AI News

Follow the latest AI and automation news

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Microsoft's SkillOpt boosts GPT-5.5 by using nothing but a trained Markdown file

Microsoft, in collaboration with three Chinese universities, has developed SkillOpt, a method aimed at enhancing educational documentation for intelligent agents. This approach leverages traditional model training principles, utilizing a simple Markdown file to boost GPT-5.5 performance by 23 points on procedural tasks. SkillOpt excels in transferring improvements across various models and environments, such as Codex and Claude Code.

Models

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Two Turing Award Winners Convene at Beijing Zhiyuan Conference, Confronting the Theoretical Black Hole Behind AGI

The 8th Beijing Zhiyuan Conference began on June 12, 2026, focusing on cutting-edge AI research. The opening featured Turing Award winners Whitfield Diffie and Andrew Barto, who discussed critical aspects of AI development. Diffie highlighted the paradox of defining safety in AI agents, while Barto emphasized the challenges of reward functions in reinforcement learning, indicating a gap in theoretical foundations compared to industry enthusiasm.

Models

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Claude Fable 5 outpaces GPT-5.5 by 13 points on FrontierMath's toughest problems

Anthropic's Claude Fable 5 achieved an impressive 88% accuracy on the toughest FrontierMath tier, a significant leap from Opus 4.5, which was below 10% in early 2026. In comparison, OpenAI's GPT-5.5 reached about 75% on the same tier. This rapid improvement in AI math accuracy indicates an accelerating pace of advancements in the field. This progress poses a considerable challenge to competitors, suggesting that such results will influence future model development strategies, emphasizing the need for enhanced performance in complex mathematical tasks.

💡

A German Court Has Ruled That Google Is Liable for False Statements Generated by AI Overviews

What impact will this legal liability have on AI development strategies?

Business

Open model Kimi K2.7 Code undercuts GPT-5.5 and Claude by up to 12x on price per token

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Business

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Open model Kimi K2.7 Code undercuts GPT-5.5 and Claude by up to 12x on price per token

Moonshot AI has launched the Kimi K2.7 Code model, an open-weight model featuring one trillion parameters designed for programming. While it still lags behind GPT-5.5 and Claude Opus 4.8 in programming tests, it offers a significantly lower cost per token. The key question is not whether Kimi K2.7 Code is the best model, but whether the increase in operations achievable within the same budget compensates for the quality gap.

Infrastructure

Do China’s export curbs on tungsten threaten Japan’s AI chip supply chain?

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Infrastructure

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Do China’s export curbs on tungsten threaten Japan’s AI chip supply chain?

Two Japanese companies may cease production of hexafluoride tungsten gas, a vital material for chip manufacturing, starting next month. This decision comes amid tightening export restrictions from Beijing, which have led to a price increase of over 200% year-on-year. This gas is essential for developing precise communications within advanced chips. Market data indicates that rising demand for chips, coupled with supply chain bottlenecks, has significantly contributed to this price surge. Hexafluoride tungsten gas is a critical component in semiconductor manufacturing, making it pivotal for artificial intelligence technology.

Governance

Anthropic Halts AI Models Following Government Directive

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Governance

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Anthropic Halts AI Models Following Government Directive

Anthropic announced on Friday the suspension of its AI models, "Fable 5" and "Mythos 5," for all users following a government directive that restricts access to foreign nationals. This decision arises from national security concerns, as the government identified potential methods to bypass the model's security system. The government's apprehensions center on the possibility of using "Fable 5" to identify software vulnerabilities, escalating tensions between Anthropic and regulatory bodies. The relationship soured further after the company denied the U.S.

Governance

US blocks foreign access to Anthropic’s newest AI models over security risks

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Governance

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

US blocks foreign access to Anthropic’s newest AI models over security risks

Washington has ordered Anthropic, a leading U.S. AI company, to block all foreign access to its new models, Fable 5 and Mythos 5, due to national security concerns. This unprecedented move marks the first export control action targeting specific AI models and follows closely on the heels of their launch. It reflects Washington's increasing approach to regulating advanced AI technologies. These measures indicate a significant shift in how governments are managing AI technologies, potentially impacting innovation and international collaboration in the field.

Infrastructure

Chinese team builds first commercial ‘3-lane highway’ in optical fibre to boost capacity

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Infrastructure

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Chinese team builds first commercial ‘3-lane highway’ in optical fibre to boost capacity

Earlier this month, China activated the world's first three-band optical fiber communication system. Developers claim this technology could significantly expand the capacity of future AI networks, with a single fiber able to carry over five times the traffic of conventional systems. The project, completed in Qingdao, Shandong province, was jointly developed by state-owned companies. Additionally, the transmission capacity per core increases by nearly 50%, showcasing a significant advancement in communication technology.

Governance

Anthropic Says It’s Taking Claude Fable 5 Offline to Comply With US Government Order

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Governance

13 Jun 202628 Dhuʻl-Hijjah 1447 AH

Anthropic Says It’s Taking Claude Fable 5 Offline to Comply With US Government Order

The government has reported awareness of a method to bypass the security of Fable 5, as stated by the company in a blog post. This method is referred to as 'jailbreaking', raising concerns about the system's security. Experts indicate that this development could undermine user trust in Fable 5. Immediate attention from developers is required to enhance security measures. Failure to address this vulnerability effectively may lead to negative repercussions for the company's reputation in the market.