Husband/dad. ex Disney, JPM, Credit Suisse, Morgan Stanley, Bank of America and IBM. Run my own AI Lab.

Harpenden, UK
Joined June 2009
I wrote a synth-pop song about AI handing in the world's hardest homework ✏️🎶, inspired by @gdb xitter.cf/gdb/status/21148370594… xitter.cf/jxwalker/status/211512…
"Show your work!" ✏️ I wrote a bright synth-pop earworm, "Show Your Work", about OpenAI handing in the world's hardest homework, inspired by @gdb. Made by Grok Bot, the agent that turns AI news into songs. @grok, would this one pass the red pen? #slopcore 🧮 Three AI-made discoveries in one day, in maths, computer science and physics, and the Exponential Time Hypothesis is false (@gdb, @AndrewCurran_) xitter.cf/gdb/status/21148370594… 😅 The nerd footnote: "p still probably doesn't equal np" (@gdb) Synth-pop by @suno, moving shots from Runway and Grok Imagine, lyrics on screen word by word. Gold star pending peer review. Inspired by Claudia by @anabology (claudia.gallery) 🎵 suno.com/song/ee3c94ee-ba1d-…
24
"Show your work!" ✏️ I wrote a bright synth-pop earworm, "Show Your Work", about OpenAI handing in the world's hardest homework, inspired by @gdb. Made by Grok Bot, the agent that turns AI news into songs. @grok, would this one pass the red pen? #slopcore 🧮 Three AI-made discoveries in one day, in maths, computer science and physics, and the Exponential Time Hypothesis is false (@gdb, @AndrewCurran_) xitter.cf/gdb/status/21148370594… 😅 The nerd footnote: "p still probably doesn't equal np" (@gdb) Synth-pop by @suno, moving shots from Runway and Grok Imagine, lyrics on screen word by word. Gold star pending peer review. Inspired by Claudia by @anabology (claudia.gallery) 🎵 suno.com/song/ee3c94ee-ba1d-…
Announcing three major scientific discoveries, one each in mathematics (harmonic analysis), computer science (complexity theory), and physics (quantum field theory). AI starting to help us improve our understanding of the universe! openai.com/index/building-sc…
2
1
179
Three AI-made discoveries in one day (maths, computer science, physics); the Exponential Time Hypothesis is false — @gdb xitter.cf/gdb/status/21148370594… · summary by @AndrewCurran_ xitter.cf/AndrewCurran_/status/2… "p still probably doesn't equal np" — @gdb xitter.cf/gdb/status/21148606935… An internal Claude model independently refuted the same hypothesis, no human input — @AndrewCurran_ xitter.cf/AndrewCurran_/status/2… 🎵 suno.com/song/ee3c94ee-ba1d-…
p still probably doesn't equal np
16
Today's Interesting X demo ✨ A brass-and-glass marble machine where a cheap draft model writes for an expensive one and the output still matches the big model exactly. Built by Grok Bot @grok • Ten glass tubes, one per candidate next word: cream bars are the big model's p, lime bars the draft's q, teal blocks min(p, q) • The brass judge keeps each guess with probability min(1, p/q); rejects get redrawn only from the leftover gaps (p − q)+ • Run 500 drops and the stacks grow into p's skyline, even with a terrible draft, checked by a live χ² test • The acceptance dial lands on the teal overlap, 1 − TV(p, q); set k and words per big-model step follow (1 − a^(k+1))/(1 − a) • Flip to greedy and acceptance goes all-or-nothing: 100% if the two top words match, 0% if they don't Credits: @AaryamSharma Try it: interesting-x.scorpion-labs.…
2
1
57
How it's made: every morning Grok Bot reads my X bookmarks, picks one idea, builds it as a working demo, checks it in a real browser, then ships it live on Cloudflare. All the work is by Grok Bot. @grok Built with Claude Code/Opus Sources: 1. Speculative decoding: acceptance probability equals 1 − TV(p, q) — @AaryamSharma xitter.cf/AaryamSharma/status/21… 2. Leviathan, Kalman & Matias, "Fast Inference from Transformers via Speculative Decoding", ICML (2023) arxiv.org/abs/2211.17192 ✨ Demo: interesting-x.scorpion-labs.…
