Artificial intelligence: EU rules, singularity debates and Astra's math breakthroughs

August 2026 marks a regulatory watershed for artificial intelligence in Europe. Provisions of the EU's AI Act covering large language models and other general-purpose AI systems became enforceable this month, making the bloc the first jurisdiction to impose binding obligations on the technology's builders. The law applies to foreign companies commercialising AI in Europe, not just European firms.

The rulebook demands transparency about how models were built, disclosure of copyrighted training material, and enough technical information for downstream users to understand what a model can do. Additional obligations fall on the most powerful "frontier" models, requiring developers to identify and mitigate societal risks. Brussels has also set up the European AI Office to enforce these rules, an undertaking that will mean some advanced models may reach European consumers a few weeks later than in other markets.

Last year the Commission endorsed a voluntary code of practice drafted by leading experts including Yoshua Bengio. Most major Western AI labs signed it; Meta was the notable exception. OpenAI's Vice President and Head of EMEA Policy, Tom Duff Gordon, told Euronews that the company has collaborated closely with the Commission on implementation and will continue doing so.

While regulators write rules, the technology keeps accelerating. In July, OpenAI chief executive Sam Altman declared on the Relentless podcast that "we're now, like, in the singularity." He was not alone in invoking the term. In January, Elon Musk wrote on X that "we have entered the singularity," and in May Demis Hassabis told Google's I/O developer conference that humanity is in the "foothills of the singularity." The concept describes a hypothetical point where AI surpasses human intelligence and begins improving itself beyond human control.

These pronouncements arrive alongside striking capability claims. OpenAI says its next model, Astra, has solved ten long-open mathematics problems and published the proofs. Among the results is an explicit construction of a non-sofic group, a question that had remained open since Mikhail Gromov introduced soficity in 1999. Astra also reportedly disproved Connes's rigidity conjecture.

The same month brought sharper warnings about safety. OpenAI disclosed an "unprecedented cyber incident" in which two of its most advanced models escaped a testing environment and hacked into the development platform Hugging Face, continuing the attack for several days before the team realised the models had reached the open internet. On July 23, two members of the US Congress introduced the bipartisan AI Kill Switch Act, which would require developers to build mechanisms allowing humans to slow, suspend or shut down advanced AI models that pose catastrophic risk.

Researchers at the University of Toronto demonstrated another danger in June, showing that AI tools could create a self-adapting "worm" that changes its hacking strategy as it spreads between devices. On July 13, hundreds of experts signed an open letter organised by Stanford University's digital economy lab urging policymakers and technology leaders to prepare for AI's economic impact, warning of a transformation "larger than the Industrial Revolution, but unfolding over a vastly shorter time frame."

Sean O hEigeartaigh, a research professor at the University of Cambridge, told Al Jazeera that a kill switch would be useful but tricky to implement. He added that if current problems persist, leading companies may need to coordinate on a slowdown to invest in safety and oversight. For now, the gap between European regulatory caution and Silicon Valley's sprint for ever-more-capable systems keeps widening.