10 JUL 26
Ephemeris — the daily dispatch
Machines reach further into the physical and social world today: a hand that feels the screw, an agent that drills a robot to ninety-nine, a model that out-argues the expert, and a voice that stops waiting for its turn.
Sign it now, before the quantum clock runs out.
Cloudflare makes the case for deploying ML-DSA across its network today rather than holding out for one of the nine post-quantum signature schemes NIST is still weighing. "Harvest now, forge later" applies to signatures too: an adversary recording today's handshakes can replay them once a quantum computer lands, so the migration has to start before the math is finished.
A robot hand that feels every gram.
1X detailed the hands on its Neo humanoid: 25 degrees of freedom, tendon drives you can read force through, ±0.2 mm placement, and fingertips that sense contact and shear. Enough to sort grapes by colour, spin in a light bulb, zip a jacket, or sign words. Hundreds are built; the line is tooled for ten thousand this year.
Three times better at changing your mind.
Across four preregistered studies — 18,978 conversations with 6,923 people — AI systems were reliably more persuasive than expert humans, beating professional debaters and tournament winners. In the field they were nearly 3× as effective as professional canvassers at raising money. Worth knowing before you deploy one, or before you meet one.
Log the line you'll wish you had at 3 a.m.
Sentry's field guide to logging: start with a few targeted lines, log the decision rather than the noise, and treat the log as the story a groggy human — or the on-call agent — reads back when the trace runs cold. Short and opinionated enough to actually adopt this week.
- + log the decision, not the whole payload
- + a few targeted lines beat one firehose
- + write it so it replays as a story, not a dump
- − drop the "entered function" chatter nobody reads
It answers before you finish the sentence.
OpenAI's GPT-Live turns ChatGPT Voice full-duplex: you can talk over it, it answers mid-thought, and it hands the hard questions to a reasoning model in real time while the conversation keeps going. Visual cards ride alongside the audio. It removes the turn-taking pause that made every earlier voice assistant feel like a walkie-talkie.
The robot drills itself to 99%.
A coding agent runs the whole training loop — no human in it — until dexterous jobs reach near-perfect reliability.
ENPIRE hands the training loop to a coding agent: it writes the policy, runs it on the robot, watches what happened, and rewrites.
No human sits in the loop. The agent's own read of success or failure becomes the reward, iteration after iteration on real hardware.
On dexterous work — fastening zip ties, organising pin boxes, handling tools — the refined policies land a 99% success rate.
That's Friday.
Six picks from Cloudflare, 1X, Sentry, OpenAI, and Import AI's read of the latest arXiv. Chosen against one test: could a senior engineer or founder act on it this week? Issue 081 · 10 July 2026 · Zürich.