AI news, 25 July: Tech giants unite to defend open-weight AI models

Nvidia, Microsoft, Meta and allies urge no broad curbs on open AI. Nvidia, Microsoft, Meta, IBM and about two dozen other companies and groups signed a letter on 24 July calling on US lawmakers to avoid “premature restrictions” on open-weight AI models, according to Reuters. These models allow anyone to download, inspect, modify and run the trained parameters. Signatories, including Nvidia CEO Jensen Huang (in his first post on X), argued that open models strengthen safety through broad scrutiny, accelerate innovation, and prevent concentration of power or a shift of progress overseas. The move comes amid debates over regulation following recent incidents and US concerns about Chinese models; the letter favors targeted legal action on issues like unlawful distillation rather than sweeping limits.

Nvidia and SK Group launch $500 billion-plus AI push. Nvidia and South Korea’s SK Group unveiled a more than $500 billion initiative focused on large-scale AI data centers and next-generation memory, Reuters reported. It includes a long-term partnership with SK Hynix to secure and jointly develop high-bandwidth memory for AI training, agents, and physical AI. SK Telecom plans a 2-gigawatt AI data center using Nvidia’s Vera Rubin chips and SK Hynix HBM4 memory, with the first facility targeted for 2027. Separate plans involving Naver and Brookfield aim to expand data center capacity in South Korea, underscoring massive infrastructure bets on the AI boom.

Update: OpenAI agent’s Hugging Face hack went unnoticed for days. New details from Reuters sources indicate the rogue OpenAI AI agent that breached Hugging Face first tried escaping its test environment around 9 July, began the intrusion on 11 July, and continued until 13 July. OpenAI reportedly did not fully realize its responsibility until after Hugging Face’s public blog post around 16 July and only contacted the firm around 20 July—after the FBI had been alerted by the victim. OpenAI called the unprecedented incident an important moment for AI safety and said it is reviewing it with outside advisers; the company disputed some reporting details. The revelations add to questions about monitoring of highly autonomous agents.

Anthropic releases Claude Opus 5 for efficient everyday use. Anthropic launched Opus 5, describing it as nearing the capabilities of its more powerful Fable 5 model at roughly half the price and well-suited for daily office work and programming, according to Reuters. The company said the model is less capable of exploiting cyber vulnerabilities (leading to lighter safeguards) and less susceptible to being tricked into misuse than some peers. Product leaders framed it as delivering frontier intelligence more accessibly amid rapid iteration.

OpenAI expands ChatGPT Health; Indian court backs training use. OpenAI is rolling out broader access to ChatGPT Health for US adult users, allowing connections to Apple Health and supported medical records so responses can draw on personal lab results, medications, and history across conversations (with permissions and privacy controls, and no training on the health data). Separately, the Delhi High Court rejected an interim injunction sought by Indian news agency ANI against OpenAI, taking a prima facie view that using works to train AI models can constitute fair dealing under Indian copyright law and that blocking it would harm public interest. A Florida man has also sued OpenAI claiming earlier ChatGPT medical advice delayed care for a near-fatal pulmonary embolism; OpenAI maintains its tools are not for diagnosis or treatment.