Crosstalk Open issue

Discoveries — draft findings for the quarterly retrospective

1. 01

   Labs is the magazine home again: quarters as issues, departments as lanes, twice-weekly filing into the open issue — with the newsletter as delivery, not a second archive.

   ↳ Introducing Labs · Jul 16

2. 02

   The field converged from every direction on one claim: the model is no longer the bottleneck — the harness, the context, and your attention are. Even the model vendors said it from their own keynotes.

   ↳ The World's Fair Report · Jul 15

3. 03

   Token costs reversed their multi-year decline mid-2026 — +29% intelligence-adjusted, +76% raw. If it holds, efficiency stops being optional and the token budget becomes a first-class design constraint.

   ↳ The World's Fair Report · Jul 15

4. 04

   Boring file architectures won retrieval: memory systems shipping with no vector database at all, embeddings reframed as cached compute, grep as the baseline.

   ↳ The World's Fair Report · Jul 15

5. 05

   Implementation got cheap, so the remaining work is getting the agent pointed at what you actually want — grill, spec, tickets with blocking edges — before anyone opens a PR to discover the plan.

   ↳ mattpocock/skills v1.1 · Jul 24

6. 06

   MCP grew a UI surface: the 2026-07-28 revision moved interactive widgets, tasks, and auth into negotiated extensions, with a stateless core so the USB port for agents can scale.

   ↳ Agents Get a UI · Jul 30

7. 07

   With A2A at v1.0 and MCP growing tasks and extensions, the protocols nearly rhyme by feature list — the real boundary is stance: MCP is a transparent workbench you compose, A2A is an opaque peer you delegate to. Choose the steps, or name the outcome.

   ↳ Where A2A Begins and Ends · Aug 4

8. 08

   Speed is a usability gate on top of quality ranking, not a substitute for it. whichllm's --speed fast floor (30 tok/s) will drop a higher-scoring quant before it will recommend molasses.

   ↳ whichllm: pick a local model that is actually fast · Aug 17

9. 09

   The phone-only project now has a front door: Grok Bot signs in with your Cursor account and hands repo work to a Cursor Cloud Agent, so the switchboard is something you text rather than wire. It only works if the environment was set up on Cursor first.

   ↳ Never Touch the Laptop · Sep 3

10. 10

    Jev is a classifier interface inside the model: you ask a fixed menu of judgments instead of continuing, completing, or calling tools. DeepSeek can return the same typed answers; speed and constraints are the difference — Jev’s p50 was 400 ms against 2,377 ms, and one wrong answer still came back at 0.98.

    ↳ Jev: Judging, Not Calling Tools · Sep 20