AI Daily · 2026-08-10

(TL;DR unavailable)

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] GitHub Models is now retired

GitHub Models has been fully retired; Simon Willison only noticed when his GitHub Actions build failed with a message about a scheduled retirement brownout that had already completed. The service previously offered a model playground and a unified API across LLM providers, with the key advantage that code running in Actions could use the existing GitHub API key for inference. No reason was given for the shutdown, though the author speculates that coding agent patterns made offering free or subsidized tokens prohibitively expensive. Willison has since migrated his workflow to an OpenAI API key and now generates summaries using GPT-5.6 Luna.

Research

⭐⭐ [Research] SQLite compressed text-history prototypes

Simon Willison prototypes two compressed text-history stores in SQLite: `WholeBlobHistoryStore` compresses the entire history as a single zstd-compressed JSON array per edit, while `ChunkedHistoryStore` splits the history into sealed compressed chunks for better scalability with long histories. Both preserve prior text and timestamps, skip unchanged replacements, and serialise writers with `BEGIN IMMEDIATE`. With help from GPT‑5.6 Sol Pro, an experiment shows 20.4MB of raw revision text from 1,000 simulated edits compresses to just 80.3KB. The chunked approach avoids the cost of decompressing and recompressing the whole array, capping each chunk at 128 revisions or 3MB of uncompressed JSON. The post offers a lightweight, data-backed approach for storing revision histories in relational databases.

Other

⭐⭐ [Other] Quoting Claude Opus 5 system prompt

Simon Willison has quoted the system prompt of Claude Opus 5, which provides the model with background on the temporary suspension and restoration of access to its Fable 5 and Mythos 5 versions due to U.S. export controls, instructing it to confirm the event matter-of-factly, avoid denial, and refer to the linked statement for further details. This reveals an internal mechanism for handling the model’s own short-lived unavailability history.