Nvidia's Jensen Huang declares "AGI has arrived"

What happened: On September 6, Nvidia CEO Jensen Huang declared that "AGI has arrived," crediting OpenAI's newly unveiled GPT-6 Astra, a model trained on more than 100,000 Nvidia Grace Blackwell GPUs with another 400,000 reportedly coming online. OpenAI itself stopped short of the claim, a notable restraint from a company not known for underselling. Although president Greg Brockman did tell reporters "welcome to the AGI era." The pushback was immediate: AI researcher and longtime industry critic (and an AIMed speaker) Gary Marcus said Huang had jumped the gun, and skeptics noted that the declaration says at least as much about Nvidia's order book as about machine intelligence. The exchange set off a week of dueling definitions, with no two commentators agreeing on what "artificial general intelligence" actually requires. AGI is a debate the field has deferred for a decade and apparently can defer no longer.

Why it matters to the AIMed audience: In my humble opinion, AGI does not have an “arrival date” and the announcement is a bit of showboating; it is akin to knowing the exact date that your adolescent son or daughter has become an adult. Whether or not "AGI" is the right word, a major hardware supplier and a frontier lab are now publicly staking claims about crossing a threshold most health-system leaders still file under science fiction. The gap between that rhetoric and reality on the wards where "AI" often means a sepsis alert with a middling positive predictive value, is exactly what breeds C-suite skepticism of vendor pitches. Two takeaways for us. First, learn to translate: when the loudest voices in AI have a financial stake in the answer, definitional claims are marketing until proven otherwise. Second, don't let skepticism curdle into complacency as the underlying capabilities are genuinely compounding, whatever we call them. A good governance committee can address claims like Huang's directly rather than letting them sit unexamined in the room.
 
Anthropic researcher publicly resigns, warning the industry is racing past safety

What happened: Jacob Coxon, a researcher who spent three years doing safety-adjacent work at both Anthropic and OpenAI, the two labs most publicly identified with responsible development, resigned. He posted a public statement warning that frontier labs are now more focused on competing with one another as well as with global competitors to the most capable model than on AI safety. He went further, noting that some in the field believe uncontrolled AI could threaten human life within the decade. Resignation-with-manifesto has become a recurring genre in AI, but this one landed differently: it came from inside the safety-branded labs, in the same week those labs shipped their most capable models as it named the competitive race itself, not any single model, as the core risk.

Why it matters to the AIMed audience: As we know, health systems don't just deploy models, these systems deploy the enterprises behind the models. When a safety researcher walks out of these labs saying the race has overtaken the guardrails, it becomes a live, quotable data point for the question every skeptical board asks: can we trust the labs building the models we're deploying? The honest answer is that trust must be structural, not reputational with contracts, independent evaluation, monitoring, and exit ramps rather than naive faith in any vendor and its culture. Healthcare has an advantage here: we already know how to treat powerful interventions as guilty until proven safe as we are risk-averse as an industry. This resignation is not a reason to halt adoption, but rather it is an opportunity to import our own field's instincts into AI governance. This is also a reminder that upstream lab culture is now part of every downstream risk assessment. Bottom line: When most if not all the AI gurus are concerned about the AI excessive velocity of capability, we should worry. 
 
Humanoid robots go consumer: China's UBTech starts delivering an emotional-AI companion this month

What happened: The humanoid era is arriving through the front door. Shenzhen-based UBTech begins first customer deliveries of its UWORLD U1 units in mid-September: a full-size "ultra-bionic" companion humanoid with more than 13,000 orders logged by its June 30 launch (priced from roughly $17,600 to $145,000 across three models.) The U1 pairs 88 degrees of freedom and lifelike silicone skin with an emotion-aware large language model and locally encrypted memory, and is marketed explicitly to China's solo dwellers and empty-nest seniors. The macro picture is equally striking: global humanoid shipments in the first half of 2026 rose roughly 272% year over year to about 20,000 units, with Chinese makers holding over 90% of the overall market share. Unitree completed mainland China's first pure-play humanoid IPO in August; and on September 2, Sam Altman confirmed OpenAI "will definitely do a humanoid."

