Ephemeris — the daily dossier
Six things worth the morning.
Your Worker can cache itself now.
Cloudflare gave Workers a regionally tiered cache that sits directly in front of the entrypoint — no zone rules, no separate service. You turn it on with one line in Wrangler and steer it with the ordinary Cache-Control headers you already write. A lower tier lives in each data centre; an upper tier aggregates fills across the network, so the first request anywhere in the world warms the cache for everyone after.
# wrangler.jsonc "cache": { "enabled": true } # then, in your handler res.headers.set( "Cache-Control", "public, max-age=3600" ) # server-rendered → behaves like static
The model writes its own GPU kernels.
KernelBench-Mega hands an agent a whole model block and a three-hour clock, then asks it to fuse the block into one megakernel. On a Kimi-Linear decode task, Anthropic's Fable posted an 18.71× speedup over an optimised PyTorch baseline — and did it with a single cooperative kernel launch per token where rival systems still fire four to fourteen. Kernel-writing is close to the metal of AI R&D, so a model that is good at it is a model that can help build the next one.
launches / token 1 (others: 4–14)
hardware RTX PRO 6000 Blackwell
agent budget 3 h wall-clock / run
A workspace the model can name.
Anthropic's new paper argues that a language model keeps a small, privileged set of activations it can report on, steer, and reason with — a "J-space" the authors read out with a Jacobian lens. It behaves like the global workspace theorists describe in human cognition: verbalisable, controllable, engaged only for the hard problems. The practical hook is auditing — you can surface a model's unspoken deliberation, and even shape behaviour by writing a concept into that workspace instead of retraining.
Five properties of the workspace
- Verbal report
- Directed control
- Internal reasoning
- Flexible transfer
- Used only when needed
One model reads a catalogue of tens of billions.
JD.com published the architecture of Oxygen AIIC, the LLM-and-vision stack that keeps its product catalogue coherent across tens of billions of SKUs. It runs on Huawei Ascend NPUs, serves more than 700 million shoppers, and folds hundreds of millions of item updates into structured knowledge every day. This is the rare paper that shows what industrial LLM deployment actually costs and looks like once it is load-bearing rather than demo-ware.
A map of which jobs bend, break, or grow.
OpenAI's workforce analysis sorts European occupations by how AI is likely to touch them — the roles facing outright automation, the ones set to grow, and the larger middle whose day-to-day simply gets rewired. It is a planning document more than a forecast: useful for a founder deciding where to add headcount, where to add tooling, and where the two collide over the next two years.
Charisma is a skill you can drill.
As trust migrates from institutions to individuals, Jeff Bullas argues the scarce asset is a learnable one: the presence and warmth that make a person hard to scroll past. The essay's real value is that it resists the algorithm's pull toward generic — and turns "be charismatic" into three moves you can actually run.
- 1Show up on a rhythm, not in bursts. Pick one channel, post weekly, hold it for two years. Trust is built on being encountered again, not on a viral spike.
- 2Give the best stuff away first. Publish your frameworks unpaywalled; the more you hand over, the more you are trusted with.
- 3Own the revenue, don't rent it. Brand deals are ~70% of a typical creator's income — build a product or community that converts the trust you already hold.
That's today.
Six picks for Tuesday, 07 July 2026 — an edge-cache primitive, a model tuning its own kernels, a workspace inside the network, a catalogue at industrial scale, a workforce map, and a note on charisma as craft.