This week focused on clarity: clearer defaults, better docs, and more robust scrapers and backups. Small UX and infra fixes made the system safer and easier to run.
Crisp posts and fresher prompt guidance make big ideas and small hacks easier to find and reuse.
- New long essay on AI bottlenecks. Added a long-form post explaining how AI bottlenecks shift across coding, agents, and docs (adb08b0). 05 Jun 2026. This helps readers judge progress and evaluate model claims practically. Takeaway: frame technical advances by the bottlenecks they move, not just by headlines. (Yes, another AI essay — but it's useful.)
- Short notes and travel anecdote. Published a compact prompt-vs-code note and an Indigo flight anecdote (adb08b0). 05 Jun 2026. These posts show practical wins from tiny experiments and human observation. Takeaway: small, concrete stories teach more than abstract claims.
- Playbook for skills when juniors vanish. Added "No juniors, no experts?" and captured changing AI opinions (c819da4). 05 Jun 2026. It gives a clear framework: switch, enforce, or level-up skill strategies. Takeaway: choose training based on the cost of being wrong.
- Better prompts and metadata for discovery. Expanded prompts, tools lists, and enriched post metadata for SEO and clarity (fb21607, e8d79af). 02–03 Jun 2026. This improves searchability and makes prompts usable. Takeaway: small metadata changes multiply discoverability.
Make secure static hosting simple to run and safe by default.
- Initial secure static pages bundle. Added a Docker+nginx+Pomerium app with scripts and README (f54222e). 05 Jun 2026. You can run public mode immediately and switch to Google-auth protected mode later. Takeaway: ship a safe, minimal default that scales to secure mode.
- Simplify secure gateway and nginx fronting. Reworked nginx configs, entrypoint, and compile script for clearer ports and binds (549e44c). 05 Jun 2026. The container now fronts Pomerium with nginx and binds services to localhost. Takeaway: explicit local binds reduce surprising exposure.
- Document Pomerium databroker files and logout UI. Added docs explaining data/databroker runtime files and a logout link in the sample site (32c2012). 05 Jun 2026. This explains what RocksDB-like files mean and how to reset them safely. Takeaway: document runtime artifacts so operators don't panic.
(Yes, you really needed a README entry about RocksDB files. Future-you will thank present-you.)
Short-form episode notes turn chat threads into listenable, searchable artifacts.
- Weekly podcast episode added. Composed the 31 May 2026 digest transcript for the group podcast (78c9a56). 31 May 2026. It summarizes cost, harness, and tooling conversations for easy reuse. Takeaway: distilling chat threads into episodes focuses product and cost debates.
Keep notes and trend signals fresh so small discoveries stay useful.
- Add May notes and refresh trends. Updated core-concepts, recent TIL entries, and refreshed trending repos TSV (7f30241). 31 May 2026. This preserves ephemeral insights and captures ecosystem signals. Takeaway: weekly housekeeping keeps knowledge leverageable.
More complete activity logs and safer backups make daily automation trustworthy.
- Personal-email & activities integration. Added personal-email support and patched activity reports to include personal mail (d9fde4f). 06 Jun 2026. This closes gaps between work and personal logs. Takeaway: include all accounts to avoid blind spots.
- More reliable backups and transcription order. Improved WhatsApp incremental scans, transcribe ordering (newest-first), and file mtimes (d9fde4f). 06 Jun 2026. Reruns miss less and are deterministic. Takeaway: deterministic ordering makes retries safe.
- New helper commands and skills; docs & dev updates. Added git-size, backuplinkedin, livetranscribe, agent skills, and dev container tweaks (d9fde4f). 06 Jun 2026. These shorten common flows and integrate skills with agents. Takeaway: small CLI helpers multiply productivity.
(Yes, your activity logger now knows your other email. No excuses for missing logs.)
Small publishing UX lifts for sharing and verification.
- Add SVG download button to statnostics. Added a download action and adjusted styles (cdfa61e). 04 Jun 2026. Users can save visuals for verification or reuse. Takeaway: make verification frictionless.
- Expanded Hack of the Day content. Added many new cards and tips to the gallery (ef94df5). 04 Jun 2026. More bite-sized safety tips for readers. Takeaway: small, repeated tips scale audience value.
Run rigorous, reproducible experiments to compare models and behaviours.
- LLM vs Stockfish chess-evaluation harness. Added an evaluation CLI, dataset builders, and tests (2a9dd9). 05 Jun 2026. It runs LLMs against Elo-limited Stockfish and logs structured artifacts. Takeaway: reproducible benchmarks reveal practical model tradeoffs.
- Update games summary after runs. Recorded more games and usage stats (e4e2c6). 05 Jun 2026. Tracking tokens and outcomes helps cost-performance analysis. Takeaway: log cost alongside accuracy to guide deployment.
User-facing tools got sturdier parsing and richer color controls.
- Colortable: Markdown, currency, reverse scale, and tests. Added Markdown parsing, currency handling, reverse-scale option, grouped D3 scales, and tests (8e670d7). 01 Jun 2026. This widens supported inputs and stabilizes behavior. Takeaway: add tests before UI changes.
- LinkedIn invite scraper robustness. Fixed description extraction and common-org heuristics; added tests (7f83d4c). 31 May 2026. Scraping now skips noise like "wants to connect." Takeaway: prefer structural cues over naive text slicing.
(Yes, the scraper now ignores the LinkedIn theater and finds real signals.)
Prune configurations to match supported generation endpoints.
- Trim model list to current backends. Removed legacy image model entries and kept current picks (74a9e9). 03 Jun 2026. This avoids broken generation pipelines. Takeaway: prune configs to reduce surprises.
- Default-to-sane: Safe defaults and clear docs prevent operator mistakes.
- Test-first UI: Add tests before feature work for stable front-end behavior.
- Log cost with accuracy: Track tokens and latency as first-class experiment outputs.
- Small content wins: Short posts and metadata improve discoverability fast.
- Document runtime state: Explain generated runtime files and reset steps.
- Include full context: Personal accounts and deterministic ordering reduce blind spots.
- Blog: Add short TL;DRs and social excerpts for new essays to increase reach.
- Auth-pages: Add a healthcheck and CI that validates secure/public startup and databroker recreation.
- Scripts: Add unit tests in CI for activities.py patch flow and backupwhatsapp edge cases.
- Tools: Deploy colortable and linkedinscraper updates to the live tools site; add an E2E smoke test.
- Research: Publish a reproducible run manifest and sample PGNs from chess-evaluation.
- Journalists: Export new cards as CSV to enable batch translations and syndication.
- llmartstyle: Run a fresh image generation sweep with the trimmed models and archive removed-model outputs.
If you want, I can draft the small CI job examples for auth-pages and scripts next.