After a year of weekends, I’m finally releasing ChipCraftX and ChipCraftBrain. You can find it at https://chipcraftx.io, or the arXiv paper at https://arxiv.org/abs/2604.19856.
I built this to fix my own problem: I want to think and iterate on new computing paradigms, fast. Today there is no single tool that lets a computer architect do that. Most of the work goes into tooling, scripting, and stitching things together instead of the ideas worth proving. Iterating fast means fast prototypes, cycle-accurate simulation, RTL, and a long list of other things all working together. ChipCraftX aims to cover that path end to end:
- Ideation to architecture
- Spec to RTL (SOTA on VerilogEval and CVDP; full results on the site)
- Simulation, synthesis, verification
- SystemC and multi-level simulation, cycle-accurate or time-approximate
- PyTorch / ONNX → RTL
- Reductions across FPGA platforms
It is not vibe-coded. Agents communicate and correct each other’s work through proper CI/CD flows with strict software discipline. The point is that a computer architect can use one suite to analyze, simulate, and optimize hardware without stitching together ten USD 100K licenses.
The World Model Link to heading
One of the special design points of ChipCraftX is that it is self-observing. A World Model watches roughly 70 perspectives and uses them to decide the next action given the world it is operating in. Twenty thousand trajectories in, and it is getting interesting. It does not learn Verilog or SystemC. It learns what to do next. This is LeCun’s idea applied to chip design, and I think it is going to change how systems get built. More on this later.
Under that sits a conventional multi-agent setup with surface and underlying layers, each with its own domain discipline. The knowledge base grows at every failed RTL or testbench generation. Five iterations per problem; usually done by the third or fourth.
Cost, and why it matters Link to heading
Twelve months of weekends and roughly USD 20K in API bills. Vibe-coded, this would have cost under USD 2K. The AI slop problem is real, and it is worth the extra money to avoid shipping it. Nobody should ship software without tests, verification, and documentation.
Early versions could not produce anything bigger than a single-core block. Hierarchical decomposition, added recently, scales to arbitrary SoCs. Enabling a feature is a matter of documenting it correctly so the agents have the right context.
The paper and the handbook Link to heading
The arXiv paper covers how ChipCraftBrain works: https://arxiv.org/abs/2604.19856. The ChipCraftX handbook is 203 pages and growing.
What I’ve built with it Link to heading
Besides reaching SOTA on the Verilog benchmarks, I put together a demo of Doom running on a RISC-V SoC on an FPGA, and watched Connoisseur optimize the system to raise the FPS. Connoisseur is the companion agent. It knows ChipCraftX inside out, ships as a VSCode extension, and has a web-portal interface.
The RTL dataset was generated by AutoChallenge, inspired by Karpathy’s AutoResearch: seed from a hand-curated open-source collection, then generate novel specs. You do not need millions of RTL lines to train a good system. A sub-30B model paired with the right logical structure is enough to produce correct RTL and testbenches. I used to pay for testbench generation. A local model does it for free now.
Why I built it Link to heading
I want to prototype new computing paradigms, and the tooling to do that at one-architect scale does not exist. Not for a million dollars, not for ten licensed tools stitched together. Commercial tooling optimizes known flows; I’m trying to build flows that don’t exist yet. I’ve been writing software to prove architectural concepts for 16 years, and ChipCraftX is the ultimate version of that work.
It’s also a bet on a thesis: the only winner of AI for chip design will be the one who gets the logic right. That means the system-level logic of how agents are wired, not any one model. Inception’s Mercury, LeCun’s World Model, AI-hardware companies like Taalas. The signs point the same direction.
If you want to collaborate, reach out.