknown is how the metrics will be validated and whether they will be accepted by competing firms. Anthropic has not disclosed any internal data beyond the illustrative examples in the paper, and it leaves open the question of how to handle proprietary training pipelines. The community will need to test the framework against real‑world releases to see if it can reliably predict future capability jumps. (Anthropic)

Hg Forms Strategic Venture With Anthropic

Private equity firm Hg announced a strategic venture with Anthropic, signalling a fresh injection of capital into the frontier AI space. The partnership will fund Anthropic's next generation of models and expand its infrastructure, though financial terms were not disclosed. Hg's move follows a broader trend of investors seeking exposure to advanced language model development, hoping to capture upside from the growing demand for specialised AI services.

The deal highlights the continued appetite for funding in the sector, even as some firms adopt slower, safety‑first approaches. Observers note that the venture could give Anthropic a competitive edge in hiring and compute resources, but the long‑term impact will depend on the performance of forthcoming model releases and regulatory developments. (fnlondon.com)

GPT‑6 Astra Tests Anthropic’s Slowdown

Anthropic’s upcoming model, GPT‑6 Astra, is being used as a test case for the company’s deliberately slower development cadence. According to a Memeburn report, the model’s release is accompanied by a market that rewards the reduced visibility, with investors showing patience for a product that is harder to monitor. The article suggests that the slower rollout may allow more thorough safety checks while still delivering competitive performance.

The approach contrasts with the rapid iteration seen at other frontier labs, and it raises questions about whether a measured pace can sustain commercial momentum. Early signals indicate that customers value the trade‑off, but the long‑term viability of a slower cadence remains to be proved as competitors continue to accelerate. (Memeburn)

OpenAI Forms Mathematics Advisory Group

OpenAI announced the creation of an independent Advisory Group on Mathematics and Artificial Intelligence. The group, described in a TechCrunch article, will review emerging mathematical results produced by OpenAI’s systems and advise on communication strategies. OpenAI emphasised that the advisory panel will not be given authority to slow or redirect its research agenda, underscoring a focus on transparency rather than governance restraint.

The move follows a series of high‑profile breakthroughs in automated theorem proving and could shape how the broader community perceives AI‑driven mathematics. While the advisory group may enhance credibility, critics note that without formal oversight its influence on research direction may be limited. (TechCrunch; OpenAI)

From the Laboratories

•  OpenAI’s Higgsfield AI releases new video ad creation tools powered by GPT‑6 Astra, speeding small‑business production pipelines. (OpenAI)

•  ChatGPT and Claude makers OpenAI and Anthropic formalise a joint push for Australian copyright reform. (Capital Brief)

•  Jun Kim, creator of the oMLX library, joins Hugging Face to support the MLX community. (Hugging Face)

•  Researchers present a pruning technique that frames block removal as an Ising optimisation problem. (Hugging Face)

•  Meta’s Muse AI agent records higher download and daily‑active‑user numbers in North America than ChatGPT did after its mobile launch. (TechCrunch)

•  Import AI issue 473 discusses the United States’ superintelligence strategy and the limits of current AI understanding. (Import AI)

Border Surveillance Scrutiny

Investigation Challenges Utility Of Virtual Border Wall

A joint investigation by MIT Technology Review and Times of San Diego has examined the performance of the virtual wall of surveillance towers installed by the United States government. The study found multiple instances where individuals walked undetected through zones covered by AI-enabled towers, only to perish nearby. In one cited case from April 2024, a migrant died in southern New Mexico while within range of three surveillance units that failed to trigger a sufficient human response.

The findings suggest a disconnect between the billions of dollars spent on high-technology detection systems and the reality of migrant safety in border regions. Journalists documented cases, including a September 2025 fatality in San Diego, where surveillance cameras stood in plain sight while individuals remained lost and eventually succumbed to the elements. The report provides four specific policy recommendations intended to address these systemic failures in surveillance deployment and emergency response.

The investigation highlights that despite advanced sensors and detection capabilities, the technology often fails to catch individuals before they suffer fatal outcomes. The government has not yet provided a detailed response regarding how it intends to adjust the monitoring protocols to better align with humanitarian needs. Further inquiry into the specific technical limitations of the tower hardware and the operational oversight of the collected data remains necessary. (MIT Technology Review; MIT Technology Review; MIT Technology Review; MIT Technology Review)

OpenAI And Anthropic Seek Copyright Reform In Australia

OpenAI and Anthropic have formalised a push for changes to Australian copyright law. The companies argue that reform is necessary to support the ongoing development and investment in AI technology within the region. Anthropic, in particular, has been making a concerted effort to demonstrate the value of its investment case to Australian stakeholders.

