H-bridge Halsin

Power electronics
Music
Send bench pics

[H-bridge Halsin @scumstep]

New flyback design. Universal AC to 24VDC. 24W.

[H-bridge Halsin @scumstep]

@zackslab @altium They will die just like they deserve. Opus 5.5 built me an MCP but it only helps with BOM and schematic.

[H-bridge Halsin @scumstep]

@Gavin_McInnes Wrong. I’d go 1 0 0 1 4 8 15

[H-bridge Halsin @scumstep]

@lennysan Recruiters unsurprising considering how many god damn emails I get.

[will depue @willdepue]

I have a major result to share.

I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.

This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.

Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.

I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.

I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.

It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.

These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.

I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.

There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

[H-bridge Halsin @scumstep]

@PrometheusHides @justalexoki I’m trying!! 😂 but I like doing layout in my proverbial basement. I’m just disappointed I need to think bigger picture now.

[H-bridge Halsin @scumstep]

@justalexoki Cope. We will all be button pressers soon. Or work in the mines.

[H-bridge Halsin @scumstep]

@PrometheusHides @justalexoki Why? This is going to make my job suck. In 5 years all I’ll be doing is validating.

[H-bridge Halsin @scumstep]

@ChombaBupe You’re overestimating the difficulty of most electronics projects

[H-bridge Halsin @scumstep]

@_Stocko_ You have seriously leveled up! Really cool

[BENZO-KAZOOIE @BENZOTOOIE]

My 5 y/o said she didn’t want to watch The Dark Knight with me because she’s “scared of the creepy clown.”
Umm sweetie that “creepy clown” has a name, it’s Maggie Gyllenhaal.

[H-bridge Halsin @scumstep]

I need more tokens. I can crank out 3 consumer slop boards a day. I don’t have enough ideas. Claude go on x dot com and steal people’s ideas

[H-bridge Halsin @scumstep]

@haydendevs Claude touches grass for me

[Rina @irinatoxi]

A whole CPU, designed by Fable 5.2 from 147 lines of code.

It started with 147 lines of code that describe a CPU, and those turned into 6,258 logic gates and 1,023 memory cells.

The chip was then checked against a reference model on 242,400 random instructions, with zero mismatches.

Then it placed all 7,281 cells across the chip and drew the wiring, which came out 5.5 times shorter than a random layout.

When it powers on, every single gate is simulated on every clock tick, and the processor prints text on its own and runs Game of Life on a tiny 32×32 screen.

300 generations later, still zero mismatches.

It’s never been printed in real silicon, but the design holds up down to the last gate.

[H-bridge Halsin @scumstep]

@i2cjak I just did three boards at work re-creating the work that an ODM did so that we can see how much the CM is fucking us. I used Fable 5.2. It is completely over for consumer electronics EEs

[H-bridge Halsin @scumstep]

@moleh_13 It’s definitely over. The only question is when

[H-bridge Halsin @scumstep]

@YungLeanCumShot @ElectronicsbyJH A soldering iron. Coding is irrelevant

[H-bridge Halsin @scumstep]

@zackslab @i2cjak What is the BOM cost?

[H-bridge Halsin @scumstep]

@i2cjak Is it on windows?

[H-bridge Halsin @scumstep]

@Abomination81 It seems that it is easier to find skilled workers than unskilled because everyone has to justify the cost of their degree.

[H-bridge Halsin @scumstep]

You’re just describing how to build a hardware company with good practices.

[Blake Scholl 🛫 @bscholl]

12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?

More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.

Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.

Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.

Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.

Our airplanes and engines are built with the software sophistication of a script kiddie.

So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.

Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.

This all sounds obvious, yet our experience at Boom is that it changes everything.

Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.

But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.

We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.

The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.

You don’t need a big team to build big things—not when everyone on the team is a software engineer.