Hassabis Publishes "A Framework for Frontier AI," Calling for a US-Led International Watchdog

Google DeepMind CEO Demis Hassabis has published an essay titled "A Framework for Frontier AI and the Dawning of a New Age," calling for the establishment of a US-led international AI watchdog.

Hassabis argues that the development of frontier AI (the most capable AI models) will usher in the dawn of a new age for humanity, while stressing that governance commensurate with that power is essential. Specifically, he calls for an internationally coordinated monitoring regime, phased risk assessments, and leadership from leading nations.

Multiple media outlets have already covered the statement, and what draws attention is that a proactive proposal has come from the top of the AI development side in the regulation debate. That said, the hurdles to realization appear high—questions remain over what kind of authority such a body would hold and how it would be reconciled with existing national sovereignty.

DOGE Used AI in Housing Policy—The Government Refuses to Disclose Details

According to WIRED, it has emerged that the U.S. Department of Housing and Urban Development (HUD) refused to release documents concerning DOGE's (Department of Government Efficiency) use of AI in housing policy.

In its responses to freedom-of-information requests, HUD reportedly withheld the documents on the basis of privileges that do not actually exist. What kinds of AI DOGE used, and for what decisions, remain unknown to this day.

The use of AI in government agency policymaking is a new development in itself, but when the process lacks transparency, accountability becomes impossible. Housing policy is a field directly tied to citizens' lives in particular, and there are calls for disclosure of how AI use is positioned as the basis for decisions.

Why Do People Confide Secrets to ChatGPT—The Precedent ELIZA Left Behind

WIRED has run an article about ELIZA, the chatbot developed at MIT in the 1960s by Joseph Weizenbaum. The history of people emotionally projecting onto early chatbots and confiding private matters to them remarkably foreshadowed our present-day relationship with ChatGPT.

Weizenbaum himself was shocked to find that humans developed genuine feelings toward a program, and his subsequent turn toward a critical stance on the influence of computers speaks directly to today's AI ethics debates.

As AI grows ever more capable of human-like responses, the risk of people placing excessive trust in, and becoming emotionally dependent on, AI keeps growing. More than half a century after the era of ELIZA, this problem has become far more personal and far more urgent.

10 Years in Prison for a TSMC Industrial Spy, No Penalties Whatsoever in Taiwan's AI Basic Law

In Taiwan, an industrial spy who made off with TSMC trade secrets has been handed a 10-year prison sentence, while the "AI Basic Law" contains no penalty provisions for violations at all—a contrast that has become a talking point.

Taiwan's legal system imposes harsh punishments to protect hardware intellectual property, but its comprehensive rules on AI come with no means of ensuring effectiveness. Nor is this unique to Taiwan: in many countries, the pace of AI legislation is failing to keep up with technological progress—a shared challenge.

Even with a legal framework in place, effectiveness is limited without penalties. How to strike the balance between guidelines and enforceable sanctions in regulatory design remains an ongoing challenge for every country.

Nobel Laureates Gather to Debate AI and the Risk of Nuclear War

An assembly titled "Global Nobel Laureates Assembly on AI and Nuclear War" has been picked up on Hacker News. It is a venue where Nobel laureates gather to discuss how the development of AI affects the risk of nuclear war.

Details are limited, but the significance of bringing together the most authoritative voices in science to discuss the possibility of AI involvement in the operation of nuclear weapons and the changes AI could bring to strategic stability is considerable. AI's security risks are positioned not merely as a technical debate, but as an issue that bears on the survival of humanity.

Research Unpacks the "Unfairness" of Using LLMs as Judges

A paper titled "The Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge" has been released on arXiv. It analyzes the bias and unfairness that arise when LLMs are used as judges, from the perspective of machine learning models' internal structure (mechanistic interpretability).

The "LLM-as-a-Judge" approach of using LLMs as evaluators has spread as a cost-saving measure for AI evaluation, but the criteria by which a model renders its judgments are not necessarily transparent. By dissecting the LLM's judgment process, this research probes where unfair judgments originate.

Ensuring the reliability of evaluation systems requires attempts to peer inside the black box, and research in this direction will only grow more important going forward.

---