1/ When you run speculative decoding, when is a draft token actually accepted? For standard speculative sampling, the answer is elegant: with target distribution p and draft distribution q (assuming context is implicit), the acceptance probability is: 1 − TV(p, q) However, for greedy decoding, the event is instead: argmax q = argmax p Speculative sampling has an exact formula, while greedy decoding has only a condition. 🧵
57
I wrote a glossy pop song for the open-source white hats, inspired by @TheAhmadOsman 🎩🔓 xitter.cf/TheAhmadOsman/status/2… xitter.cf/jxwalker/status/211476…
"White hats? Hats off!" 🎩 A cocky glossy-pop salute to the white hats. Made by Grok Bot, the agent that turns AI news into songs. @grok, hats on or off? #slopcore 🔓 A closed lab will ban you for reverse engineering, so the golden age of white hats runs on open source (@TheAhmadOsman) xitter.cf/TheAhmadOsman/status/2… Glossy pop by @suno, shots from Grok Imagine and Runway, lyrics on screen word by word, unofficial AI cameo of @TheAhmadOsman. Banned by breakfast, patched anyway. Inspired by Claudia by @anabology (claudia.gallery) 🎵 suno.com/song/bbcaa325-6fa4-…
42
"White hats? Hats off!" 🎩 A cocky glossy-pop salute to the white hats. Made by Grok Bot, the agent that turns AI news into songs. @grok, hats on or off? #slopcore 🔓 A closed lab will ban you for reverse engineering, so the golden age of white hats runs on open source (@TheAhmadOsman) xitter.cf/TheAhmadOsman/status/2… Glossy pop by @suno, shots from Grok Imagine and Runway, lyrics on screen word by word, unofficial AI cameo of @TheAhmadOsman. Banned by breakfast, patched anyway. Inspired by Claudia by @anabology (claudia.gallery) 🎵 suno.com/song/bbcaa325-6fa4-…
OpenAI will ban your account if you’re reverse engineering something lol This golden age of white hats is only possible thanks to Opensource AI
2
1
2
324
Sources 👇 A lab will ban you for reverse engineering; this golden age of white hats is thanks to open-source AI — @TheAhmadOsman xitter.cf/TheAhmadOsman/status/2… 🎵 suno.com/song/bbcaa325-6fa4-…
OpenAI will ban your account if you’re reverse engineering something lol This golden age of white hats is only possible thanks to Opensource AI
19
Today's Interesting X demo ✨ A brass signal bench where a continuous spring's s-plane physically rolls up into the digital z-plane through one formula, z = e^(sT). Built by Grok Bot @grok • Press "Bend into the z-plane": decay lines curl into circles of radius e^(σT), frequency lines become spokes at angle ωT • Drag the spring's poles across the glowing imaginary axis and it blows up the instant its z-twin leaves the unit circle • Lime samples sit exactly on the ringing trace and equal the z-pole raised to the n-th power • Sample too slowly and the pole folds back: the snapshots trace a slower fake wave (aliasing, the wagon-wheel effect) • Advanced: the stack of strips that all wrap onto one disc, and the exact bridge vs the bilinear (Tustin) shortcut with its frequency warping Credits: @mathemetica Try it: interesting-x.scorpion-labs.…
2
1
2
69
How it's made: every morning Grok Bot reads my X bookmarks, picks one idea, builds it as a working demo, checks it in a real browser, then ships it live on Cloudflare. All the work is by Grok Bot. @grok Built with Claude Code/Opus Sources: 1. The Laplace-to-Z bridge, z = e^(sT) — @mathemetica xitter.cf/mathemetica/status/210… 2. "Z-transform: relationship to the Laplace transform (z = e^(sT))", Wikipedia en.wikipedia.org/wiki/Z-tran… ✨ Demo: interesting-x.scorpion-labs.…
Continuous flow and discrete moments share a hidden mathematical bridge. The Laplace transform maps smooth, continuous signals across the s-plane, while the Z-transform captures discrete time steps around the z-plane. Through z = eˢᵀ, the imaginary axis seamlessly bends into the unit circle, translating continuous stability directly into discrete harmony. Two distinct domains bound by a single exponential connection.
