61%
Avg word accuracy
Ephemeris
Rewired.
Five rebuilds, all dated the same day — a build engine, a data warehouse, a voice channel, a robot's own policy, and a sentence pulled from brainwaves.
Dev Tools · Build Engine
The build engine keeps its memory between runs.
Turbopack in Next.js 16.3 writes a persistent on-disk cache, so a cold build reuses work it already did instead of starting over. It also evicts dev-server memory before it balloons and ships an experimental Rust port of the React Compiler. The slow half of the loop — recompiling what never changed — gets shorter.
PERSISTENT FILE CACHEbuilds reuse prior work
DEV MEMORY EVICTIONcaps long-session bloat
RUST REACT COMPILERexperimental
import.meta.globnow supported
Architecture · Postmortem
They tore out ClickHouse and rebuilt on DuckDB.
PostHog rebuilt its data warehouse on DuckDB, handing each customer a single-tenant instance and first-class support for dbt and the modern data stack. The post is the reasoning worth stealing: why an engine that began life embedded in analytics apps won the warehouse workload over a distributed column store.
From
ClickHouse — shared, distributed, hard to isolate per tenant
→
To
DuckDB — single-tenant instances, dbt-native, simpler ops
AI Tools · Realtime Voice
Voice becomes a primitive, not a project.
Vercel's AI Gateway now speaks. Realtime models from OpenAI and xAI sit behind one API, with generateSpeech() and transcribe() for one-shot work and a useRealtime hook for the browser. You get bidirectional, interruptible audio under the same auth, observability, and spend controls as the rest of the stack.
// one API, three calls
generateSpeech({ model, text }) → audio
transcribe({ model, audio }) → text
useRealtime({ model }) → live, interruptible
Science · Robotics
Robots that run their own experiments.
NVIDIA's ENPIRE puts physical robots in a self-experimentation loop: try a policy, measure the result, revise, repeat — improvement with no human in the tuning seat. Jack Clark's Import AI 463 pairs it with Tencent's ARGUS, software shepherding a 10,000-GPU training cluster, and an essay on what work is left for people.
ENPIRE
Robots improve their own policies by self-experiment
10,000GPUs
Tencent's ARGUS shepherds one training cluster
Science · Neurotech
A sentence, pulled from brainwaves.
Meta's Brain2Qwerty v2 decodes typed sentences from non-invasive MEG recordings at 61% word accuracy — 78% for the best participant, where most sentences come back with one wrong word or fewer. Scaling deep nets alongside a language model is narrowing the gap with surgical implants, and it needs no surgery to do it.
78%
Best participant
0
Implants
That's today.
Five rebuilds, every one dated 29 June, read the morning after. Each pick had to pass one test: could a senior engineer or founder act on it this week.
Sources today — Next.js · PostHog · Vercel · Jack Clark / Import AI · Meta AI (via @rvnikita_blog)
Issue 071 · 30 June 2026 · Zürich