The dossier · six entries
Ephemeris.
Six places AI showed up this week the official ledgers don't track yet — a free image corpus, a million-token model, the SEO that does nothing, the watching your own AI now needs, a whale lane, and the jobs that bend first.
A corpus of images you're cleared to train on.
Stanford's Vision Lab released GPIC — roughly 28 trillion pixels across 100 million images, each captioned by a vision-language model, deduplicated and safety-filtered, and licensed for both research and commercial use. It's on Hugging Face today, which makes it one of the few large image sets you can fine-tune a generative model on without a lawyer in the room.
- ≈28T
- pixels in the corpus
- VLM
- caption on every image
- License
- research + commercial
- Host
- Hugging Face, now
The schema you added did nothing.
Ahrefs tracked 1,885 pages that bolted on JSON-LD schema and matched them against 4,000 that didn't. Across Google AI Overviews, AI Mode, and ChatGPT, the citation needle barely moved — and on Overviews it slipped backwards. AI-cited pages do tend to carry schema, but that's site quality showing through, not the markup doing work.
- 1,885 pages added schema; matched to 4,000 that didn't.
- Every AI surface landed at the edge of zero uplift.
- Cited pages carry schema 53% of the time — quality, not cause.
- The strongest signal they found was YouTube, not markup.
- +2.4%
- Google AI Mode
- +2.2%
- ChatGPT
- −4.6%
- AI Overviews
The biggest Qwen yet ships closed.
Qwen3.7-Max is Alibaba's new flagship: a one-million-token context window, output around 208 tokens a second, and seventh place on the Artificial Analysis intelligence index. It's API-only with the weights held back — a turn for a team known for open releases. If you need a long-context agent this week, it's a live option to benchmark against your stack.
Observability, the AI-shaped version.
An LLM feature fails in ways your APM never modelled: a prompt regresses overnight, a token bill balloons, a tool call returns confident garbage. PostHog's field guide walks the building blocks — tracing each call, tracking cost, capturing errors, wiring evals — and lays the main tools side by side so you choose before the incident rather than during it.
Wire these before launch
- Trace every model call, end to end
- Cost per call, per user, per feature
- Capture errors and bad tool calls
- Evals on the prompts that matter
PostHog · Langfuse · LangSmith · Helicone
A camera that reads a whale by its heat.
WhaleSpotter pairs thermal cameras with a neural network trained on hundreds of thousands of heat images to pick gray whales out of San Francisco Bay — the roughly two-degree warmth of a blowhole against cold water. It flags them up to seven kilometres out and hands each detection to a human before alerting nearby vessels, buying captains time to change course.
“…as far as seven kilometres away.”
- ≈2 °C
- blowhole vs. cold water
- 7 km
- early-warning range
- Human
- verifies before the alert
Which jobs bend first.
An analysis reframes automation risk around job "dimensionality" — how many interdependent tasks a role bundles. Low-dimensionality work, where tasks don't lean on one another, takes the brunt; high-dimensionality roles can see wages rise if demand for their output is elastic. It's a sharper lens than "will AI take my job" for deciding where to put a career or a team.
| Low dimensionality | few interlocking tasks → most exposed to automation |
| High dimensionality | many interlocking tasks → wages may rise if demand is elastic |
That's today.
Six picks from a saturated week — a Stanford image corpus, Alibaba's closed flagship, a marketing study that clears the air, the observability an AI feature earns, a thermal eye on the shipping lane, and a cleaner way to think about which jobs go first.