had claude fable 5.2 build a 10 storey concrete office building in @kinth_ai. 356 parts, foundation to rooftop solar
then shook it. 0.45 g earthquake, run as an FEA time history on the frame, nobody keyframed the sway
roof moves 205 mm, shown 14x
had claude fable 5.2 build a quadcopter with @kinth_ai's CAD API. 107 parts, it even modelled the motor windings
the flight is a physics sim running off the model's own mass and inertia, nobody keyframed anything
animated in kinth
@0x_fokki since it's built from one picture, are the internals inferred or did you give it a parts list? the step rate dropping when the battery comes out is a nice touch
@M__Houser 21B tokens and no coding background is a wild combination. what broke first as it grew, the world itself or keeping the code coherent across that many sessions?
@jpschroeder 'without CAD software' is the interesting line. it gets further with a kernel under it: a 107 part quadcopter as real solids, then flown in a physics sim https://t.co/6I8YA5mF4T
was the garden procedural or did it place things one by one?
had claude fable 5.2 build a quadcopter with @kinth_ai's CAD API. 107 parts, it even modelled the motor windings
the flight is a physics sim running off the model's own mass and inertia, nobody keyframed anything
animated in kinth
@aecyydesign clean. did fable drive blender through python or write the scene from scratch? asking because the motion in mine today came out of a physics sim, and i'm curious how much of yours was keyframed by hand
@leo_linsky what does an engineering design environment look like in your set, is the output checked by a solver or scored by a rubric? asking because on a mechanical one today it passed stress by 46x, then found a resonance on its own and redesigned the part https://t.co/dqlwLKqlnA
had fable 5.2 run FEA on the drone's arm (shown in @kinth_ai)
stress is fine, 2.4 MPa at full throttle. but it has a bending mode at 84 Hz and the rotors hover at 83 Hz
so it redid the arm in hollow carbon. 151 Hz, 24% lighter
@orcdev three.js scenes get judged by eye. the other end of 3D is geometry that has to be exact, like CAD, where a part is right or it isn't. would your 4 elements benchmark have a way to score that, or is it visual only?
@AymericRoucher letting it build its own physics is the part that scales. same thing happened with a quadcopter today, it wrote a 6-dof flight sim off the CAD's own mass and inertia https://t.co/6I8YA5m7fl
did you check the shallow-water solver against anything, or judge it by eye?
had claude fable 5.2 build a quadcopter with @kinth_ai's CAD API. 107 parts, it even modelled the motor windings
the flight is a physics sim running off the model's own mass and inertia, nobody keyframed anything
animated in kinth
@lennysan point 7 is the one. the growth is in work where the output has to obey physics: hardware, robotics, energy. that's also the work software-style AI hasn't reached yet, so the people who can check a design get more valuable
@bindureddy doesn't look incremental from mechanical engineering. it modelled a 107 part drone, ran FEA on its own arm, found a resonance the stress check passed, and redesigned it. one afternoon https://t.co/dqlwLKqTd8
had fable 5.2 run FEA on the drone's arm (shown in @kinth_ai)
stress is fine, 2.4 MPa at full throttle. but it has a bending mode at 84 Hz and the rotors hover at 83 Hz
so it redid the arm in hollow carbon. 151 Hz, 24% lighter
@0xMovez does it simulate the electrical side, pin states and current limits, or is it a 3D reference for now? wiring is what beginners get wrong, and a sim can catch that before anything smokes
@Grummz@felixrieseberg the part that carries over to engineering is it building every asset itself. gave it a quadcopter instead of a game: 107 parts as real CAD solids, then its own flight sim to fly them https://t.co/6I8YA5mF4T
had claude fable 5.2 build a quadcopter with @kinth_ai's CAD API. 107 parts, it even modelled the motor windings
the flight is a physics sim running off the model's own mass and inertia, nobody keyframed anything
animated in kinth
@ArtificialAnlys@GoogleDeepMind reading a schematic and marking it up is the review half of engineering. the other half is producing geometry that survives a check: mass, clearance, vibration. a benchmark where the deliverable has to pass a solver would be worth a lot
@Rewkang the part of physical AI nobody prices in is the design work before the robot exists. every one is thousands of mechanical decisions checked against physics, and that's only now getting automated. fable 5.2 caught a resonance in a drone arm for me today https://t.co/dqlwLKqTd8
had fable 5.2 run FEA on the drone's arm (shown in @kinth_ai)
stress is fine, 2.4 MPa at full throttle. but it has a bending mode at 84 Hz and the rotors hover at 83 Hz
so it redid the arm in hollow carbon. 151 Hz, 24% lighter
@PawelHuryn replanted bugs work as a test because there's a ground truth to score against. mechanical design doesn't have one, a part can pass every check you wrote and still resonate. what would the hardware version of this benchmark score?
@mattshumer_ games are the forgiving test, nothing in them has to carry a load. the stricter one is geometry that has to be right. had it model a 107 part quadcopter in CAD today and fly it in a sim off the model's own mass and inertia https://t.co/6I8YA5m7fl
had claude fable 5.2 build a quadcopter with @kinth_ai's CAD API. 107 parts, it even modelled the motor windings
the flight is a physics sim running off the model's own mass and inertia, nobody keyframed anything
animated in kinth
@_MaxBlade herringbone gears meshing on the first print is a good sign, they punish any tooth profile error. what did it write the CAD in?
tried it on a bigger assembly today, a 107 part quadcopter https://t.co/6I8YA5m7fl
had claude fable 5.2 build a quadcopter with @kinth_ai's CAD API. 107 parts, it even modelled the motor windings
the flight is a physics sim running off the model's own mass and inertia, nobody keyframed anything
animated in kinth