Why it matters to the AIMed audience: On my last several trips to China, I have been increasingly impressed with the accelerating evolution of the humanoid robots there. Embodied AI is entering homes before it enters hospitals, and it is entering as a caregiver-adjacent product. A companion robot running an emotional LLM, sold to isolated seniors, is a healthcare intervention in everything but regulatory category: it will shape mood, adherence, loneliness, and eventually when and whether its owner seeks care. Clinicians will encounter these robots the way we encountered consumer wearables and through patients, not procurement. Our field should get ahead of the questions now: What does an emotion-aware machine owe a user with early dementia or depression? Who audits its advices? What happens to the encrypted "memories" when the owner is hospitalized or dies? And with Asia commanding production while the U.S. races to respond, the geography of embodied AI will not be decided in American health systems. We should monitor these robots the way our cardiologists have watched the smartwatch: skeptically.
 
Google and the Johnson & Johnson Foundation commit $5M to train 250,000 nurses on AI

What happened: Google and the Johnson & Johnson Foundation announced a three-year, $5 million grant to build AI skills across major nursing organizations, with a target of reaching 250,000 nurses and an explicit priority on underserved and rural regions. The arithmetic is worth pausing on: $5 million across a quarter-million nurses is $20 a head, so this is seed capital for curricula and train-the-trainer infrastructure delivered through existing nursing bodies, not a tuition program. Still, the choice of workforce is the story. Nursing is the largest licensed profession in healthcare, the one with the most patient-facing hours, and historically the last to be consulted on the technologies it is asked to absorb. Two consumer-scale brands just declared, with money, that nursing AI literacy is a national-scale challenge, and that rural and underserved settings should be first in line rather than last.

Why it matters to the AIMed audience: I personally am very excited about this commitment to fund nursing AI education as have seen increasing numbers of nurses at both the ABAIM courses and the AIMed meetings, but I am just not sure how the financing of this ambitious program will be executed. This is direct evidence that workforce AI-literacy funding is scaling nationally, and it lands squarely on AIMed26's nursing track: the demand side of AI education is now being subsidized, which strengthens the case for positioning the conference as a training and CE destination rather than a vendor showcase. For health-system leaders, it is also a benchmark: if external foundations are funding AI literacy for 250,000 nurses, an internal budget that allocates nothing for the same purpose is officially very much behind the field. And there is a deeper point about work design. AI deployments fail most often at the workflow layer, and nurses own the workflow the most. Educating the people who will actually supervise, override, and troubleshoot AI at the bedside is not a soft investment; it is the difference between a tool that gets used and a tool that gets worked around. Lastly, as there are about 5 million nurses in the United States alone, it may be a better strategy to fully educate about 5,000-10,000 nurses or so early adopters and allow these nurses to be the educators for the rest. 

Nearly 50 hospitals exposed clinician schedules via public links — and AI was the discovery tool

What happened: A Philadelphia Inquirer investigation, reported by Becker's Hospital Review, found that nearly 50 U.S. hospitals and health systems- including Penn Medicine, Children's Hospital of Philadelphia, ChristianaCare, Johns Hopkins Medicine, UCSF, and Cedars-Sinai- had publicly accessible QGenda workforce-scheduling pages exposing clinician schedules and, in some cases, personal cellphone numbers. The pages never surfaced in ordinary Google searches, which is precisely why they sat unnoticed for so long. The Inquirer's reporters instead used OpenAI's Codex to systematically identify live public links nationwide- an AI agent doing in hours what manual searching could never have accomplished. After being contacted, at least 10 systems removed access or added password protection. QGenda, the scheduling vendor at the center of the story, did not respond to requests for comment.

Why it matters to the AIMed audience: This is the rare story in which AI was the discovery instrument rather than the vulnerability issue. The data sat in plain sight for years but was practically invisible until someone pointed an AI agent at it- and journalists got there first. The uncomfortable corollary here: the same scan is trivially repeatable by anyone at any time. For systems still in the "exploring" stage of AI adoption, this is a concrete, name-brand prompt to audit every workforce platform, portal, and vendor-hosted page for public exposure before someone else does. It should also reframe the governance conversation: an AI risk assessment that only inventories the tools you deploy misses the AI tools pointed at you. Red-teaming your own digital perimeter with the same agents an adversary would use is now table stakes. We are just at the beginning stage of AI and cybersecurity challenges[ there will be a cybersecurity module at AIMed26.