A Slightly Late Hello
We have been quiet for a few weeks. This was partly intentional and partly the standard summer phenomenon where every sensible plan is overtaken by travel, deadlines, and the peculiar conviction that the World Cup only needs checking once more before bed.
The tournament has also become one of those things that seems to be everywhere at once. I have noticed World Cup information appearing more readily inside Gemini and ChatGPT, alongside the usual score apps, group chats, and half-informed opinions from people who have suddenly become experts on the offside rule. It is a small change, but a fun one. The AI apps are increasingly becoming places people go for whatever is happening right now, not just for writing an email or asking a complicated question.
While all of that was going on, the AI release cycle kept moving at its usual slightly unreasonable pace.
The Models Did Not Take a Summer Break
OpenAI released GPT-5.6, which is less interesting to me as a leaderboard event than as another sign of where these products are going. The emphasis is on longer work: tool calling, carrying context across turns, caching, multi-agent tasks, and more explicit controls over what a system can do without asking first. That is a lot more useful than a model that simply sounds impressive for three paragraphs.
Anthropic had its own brief bit of drama. Fable 5 launched in June, was suspended a few days later, and was redeployed on July 1. We are still experimenting with it, but so far it has a fairly heavy set of guidelines and boundaries that seem to trigger on the slightest version of several use cases we have tried. For now, that has made it less useful than it might otherwise be. I am not especially interested in turning every product interruption into a grand lesson, but it is worth remembering that the model is only one part of what we use. Access, reliability, the surrounding tools, and the guardrails matter too. Anyone who has had a workflow quietly break after an update will recognize the feeling.
The More Interesting Story Was in Utah
The update I found most worth reading came from Utah's pilot with Doctronic. It is refreshingly unflashy. The system is being tested only for renewing a limited group of already prescribed medications. It cannot issue new prescriptions, change a treatment plan, or handle controlled substances. A licensed clinician still has to authorize every request.
That narrow scope makes the early outcomes report genuinely useful. In 72% of reviewed cases, the system recommended a renewal. The remaining 28% were escalated, often because someone needed laboratory testing, had a complication, or had gone too long without a visit. Physicians agreed with 91% of the renewal recommendations on first review. They thought 69% of the escalations were appropriate, while 31% were overly cautious.
These are early data, not a victory lap. The pilot is small, its initial reviews were performed by company physicians, and Utah has started an independent review. The state says it does not yet have robust evidence of benefit. Still, I like the basic shape of the experiment. Give a system a specific job. Be clear about what it cannot do. Keep the clinician in charge. Then show us where the system and the clinician disagree.
That is a far more useful way to learn than announcing that a model passed another exam it will never actually take.
Back to BeamPath
We have a fair amount to catch up on, and there is plenty worth covering beyond the release announcements. Over the coming weeks, I want Beam Notes to stay useful in the way BeamPath was meant to be useful: a place to sort through what is new, what is actually interesting, and what is mostly noise wearing a very expensive product launch.
For now, enjoy the final week of the World Cup. And if you have found an AI tool, paper, or clinical use case that changed how you are working, send it over. The notebook is still growing.
- Ramez