China implants the first approved invasive brain chip

Today’s most important AI news

Today’s most striking headline comes from China, not Silicon Valley: the country has just become the first in the world to officially approve an invasive brain chip for use in patients. In parallel, NVIDIA presented Cosmos 3, an open model designed specifically for reasoning about physical systems. And in the world of agents, new research reveals how badly LLMs handle tasks as everyday as operating financial spreadsheets, which has direct implications for any company considering automating processes.

China approves the world’s first invasive brain chip: what this changes

Last October, Dong Hui — a 39-year-old man with cervical paralysis caused by an accident — tried to hold a pen and write, and succeeded thanks to a chip implanted directly in the brain. This week, MIT Technology Review confirms that China has just obtained regulatory approval for this device, becoming the first country in the world to give official green light to an invasive BCI (brain-computer interface) for clinical use.

This puts China directly on par with — and in some respects ahead of — initiatives like Neuralink in terms of regulatory speed. Where the United States is still in experimental stages with partial FDA approvals, China has created an institutional precedent that can accelerate clinical trials, private investment and hardware development across Asia. The move isn’t just medical: it’s geopolitical.

For Latin American professionals in digital health, medtech or venture investment, this opens a concrete question: will the region’s health systems align with Western regulatory standards or will they also look toward Asian models? The Chinese approval could become an alternative reference, especially for countries with more agile regulatory frameworks.

NVIDIA Cosmos 3: the first open model for AI to reason about the physical world

NVIDIA announced the launch of Cosmos 3 on Hugging Face this week, describing it as “the first open omnimodal model for reasoning and action in physical AI”. Unlike conventional LLMs trained mainly on text and static images, Cosmos 3 is designed so that autonomous systems — robots, vehicles, industrial machinery — can reason about physical environments in real time and make action decisions.

The data point that matters: research published this week on ArXiv (on AI agents for nuclear reactor control) documents that today’s best vision-language models reach only 50-53% accuracy on basic quantitative physics tasks, behaving like “approximate guessers” that preserve semantic coherence but violate real physical constraints. Cosmos 3 aims directly at that problem, and the fact that it’s an open model — available on Hugging Face — puts it within reach of teams without access to NVIDIA’s proprietary infrastructure.

For developers and product teams working in manufacturing, logistics or robotics, Cosmos 3 is probably the most relevant launch of the week. Not because it’s perfect, but because it’s the first model of its kind they can download, test and adapt without depending on a closed API. The most immediate use case: simulating physical environments to train agents before deploying them on real hardware.

LLMs and financial spreadsheets: a bigger gap than expected

BlueFin, a benchmark published this week on ArXiv, evaluates how well LLM agents handle spreadsheet tasks in professional finance: data synthesis, workbook manipulation and understanding complex structures. The results are revealing by contrast: while LLMs show solid capabilities in code generation or writing, their performance on real financial spreadsheets — with multiple tabs, nested formulas and domain conventions — drops significantly.

This matters because spreadsheets are the invisible operating system of corporate finance in Latin America. According to the paper’s authors, there are “orders of magnitude more users of spreadsheet software than professional developers in the world”, and yet existing benchmarks barely cover them. BlueFin is the first serious attempt to measure this capability rigorously.

The practical implication is direct: if you’re evaluating whether an AI agent can automate financial analysis at your company, don’t assume performance on code benchmarks translates into performance in Excel or Google Sheets with real data. BlueFin’s tests suggest there’s still a significant gap that product teams should measure before committing critical workflows.

In a nutshell

Today’s pattern isn’t technological, it’s about where the limits are moving. China has just shown that regulatory advantage can be as decisive as technical advantage. NVIDIA is betting that the next frontier isn’t language but physics. And BlueFin reminds us that LLMs still have blind spots on tasks that seem trivial but move millions of dollars every day. What ties these three topics together is the same tension: the distance between what AI can do in a lab and what it can do reliably in the real world remains bigger than the headlines suggest.

---

Sources used: https://www.technologyreview.com/2026/06/01/1138133/china-world-first-brain-chip/, https://www.technologyreview.com/2026/06/01/1138207/the-download-china-bci-brain-implant-nvidia-ai-chips-laptops/, https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai, https://arxiv.org/abs/2512.23292, https://arxiv.org/abs/2605.30907