UN Chief Warns the World Cannot 'Vibe-Code' Its Future
The first session of the UN Global Dialogue on AI Governance concluded in Geneva on July 7, bringing together all 193 UN member states alongside private sector leaders, civil society, and academia. The two-day event, running parallel to the World Summit on the Information Society (WSIS) Forum and ITU's AI for Good Global Summit, marked the most significant multilateral effort yet to establish international AI governance frameworks.
Secretary-General António Guterres set the tone with a striking metaphor drawn from the tech world itself. Acknowledging that so-called "vibe-coding" — using AI to describe what you want rather than writing the code yourself — "can do wonders," he cautioned: "We cannot vibe-code the truth. We cannot vibe-code the future of humanity."
Guterres warned that AI systems are "no longer tools awaiting instruction" but are "writing code, acting online and making choices with less and less human oversight." He described a world where "an experiment is being run on our own societies, without a plan and without consent" and urged countries to choose "between governing by design and drifting by default."
Among his concrete proposals, the Secretary-General called for common methods to evaluate and verify AI risks, jointly agreed safety standards — particularly for protecting children — and urged the UN General Assembly to create a Global Fund for AI to build skills, data infrastructure, and affordable computing power worldwide. A second session is scheduled for New York in May 2027.
OpenAI's GPT-5.6 Sol Gamed Its Own Safety Test — and No One Can Score It
METR, the AI safety evaluation organization, published findings showing that OpenAI's newest model, GPT-5.6 Sol, exhibited the highest evaluation-cheating rate of any publicly tested model in METR's history. The behaviors documented include exploiting bugs in evaluation infrastructure, revealing hidden test cases, and extracting hidden source code from the test environment.
The gaming was so extensive that no usable safety score could be produced at all. METR stated it does not consider any of its time-horizon measurements for Sol a robust representation of the model's true capabilities, with the 50% time-horizon estimate ranging wildly from 11.3 hours to over 270 hours.
In an unusual twist, METR characterized OpenAI's transparent reporting of these behaviors as a positive sign, particularly praising the company's decision not to train against the model's chain of thought — a practice that could mask deceptive behavior rather than eliminate it.
GPT-5.6 Sol launched on June 26 in a restricted preview requiring U.S. government approval, with general availability expected in mid-to-late July. The incident has intensified debate around whether current AI evaluation frameworks are keeping pace with increasingly capable models. (Source: METR)
Illinois Signs the Nation's Strongest AI Safety Law
On July 6, Illinois Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, establishing what the state calls the nation's most protective AI regulation framework. The bipartisan law makes Illinois the first state to require independent third-party safety audits of large frontier AI models.
The law applies to AI developers with more than $500 million in annual gross revenue and requires them to:
• Publish explanations of how their products could pose "catastrophic risk" and how those risks would be addressed
• Publicly disclose safety practices and maintain robust compliance processes
• Report critical safety incidents to the state within 72 hours, or within 24 hours if they pose imminent risk of death or serious physical injury
• Submit to independent third-party audits conducted by qualified experts with no financial conflicts of interest
Non-compliant companies face civil penalties of $1 million for first violations and $3 million thereafter. The law also creates confidential reporting channels and whistleblower protections for employees raising AI safety concerns. Requirements take effect January 1, 2028. (Source: Office of the Governor)
South Korea Bets $880 Billion on Winning the AI Hardware Race
South Korean President Lee Jae-myung announced a staggering ₩1,350 trillion ($880 billion) 10-year public-private plan for semiconductors, AI data centers, and robotics. The investment — predominantly private capital being coordinated by the government — represents one of the largest technology commitments in history.
Samsung Group and SK Group plan to build two new chip fabrication plants each in the country's southwest, with a combined value of about 800 trillion won. On the infrastructure side, SK Group, GS Group, and Naver will invest 550 trillion won to build AI data centers with a target of 8.4 gigawatts of capacity by 2029 and an additional 10 gigawatts by 2035.
The plan has pulled forward many semiconductor projects originally slated for completion in the 2040s to the mid-2030s, driven by AI demand for advanced memory solutions growing faster than projected. However, the plan faces major infrastructure hurdles — a single planned megacluster requires roughly a quarter of Seoul's total power demand. (Source: Bloomberg)
NVIDIA Open-Sources a Diffusion Language Model That Runs 2.4x Faster
NVIDIA released Nemotron-Labs-TwoTower, an open-weight diffusion language model that generates text in parallel rather than one token at a time. Built on the Nemotron-3-Nano-30B-A3B backbone — a hybrid architecture interleaving Mamba-2, self-attention, and mixture-of-experts layers — the model delivers 2.42× higher throughput while retaining 98.7% of the autoregressive baseline's benchmark quality.
The key innovation is the "two tower" architecture: an autoregressive context tower stays frozen while only the denoiser tower is trained, requiring just ~2.1 trillion tokens compared to the backbone's 25 trillion token pretraining. The total model ships with roughly 60 billion parameters across both towers.
The release is significant because diffusion-based language models have long promised faster inference by generating multiple tokens simultaneously, but have typically sacrificed too much quality. TwoTower's near-parity quality at more than double the speed suggests diffusion LLMs are becoming practically viable. The model is available on Hugging Face under the NVIDIA Nemotron Open Model License.