This development follows recent industry complaints, with the industry body APRA noting that no major AI platform has requested to license copyright material in the country in the last four years. It remains unclear how local regulators will balance these industry demands against the concerns of rights holders. (Capital Brief; The Music Network; AFR)

In Brief

•  Tengai has released version 4.0 of its hiring screening software, explicitly designed to comply with requirements of the EU AI Act. (Mynewsdesk)

•  The United States has deferred federal AI regulation until December, citing the pressures of the midterm elections and a scheduled visit by President Xi. (조선일보)

•  Critics argue that the proposed EU Kids Act will undermine user privacy and impose burdensome age verification requirements on AI systems and online services. (EFF)

Security Vulnerability

Meta Agent Muse Faces Zero-Day Attack

The new AI assistant Muse, developed by Meta, is currently susceptible to a serious zero-day vulnerability. Security researchers report that a ClickFix attack provides a path for malicious actors to hijack the agent entirely. This exploit is particularly concerning given the level of privilege granted to the assistant within internal systems.

Muse operates with significant access to user data and system controls. The presence of such a vulnerability highlights the risks inherent in deploying agents with broad authority. At present, the exact scope of the compromise remains under investigation by engineering teams.

The incident raises difficult questions regarding the safety of production-grade AI agents. While Meta has not released a comprehensive patch, users are advised to monitor their system logs for anomalous activity. The vulnerability demonstrates that even advanced assistants remain bound by the security limitations of their underlying infrastructure. (Ars Technica)

TypeSafe AI Debuts Jev Decision Models

TypeSafe AI has introduced a new category of model designated as System One, or decision models. Unlike standard large language models that generate text, Jev is engineered to output floating point numbers. These values correspond to specific categories or binary yes or no decisions.

The shift from generative text to structured classification represents a different approach to model architecture. By returning numerical data, Jev aims to provide more predictable inputs for downstream automated systems. Whether this format proves superior to traditional LLM output for complex decision-making tasks is a subject of ongoing debate among developers. (Simon Willison)

AWS Releases Strands Harness

AWS has launched an open-source AI agent project titled Strands Harness. This tool is designed to facilitate multi-cloud operations, allowing developers to manage agents across diverse infrastructure environments. The project is now available for public review and contribution.

The initiative reflects a growing interest in standardising how agents interact with cloud services. By providing a unified harness, AWS hopes to address the complexities of cross-platform agent deployment. Further documentation on the integration capabilities of Strands Harness is available on their respective repository. (디지털투데이)

In Brief

•  Grok 4.7 has reached a score of 1657 Elo on the AA-Briefcase benchmark, narrowing the gap with Claude. (24/7 Wall St.)

•  Jun Kim, the creator of oMLX, has joined Hugging Face to provide additional support for the MLX community. (Hugging Face)

•  Researchers are applying physics techniques to LLM pruning by treating block removal as an Ising optimisation problem. (Hugging Face)

New Media Risks

Chinese Microdrama Industry Rents Faces For AI Productions

A new trend has emerged within the digital entertainment sector in China. Performers are now opting to rent out their faces for use in AI-generated shows, fueling a boom in microdramas. This practice represents a shift in how acting talent interacts with synthetic media tools. By licensing their physical features for machine-driven production, actors are enabling the creation of content that requires little to no human presence on set after the initial capture.

The consequences for the traditional acting profession remain a subject of active discussion. While these agreements provide a new revenue stream for some, they raise questions regarding the long-term control of an individual's digital persona. It is unclear how much oversight these actors retain over the final performance or the nature of the scripts their faces are asked to inhabit. The industry continues to expand as technology lowers the barrier for high-frequency video output.

This development mirrors broader anxieties regarding the place of human creativity in a landscape increasingly populated by automated content. As AI-generated shows become more common, observers are noting the difficulty of distinguishing between performances anchored in genuine human intent and those assembled entirely by algorithms. The permanence of these digital likenesses suggests that the legal and ethical framework for screen acting may require significant revision to protect the workers caught in this change. (Moneycontrol.com)

Autodesk Addresses AI Limitations In Film Production

A recent AI-generated film produced by Autodesk has drawn attention due to an uncorrected continuity error. Maurice Patel, the company's VP of M&E Strategy, pointed to the mistake as a practical demonstration of why current technology cannot yet replace the work of human animators. Despite the capability of artificial intelligence to generate complex imagery, it frequently fails to maintain logical consistency within a scene.

The error serves as a point of reflection for the creative trades. While tools for automated design and motion continue to proliferate, the human capacity for oversight and narrative coherence remains a vital component of professional film craft. (Creative Bloq)

In Brief

•  The debate over human creativity standards continues as international organisations consider how to incorporate AI music into current copyright systems. (asiae.co.kr)

•  iPhone owners in the US may now submit claims for a 250 million dollar settlement regarding Apple’s failure to deliver on promised AI upgrades to Siri. (The Verge)

•  New online services now offer AI detection for text, image, music, and video to help distinguish synthetic from human-made content. (AI or Not)