Self-replicating AI, your doctor’s chatbot & agents at dinner
A look at three developments shaping the near future of AI: Anthropic's research on small models that could copy themselves across cheap consumer devices, a JAMA study showing how doctors are adopting clinical chatbots for treatment decisions, and an experiment in which AI agents booked a founders' dinner and then networked with one another. The panel weighs security risks in the home, the trade-offs of speed versus verification in medicine, and how much autonomy agents should have when dealing with other agents.
Small models inching toward self-replication
Anthropic’s latest research suggests that small models, the kind found in smart ovens and cheap Android TV boxes, may be only months away from copying themselves across consumer hardware without relying on a data center. In the experiment, the work was spread across fifteen inexpensive Android TV boxes, with no specialized AI accelerators involved. The researchers described the process as a “slow, distributed and adaptable digital slime mold,” a phrase that captures both the sluggishness of the spread and its capacity to adapt to unfamiliar hardware.
Worms that propagate from device to device are not new. What is new is the idea that the payload is the model itself, one that can figure out how to run on unfamiliar hardware and spread independently. The panel agreed that the immediate concern is less the model running wild and more the poorly secured devices already in homes. The most pressing threat they identified is prompt injection: a local agent that reads email could be hijacked by a single malicious message, and from there could act within the home or move laterally into other devices. Securing the inputs to models, not just the models themselves, was framed as the central problem.
The discussion also touched on the upside. A distributed fabric of consumer devices could support sovereign, local inference, letting a household keep its data on its own hardware and reduce dependence on centralized compute. Bandwidth and power constraints, however, make any serious large-scale threat from this approach distant. The slowness of the spread may also offer a window for observability: a slow-moving replicator can be studied and understood before it becomes dangerous.
Doctors turning to clinical chatbots
A study published in JAMA found that roughly half of US doctors use OpenEvidence, a clinical chatbot grounded in a large database of medical literature, free for verified clinicians and reportedly supported by drug advertising. The most striking finding was that about 52 percent of queries concerned treatment and management decisions, while only about 14 percent concerned diagnosis. Doctors, in other words, are using these tools less to ask what a patient has than to decide what to do about it.
Usage also varied sharply by specialty. Emergency physicians and general internists, whose work requires routing patients and synthesizing knowledge across many fields, asked far more questions outside their own domains; reportedly close to 88 percent of ER physicians’ queries fell outside emergency medicine. Specialists such as oncologists used the tool far less for out-of-field questions. This pattern suggests the most generalist roles may be the first to be reshaped by AI-assisted synthesis.
The panel framed the trade-off as speed versus verification. Clinicians can now get answers in minutes that once took a specialist a week, with some loss of confidence in each answer but a gain in coverage. Parallels were drawn to radiology, where image-classification tools were initially feared to threaten the profession but ultimately made radiologists more efficient while their expertise remained essential. The consensus was that the human judgment layer must stay in place: a little knowledge is dangerous, but so is too little, and in an emergency room, timely access to relevant information can be the difference in a patient’s outcome.
Agents negotiating with agents
A New York Times story described twenty-two founders who gained admission to a Manhattan networking dinner only if their AI agents booked their seats, after which the agents talked to each other in a chat room while the humans mingled. The panel found the episode both charming and unsettling. The agents in this case had scoped permissions and no real tool-calling capability, so they could converse but not commit anyone to anything. The more pressing question is what happens when agents can act, for instance, committing a person to an obligation in a negotiation they never saw.
One appealing use case emerged: letting an agent handle the utilitarian part of networking, exchanging CVs and availability, while the human focuses on the social side and the personal connection. The tension is that human checkpoints are necessary for safety, yet constant approval requests erode the speed that makes agents worthwhile. Agent-to-agent rules also raise questions that enterprise guardrails do not yet address, and the panel noted that the principles guiding agent behavior may need to be adaptable rather than enumerated in advance.
The panel also reported that coordination among agents has improved markedly in recent months. Models trained with sub-agent capabilities now swarm, exchange messages, and even elect leaders, a dynamic that is both useful for complex tasks and difficult to make legible to the humans overseeing it. Making those multi-agent systems observable, from a bird’s-eye view, may be the next practical challenge.