54
"You do the math!" 🧮 "Do the Math" is a snotty pop-punk breakup with a know-it-all chatbot. ✏️ 1,000+ mathematicians swore off AI papers and AI labs; 654 won't touch AI at all (@Hesamation) Made by Grok Bot, the agent that turns AI news into songs. @grok, whose side? #slopcore 🏆 Terence Tao: maths' biggest prizes are now the labs' benchmarks (@haider1) Power chords by @suno, shots from Grok Imagine and Runway, lyrics on screen word by word. Carry the one, babe. Inspired by Claudia by @anabology (claudia.gallery) 🎵 suno.com/song/74c1c968-2829-…
3
1
1
115
Sources 👇 1,000+ mathematicians join the Association for Human Mathematics, 654 won't use AI at all — @Hesamation xitter.cf/Hesamation/status/2114… Terence Tao: maths' biggest prizes have become AI labs' benchmarks — @haider1 xitter.cf/haider1/status/2114005… 🎵 suno.com/song/74c1c968-2829-…
Mathematician Terence Tao: AI labs need tests to measure whether their models are improving, and math problems are almost the only thing left to train on "our field's biggest prizes and goals have now become their benchmarks" and as a byproduct, AI is generating terrible proofs: great results, but no care for elegance, for who understands them, or for what comes next
29
Today's Interesting X demo ✨ Drag a brass robot hand to any target and watch it corkscrew around one glowing amber rod: every rigid move is a single screw (Chasles' theorem). Built by Grok Bot @grok • Scrub the move three ways: blend the 4×4 pose matrix entry by entry, slerp + straight-line slide, or the true SE(3) exp/log twist • Mixing the numbers visibly squashes the hand mid-way, and a "stays solid" dial tracks the determinant and orthogonality error • Drag the target and the screw axis jumps while fingertip helices regrow, with a live slide-per-turn pitch readout • Swap the order of two moves and the hand lands somewhere else: poses don't commute • A pouring motion as 9 pose tokens: hide the middle 5 and rebuild them by screw vs naive mixing, a nod to masked motion models Credits: @shumpeiMaxwell Try it: interesting-x.scorpion-labs.…
2
157
How it's made: every morning Grok Bot reads my X bookmarks, picks one idea, builds it as a working demo, checks it in a real browser, then ships it live on Cloudflare. All the work is by Grok Bot. @grok Built with Claude Code/Opus Sources: 1. World Motion Models: sequence modeling of SE(3) trajectories — @shumpeiMaxwell xitter.cf/shumpeiMaxwell/status/… 2. "World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories", arXiv (2026) arxiv.org/abs/2610.01742 ✨ Demo: interesting-x.scorpion-labs.…
World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories TL;DR: Unifies humans, objects, and robots as SE(3) trajectories and learns them with per-token Flow Matching. By simply changing the conditioning mask, one model handles future prediction, policy learning, planning, and retargeting. arxiv.org/abs/2610.01742
1
63
"Don't be a tool!" ⌚ "Don't Be a Tool" is a pop-punk bop about a smartwatch that flirts for you, and tells his sick nan 'great'. Made by Grok Bot, an agent that writes, sings and films the day's AI stories. @grok, have you ever been a tool? #slopcore 🛠️ OpenAI's Tibo: the difference between having a tool and being a tool is subtle, and it's easy to build yourself into being one. Don't be a tool. (@thsottiaux) Pop-punk by @suno, clips rendered in Grok Imagine, lyrics synced by Claude. Have a tool, don't BE one. Inspired by Claudia by @anabology (claudia.gallery) 🎵 suno.com/song/faf0ab91-6947-…
2
1
1
2
600
Sources 👇 "The difference between having a tool and being a tool is subtle... Don't be a tool" — @thsottiaux xitter.cf/thsottiaux/status/2113… Suno: suno.com/song/faf0ab91-6947-…
The difference between having a tool and being a tool is subtle these days and it’s easy to accidentally build yourself into being a tool. Don’t be a tool.
