Head of Applied AI at Dext. Building https://xitter.cf/t.co/DANgmpZvSA | Quantum physics PhD from Oxford | Former Italian blindfolded Rubik’s cube record holder
London, England
Joined May 2015
- Tweets1.5K
- Following2.4K
- Followers2.1K
- Likes2.6K
Pinned Tweet
Dentist did a 3D X-ray of my jaw before a root canal and said I wouldn't be able to open the raw data, it needs specialized software. It's 800 DICOM files. Asked Claude Code to make me a viewer. Two prompts later... this is nicer than what he showed me on his screen.
Benchmarks are nice, so i gave pi real work instead: a Snake game, a JSON parser vs 167 tests, a feature in a real repo, checked by running the code!
Flash-Next, one DGX Spark, 3 runs each: @ashxhart TensorFold + @MiaAI_lab’s patches 17.1 min, vLLM 23.6, stock TensorFold 24.2, Flare 32.7
All 4 engines passed all 36 tasks. Flare’s mean includes one 47-min Snake run where it got stuck, its other runs were 17 and 25 min.
pi, thinking on, temperature 0.7, up to 3 fix rounds. Recipes: github.com/MiaAI-Lab/Qwen3.8…, github.com/sxuff/qwen38-flas…
Ran a proper bake-off for Qwen3.8 Flash-Next on one DGX Spark: 4 engines, 3 runs each!
Agent test, 30K context:
TensorFold + @MiaAI_lab’s patches 36.2 tok/s, @NVIDIAAI Flare 31.2,
@ashxhart 's stock TensorFold 30.2,
vLLM 26.4
Turns out what the model writes matters more: JSON 67–70 tok/s, prose 31–34, on every engine
Recipes: Flare by @filicroval github.com/sxuff/qwen38-flas…, @MiaAI_lab TensorFold patches github.com/MiaAI-Lab/Qwen3.8…
NVIDIA FP4-Local weights (stock TensorFold can’t read them, so it ran local-inference-lab NVFP4). One request at a time, greedy. Real agent tasks next!
1/ I’ve been using @claudeai to work on Linear A since August. Having read @PlasmaAI’s release of @IdaeanDaktyl’s proposed Semitic reading, I think it’s more a hypothesis than a decipherment. Some notes from testing their word-matching engine 🧵
Linear A, the script of Bronze Age Crete, has puzzled researchers for over a century.
Today, we’re releasing the code and data behind Plasma researcher @IdaeanDaktyl’s proposed decipherment. Explore the evidence and test the readings.
plasma.ai/research/linear-a
5/ None of the words the article highlights (IO Za 2 prayer, HT 31 vessels) gets a significant match; many match too easily to judge. SU-PU (“basin”) matches a root in their list exactly, but so do 947 of the engine’s 1,000 fake decoys of that shape. A method limit, not a knock.
i didn’t think this was possible… I got voice transcription running INSIDE the chip of my @EvenRealities G2 🤯
Fully local, in less than half a megabyte of memory, words on the lens 1-2 seconds after I say them. My phone isn’t even connected, nothing leaves the glasses!
Inspired by @wesbos photo-frame project, I got @OmarchyLinux running natively on a $50 digital photo frame.
Allwinner A33, 1 GB RAM, Mali-400 GP, Arch Linux ARM booted from microSD.
I had to build a custom SD boot path, adapt Omarchy for armv7, fix display/touch rotation, bring up the AIC8800 Wi‑Fi driver, and manually enable the UART Bluetooth controller for BlueZ.
Now it boots straight into Hyprland/Omarchy with touch, on-screen keyboard, Wi‑Fi, Bluetooth LE mouse/keyboard, and the original eMMC still untouched.
Has anyone else run Omarchy on one of these? @dhh
I’m I the first running a firmware update on a flight with Starlink? @elonmusk
Used Claude + my EvenRealities G2 to try the flash-suppression trick from @doristsao's new consciousness preprint, then add my own twist!
The G2 has a display per eye: one eye gets Einstein, then Curie flashes into the other. Curie took over, like their volunteers reported… but after a moment i could pull Einstein back!
Then i flipped the two faces ~15 times a second, the rate their neurons switch at. Result: i just saw both, flickering, no winner. Not a replication (n=1, i knew what was coming), but crazy what you can build at home with AI now
1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness.
New preprint: biorxiv.org/content/10.64898… (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)
Can AI actually count? I tried Glimpse, the first model from @ElorianAI and @AndrewDai, a team that led Google’s Gemini and Apple’s MM-1, on six mosaics.
It counted every tile exactly, 345 for the owl. Gemini 3.8 Flash guessed low every time: 0/6.
It doesn’t eyeball it, it numbers every tile so you can check. That’s Elorian’s visual thinking.
I absolutely cracked the @EvenRealities G2!
I can now stream YouTube to them!!!
here's the lastest @mkbhd video:
Full resolution, ~15 frames a second, and the sound plays in sync in my AirPods!
Just managed to get Wispr Flow-style dictation running fully LOCALLY on a £50 photo frame running native @Omarchy!
I talk, and a speech model on the frame’s own 1GHz 32-bit chip with 1GB of RAM transcribes it and types it out, no cloud at all. Turns out the frame has no mic, so my Mac just sends the audio over and the frame does all the AI, just around 1s of latency!
Huge thanks to @wesbos, whose photo-frame-dashboard repo cracked open this exact frame and got me started, and to @dhh for Omarchy! Model is Moonshine tiny via sherpa-onnx, all on the frame’s CPU. Open source: github.com/pangoleen/ormarch…
