Daily AI Zine
Wednesday, August 12, 2026
Issue No. 027 · Tokyo · full edition

✦ Today's Big Thing

Copilot gets memory while GitHub makes AI spend easier to see

The practical shift today is less about a new frontier model and more about stateful coding assistants, local-model options, and model-level cost visibility.

4 min read · 7 sections

In Brief
  1. GitHub Copilot for JetBrains now has persistent memory, local Ollama access, enterprise controls, and MCP reliability fixes.
  2. MAI-Code-1.1-Flash is rolling into GitHub Copilot with native vision support for image understanding.
  3. GitHub’s AI usage report now breaks token spend down by model, including input, output, and cache tokens.
  4. Runway’s Gen-4.5 claim is worth a controlled creative test, not a pipeline switch.

Today's Big Thing

The one thing that matters

Test it

Copilot memory makes coding assistants more stateful

GitHub Copilot for JetBrains now has persistent memory, local model access through Ollama, more enterprise controls, everyday chat workflow improvements, and reliability fixes across MCP servers. This is not a broad frontier-model moment, but it is a useful shift: IDE agents are moving from one-off chat helpers toward remembered, governed work partners that can also touch local models. The immediate value is not JetBrains itself. It is the operating pattern. Test whether persistent memory improves repeat coding tasks without letting old assumptions leak into new work. Use a contained repo, define what Copilot is allowed to remember, and compare its next session against a clean-session baseline.

T1GitHub Changelog08/11

My AI Ecosystem

Your actual stack

Test it

Copilot adds a smaller vision-capable coding model

MAI-Code-1.1-Flash, Microsoft’s latest small-tier coding model, is rolling out in GitHub Copilot with native vision support for image understanding and reported coding-quality improvements. The previous MAI-Code-1-Flash is scheduled for deprecation on September 10, 2026. Test it on screenshot-to-fix tasks and simple UI bug reports before trusting it for larger code changes.

T1GitHub Changelog08/11T1GitHub Changelog08/11
Use it

GitHub usage reports now show model-level token spend

GitHub’s AI usage report now shows a per-model token breakdown behind AI credits, including input, output, and cache tokens. That turns Copilot spend from a blended total into something you can inspect. Use it to spot which models are eating budget, which workflows benefit from caching, and where a cheaper model can handle routine coding tasks.

T1GitHub Changelog08/11
Test it

Hot signal: Runway’s Gen-4.5 is a creative-tool watch item

Runway is presenting Gen-4.5 as a video model focused on motion quality, prompt adherence, and visual fidelity. The available summary is thin, so do not treat this as a production switch. Treat it as a creative-tool signal. Run one controlled storyboard beat through the current video stack and Gen-4.5, then compare movement consistency and prompt following.

T1Runway Research

CoWork Corner

Claude CoWork, day to day

CoWork: keep MCP installs boring and named

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows.

No meaningful CoWork product change appeared in today’s candidate set. The useful revisit is connector hygiene. Claude Desktop Extensions have documented one-click MCP server installation, so keep each installed connector tied to one named project job and one owner note. If a connector cannot explain the job it performs, leave it out of CoWork.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Use the Circles case as an OpenAI revenue workflow template

OpenAI reports that Circles uses the OpenAI API and Codex for AI-native telco experiences, with higher ARPU, lower churn, and better development efficiency. Adrian should care because Grey Group OS can copy the pattern without copying the telco business: connect customer personalization, support follow-up, and implementation work in one measured loop. Today, pick one Goodsense or SET customer journey and define the one signal that should trigger a GPT-assisted follow-up.

T1OpenAI News08/03
Use it

Give Codex one modernization queue, not a blank repo

OpenAI’s field report says scientists are using AI coding agents to modernize scientific computing and accelerate software development. Adrian should care because Grey Group OS and Daily Ops probably have older scripts, templates, or repo habits that slow delivery. Do not ask Codex to redesign everything. Today, choose one brittle workflow and have Codex inventory dependencies, identify the smallest safe refactor, and open a reviewable change.

T1OpenAI News07/28

Small Money Systems

Small, repeatable, real

Use it

Small system: Copilot spend tune-up for AI-heavy teams

System: build a one-page Copilot credit usage tune-up using GitHub’s new per-model token breakdown. Customer: small product teams, production companies, or founder-led operators already paying for Copilot. Offer: they receive a 30-minute review that identifies expensive models, cache usage, and simple routing rules. Price: ¥25,000 for the first review. Existing assets: Grey Group OS, GitHub workflow habits, Claude Code notes, and Adrian’s AI-ops judgment. AI workflow: Claude Code turns exported usage notes into a short report, and Codex drafts any repo instruction changes. First action: open one GitHub usage report and capture the per-model token table. Repeatability: the same report template can run monthly. Effort: one hour. Expected value: recurring audit revenue and lower AI credit waste for the customer.

T1GitHub Changelog08/11

Build Next

Deployable now

Use it

Build a GitHub AI credit ledger

What to build: a small repo note or dashboard that records weekly GitHub AI usage by model, splitting input, output, and cache tokens. Why now: GitHub’s usage report now exposes the per-model token breakdown behind AI credits. Effort: one hour. Expected impact: catches model spend drift and makes it easier to choose when a cheaper coding model is enough. Dependencies: access to the GitHub AI usage report and one repo where Claude Code or Codex can store weekly snapshots.

T1GitHub Changelog08/11

Try This Today

One action, right now

Before the next coding session, open GitHub’s AI usage report and write down the top model by input, output, and cache tokens. Then decide which one task today should move to a cheaper or smaller model.