self.md radar — 2026-09-09
the rough mark gets its own input field
A sketch box has appeared beside the image prompt. A memory benchmark has started charging for the history it replays. A Mac bot app gives each browser worker a separate Linux screen.
1. OpenAI lets you draw the instruction first
ChatGPT Images 2.5 adds Sketch: draw directly in ChatGPT, then use that drawing as reference for the image you actually want. The release also divides the API into two named models. GPT-Image-2.5 Flare is positioned as the default, with OpenAI claiming 50% lower latency than GPT-Image-2; Sunburst takes longer in exchange for more control over detailed edits.
That is a useful change even before the quality claims settle. A crooked line, a blocked-out layout or a badly drawn arrow can enter the job as an object, rather than being translated into a text instruction.
reading: OpenAI’s Images 2.5 announcement
2. Agent memory gets a cost column
The MERIT paper asks a more irritating question than “can the agent remember the chat?”: does memory improve a tool-using task enough to justify what it costs? Its experiments report that swapping memory implementations can move task success by up to 60 points. The best condition in each domain returned 2.7–3.9 times the marginal utility per dollar; full replay never won on economics.
The paper does not establish a best memory design for every agent. It does expose a common weak default: retaining every conversation without testing whether the retained material changes the task enough to pay for itself.
reading: “When Does Memory Help?”
3. A persistent bot comes with a separate desktop
Routi Bot
is a Mac app for bots that stay around: each has a personality, chosen model and, for browser work, its own Linux desktop in Docker. The README calls it one screen per bot. The setup supports Claude or ChatGPT subscriptions or an API key; its uninstall notes say keys are removed from Keychain, while the bots and their conversations remain in ~/.routi.
That retention detail matters as much as the container. A bot that lives on after its window closes needs an inspectable working surface and an explicit cleanup path. Docker supplies a boundary, not a cleanup decision.
reading: Routi Bot repository
left on the table
OpenAI’s Navier–Stokes announcement is too large to wave through as a Radar item on its own press release. The company has posted a paper and Lean formalization; the mathematical verdict belongs to people who can actually audit them.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Agent memory systems — what agents should remember, what belongs in logs, and how to avoid memory sludge
- Claude Code pricing — plan choice, cost rails, and when usage gets weird