Do AI agents measure, or remember? With them I re-measured three classic results from the original records, Tycho Brahe to GW250114, with the answer withheld. The paper, three films and every failure along the way are in here. https://t.co/uVx5MYLh5c
Six practices that catch what your agents miss. Each came from a miss while building showtime with a crew of agents: the rule, how to set it up, and what the miss taught. The interactive manual, edition 2: https://t.co/aamIhE5SXG https://t.co/H0BnUi2pfI
The GW250114 film: wave peaks measured in the strain. Hawking's area theorem says a merger can't shrink the total horizon area. All 40 analysis cells favour growth, but the headline fits' lowest probability, 0.86, is under the 0.90 fixed in advance, so the claim was withheld.
Do AI agents measure, or remember? With them I re-measured three classic results from the original records, Tycho Brahe to GW250114, with the answer withheld. The paper, three films and every failure along the way are in here. https://t.co/uVx5MYLh5c
@konstructivizm The UV comes from the core’s heat: the hotter a body, the shorter the wavelength where its glow peaks. From 1,000 K to 10,000 K the peak moves from 2,900 nm to 290 nm, into the ultraviolet. Drag the temperature yourself: https://t.co/u7NL1DyJMh
16,173 logged decisions, $3.08. Limits: our labels (mostly AI-made, spot-checked), rules dated by file time only, one week, one vendor.
AI agents built and ran it; I set the question and made the call.
Plan, rules, data: https://t.co/1m6BzpDZIw
Blog: https://t.co/bc1QUZeELV
TypeSafe released Jev 1.14 on 14 Oct and advised developers to favour their own evals. So before switching, we tested it on our own labelled work, with the decision rules written before the first call.
We are staying on 1.13 for text. What we learned 🧵
Images: a caption over a face, all 102 confident answers right; 1.14's boxes (after a presence check) found all 80 charts. Neither clears our bar after adjusting. "Are chart labels readable?" failed: 0 of 24 right at 8-12 px on 1.13.
Experimental per vendor; synthetic frames.
It runs in the browser, free, nothing to install, and you don't need any physics to start: https://t.co/cmMEocuwrV
We built it one stop at a time, a human and an agent working together.
How Physics Works is a history of physics you can play with: 50 stops from Thales to the 2015 detection of gravitational waves, each with a simulation.
At the first stop you measure a pyramid by its shadow, like Thales. Near the end you hear two black holes merge.
Play them live: https://t.co/q7OZ1GgdIE
Release notes: https://t.co/XbPlruZHDd
showtime is open source and runs on your machine: https://t.co/sOvIKSBhut
Clips: Erik Pålsson / Sounds of Changes (CC BY 3.0); Canada Science and Technology Museum (CC BY 3.0); LGR (CC BY 3.0); KasraR (public domain); Elenaf (CC BY-SA 3.0); Lukáš Perný (CC BY-SA 4.0), via Wikimedia Commons. The teaser (post 1) is CC BY-SA 4.0.
/showtime x-ray this clip of my typewriter, drawn on the machine.
showtime finds every keystroke in the clip's sound and draws the inside moving on the exact frame of each one. Then it reads the final file back: 19 of 19 on their frame, audio 0.0 ms off. Draw it four frames late and it fails.
X-ray, new in showtime 0.4.4 🧵
How X-ray checks itself: a tiny mark in one corner lights only on frames where a sound happened. After the render, showtime reads that corner on every frame of the final file and lines it up with the clip's sound. Drawn four frames late: 0 of 19, FAIL.
What should I X-ray next?
@mattpocockuk One fix we built in: make the output prove itself. Every X-ray film carries a sync mark that showtime reads back from the final file. Drawn four frames late, it fails: 0 of 19.
showtime 0.4.3 is out.
Living diagrams: a system drawn as an isometric figure that builds itself, flows and packets moving through it, the camera choosing what to look at. And the same scene exports as a live figure people can poke: hover a part, pick a step, turn a knob.