Today’s radar brought the announcement of a 501-billion-parameter open-weight model, a search API for agents, and an AI case with a reputation problem. But what stayed with me was a study of a math tutor. I’ll start there.

People who work in product learn an uncomfortable lesson early: shipping a feature is the easy part. Getting someone to use it for real, every day, is harder.

I was reminded of that while reading a study about Khanmigo, Khan Academy’s AI tutor. It was a two-year randomized trial across 18 Tennessee middle schools. The tutor was configured to coach students rather than hand them the answer, and math achievement improved measurably.

Another number caught my attention: the median student used the tool on only one-third of the planned practice days.

In other words, in this case, the AI delivered results and access was available, but that still wasn’t enough to create a habit. The bottleneck wasn’t the model’s capability; it was turning usefulness into repeat use.

I actually find that encouraging. It shows that, in this case, the technology can already produce results. The next gains come from product work: onboarding, showing up at the right moment, fitting into a person’s existing workflow, and giving them a clear reason to come back tomorrow.

To me, this is central to AI product adoption: buying a license doesn’t finish the project. It starts the work that needs product people watching usage data and adjusting the experience until it becomes a routine. That work belongs in AI product management.

This applies to schools, banks, and any company putting an AI assistant in its team’s hands. The model may work, but the experience still has to earn a place in people’s day-to-day work.

If you’d like to read the full study, here’s the link: the Khanmigo study.

The rest of the radar

  • Beam, Reflection’s announced open-weight model — the 501-billion-parameter model adds to the build-versus-buy conversation. Reflection says it plans to release the weights under Apache 2.0 later this month; if that happens, teams will have another option and closed API pricing could face more pressure. Read the announcement.
  • Cloudflare Web Search API (beta) — makes it simpler to give agents real-time information without integrating each search provider separately. See the changelog.
  • Claude Opus 5.5 agents and magnetic semiconductors — a long, reproducible research effort carried out with agents, with calculations, code, and limitations published. It is a useful reference for operating AI agents in production in specialized domains. Read the study.
  • ChatGPT and fake New Yorker signatures — a report highlights the reputational risk when generated images carry real cartoonists’ signatures, bringing guardrails, attribution, and creators’ identities into the product discussion. Read the report.

That’s it for today. Tomorrow brings more, and the radar will keep filtering what deserves your attention.