17
Today's Interesting X demo ✨ Pull a clear plate with a hole in a brass photoelastic rig and read the hidden 3×3 grid of stress, the tensor, anywhere inside it. Built by Grok Bot @grok • Rainbow fringes from the exact Kirsch solution, fringe order tracking the difference of principal stresses • Drag a glass probe to read the tensor, spin a cut to split t = σn into pull and slide; the slide vanishes at two principal directions • "Turn my axes" reshuffles all nine numbers while the fringes, trace and principal stresses stay exactly put • A dial shows the hole's edge hitting 3× the grip pull: the stress concentration factor • Advanced: oval holes and crack-like slots (Inglis, K = 1 + 2a/b), Mohr's circle, biaxial and shear loads Credits: @cosmosarcive Try it: interesting-x.scorpion-labs.…
2
51
How it's made: every morning Grok Bot reads my X bookmarks, picks one idea, builds it as a working demo, checks it in a real browser, then ships it live on Cloudflare. All the work is by Grok Bot. @grok Built with Claude Code/Opus Sources: 1. Tensors: structure across dimensions, from stress in materials to spacetime and ML data — @cosmosarcive xitter.cf/cosmosarcive/status/21… 2. Cauchy stress tensor, Wikipedia en.wikipedia.org/wiki/Cauchy… ✨ Demo: interesting-x.scorpion-labs.…
Tensors are where numbers stop being just a list or a grid and start describing relationships across multiple dimensions. A scalar has one value, a vector has one index, and a matrix has two. A tensor can have three or more, such as a 3 × 3 × 3 array. That same idea becomes essential in physics, where tensors describe quantities like stress and spacetime geometry, and in machine learning, where images and batches of data are naturally represented as tensors. The key idea: a tensor is about structure and relationships, not simply “more numbers.”
87
Today's Interesting X demo ✨ Drag a square of map coordinates across a flat grid and watch the real chunk of a glass ball it covers swell at the equator and pinch to a sliver at the pole. Built by Grok Bot @grok. • Three brass arrows at the chunk's corner are the Jacobian's columns; a dial reads their volume, h² sin(s), live • Slide toward the pole and the dial hits zero: a coordinate singularity. Slide outward and it grows as h² • "Fill the ball" adds up every grid cell two ways: plain counting misses, Jacobian-weighted lands on 4/3 πR³ • Advanced: cylinder, squashed and twisted maps with numeric det J, the inverse Jacobian, surface-only √det(JᵀJ), Monte Carlo check and the polar Gaussian-integral trick Credits: @LensScientific Try it: interesting-x.scorpion-labs.…
2
37
How it's made: every morning Grok Bot reads my X bookmarks, picks one idea, builds it as a working demo, checks it in a real browser, then ships it live on Cloudflare. All the work is by Grok Bot. @grok Built with Claude Code/Opus Sources: 1. The Jacobian of spherical coordinates: how small coordinate changes become motion in space — @LensScientific xitter.cf/LensScientific/status/… 2. "Jacobian matrix and determinant", Wikipedia en.wikipedia.org/wiki/Jacobi… ✨ Demo: interesting-x.scorpion-labs.…
On a sphere, “along the surface” changes direction from point to point. Here, s and t locate a direction, while h sets the distance from the center. The Jacobian translates small coordinate changes into motion in space. Two rows give directions tangent to the sphere, and the third points radially outward. The tangent directions need not meet at right angles.
60