Open LinkedIn on any given morning and you will find two parallel universes.
In one, AI is rewriting medicine from the ground up. Every other post is a breathless announcement -- a new model, a new benchmark, a new startup that will "revolutionize" something. The tone is that of a sports commentator calling a no-hitter in the second inning.
In the other universe, AI is a party trick that gets drug interactions wrong and hallucinates journal citations. A new study drops showing an LLM performing worse than a second-year resident on some narrow task, and the comments fill with vindication. "See? I told you this was all hype."
The truth, as it tends to, lives somewhere in the middle. And I think it is worth spending a few minutes there.
The Race for Your Health Data
The biggest development in the past six weeks is not a model release -- it is a land grab.
In January, OpenAI launched ChatGPT Health, a dedicated tab that lets users upload medical records and connect wellness apps like Apple Health and MyFitnessPal. They reported that 230 million users were already asking health questions each week. That is not a beta test. That is a population.
Within days, Anthropic countered with Claude for Healthcare, a suite of tools targeting providers, payers, and health systems -- announced at JPMorgan, because in healthcare, you launch where the money is.
What is worth noticing: neither company is building a "medical AI." They are building general models and pointing them at healthcare. This distinction matters. The question is not whether these tools are smart enough. It is whether the guardrails are strong enough, and whether physicians will have a seat at the table in designing how these systems are deployed -- or whether we will simply receive them, pre-configured, from vendors in Cupertino and San Francisco.
(If you want to think more deeply about that, read the Nature Medicine piece we just added to BeamPath: Physicians as Context Engineers. It argues, convincingly, that the central professional task of the physician of the future is not using AI. It is designing the context in which AI operates.)
The Model Carousel Keeps Spinning
For those keeping score, here is what dropped in the past two weeks:
Anthropic released Claude Sonnet 4.6 (February 17). Opus-class performance at Sonnet pricing. One million token context window in beta. For those of us using Claude daily, the jump is noticeable -- it follows instructions more precisely and handles longer documents without losing the thread. The pace of improvement is genuinely startling; Sonnet 4.5 came out just weeks ago.
Google upgraded Gemini 3 Deep Think (February 12), their reasoning-focused model designed for hard science, math, and engineering problems. It is not a general chatbot update -- it is specifically for problems where you want the model to slow down and think carefully.
xAI launched Grok 4.20 in beta (February 17), with a multi-agent architecture where four sub-agents tackle a problem independently before synthesizing a response. Musk being Musk, the benchmarks are promised "next month."
DeepSeek V4 is expected any day now, reportedly with one trillion parameters and a focus on coding tasks. Whether this reshapes the competitive landscape or simply adds another row to the benchmark tables remains to be seen.
If this feels like a lot, it is. The release cadence is now measured in weeks, not quarters. For clinicians, the practical takeaway: the tools you evaluated in December are already outdated. The ones you ignore today will be dramatically better by summer.
A Book Worth Your Time
If you read one thing about AI this month that is not a model announcement, make it the work of Arvind Narayanan and Sayash Kapoor. Their Substack, AI as Normal Technology, and their book AI Snake Oil, make a deceptively simple argument: AI is transformative technology, but it is not magic, and it is not superintelligence. It is a tool. A powerful, flawed, rapidly improving tool -- the kind of thing that should be evaluated with the same rigor we apply to a new drug or a new surgical technique.
Everyone in this space has an agenda. The companies selling AI want you to believe it is indispensable. The skeptics want you to believe it is dangerous. The researchers want you to believe their benchmark is the one that matters. Narayanan and Kapoor cut through all of it. We should educate ourselves and help build what works for us. We can now, with AI.
OpenClaw: The AI Agent That Actually Does Things
If you have been anywhere near tech Twitter in the past few weeks, you have probably seen OpenClaw -- the open-source AI agent that went viral in late January and has since been covered by CNBC, IBM, and seemingly every AI newsletter. What makes it different from the chatbot you already have open in another tab is that it actually does things: it runs on your machine, connects to your email, calendar, browser, and files, and carries out tasks autonomously. I have been running it for the past three to four weeks on its own Mac mini -- it set up its own Gmail, GitHub, and development environment -- and I am using it to help keep BeamPath updated, curate resources, and build personalized research newsletters that scan journals and synthesize evidence tailored to exactly what I care about. More detailed write-ups on that workflow are coming. The setup is not trivial, but if the trajectory holds, the gap between "AI assistant" and "AI colleague" is closing faster than most people realise.