The AI community building the future of chip design.
ActGen is where engineers, researchers and companies collaborate on the models, datasets, evals and agents that teach AI real silicon work — created by experts, checked by machines, credited forever, and shared under consent. Knowledge that used to be locked inside companies becomes something the whole field can build on.
Trending on ActGen this week
The same live chart for every reader — computed from the hub's own activity, never curated by hand.
What ActGen is
The rarest data in AI is real chip-design work. Experts here create it — in the open.
Every other field got one place where its data, the benchmarks that grade it, and the models trained on both live together. Silicon never did — the knowledge stayed locked inside companies. ActGen exists to change that, and the engine is one loop: working with the labs pushing the AI frontier, we build the rubrics and evals that measure what their models can and cannot yet do in chip design; where the evals expose a gap, our community of experts creates the rare data that closes it — real corrections, real tool traces, all the way down to production layout — and that data goes back into training, so the next generation of models is measurably better at silicon. Measure, create, improve, re-measure. Every contributor is credited, and consent is honored on every rebuild: a contributor who withdraws is withdrawn, everywhere, every time.
Organizations sponsor the datasets; experts produce them through real design work. The public release advances the whole field, sponsors receive the held-out partitions that referee their own models, and the labs training on the data make every tool here better — with each contributor named on the record.
The referee inside chip companiesDesign teams use the same versioned, runnable rubrics and evals to referee the agents they build in-house and the ones vendors sell them — across pre-silicon, post-silicon, and everything between — so adoption decisions rest on measurements, not demos. We referee; we don't take sides.
From workload to packaged siliconTeams that have never owned a chip flow arrive with the software workload they wish ran in hardware, work in a tight loop with the platform's experts and agents, and leave with a complete packaged part that plugs into their infrastructure — humans finishing whatever the agents cannot.
Sponsoring a dataset is how organizations join: fund the open benchmarks, bring the tasks today's AI cannot solve, and the whole field — including you — gets to measure progress against them.
The hub for chip design
Versioned, ownable artifacts you can browse, publish and pull — filed by their role in the design flow rather than by modality. Blocks are silicon's own kind, and the one a general-purpose model hub has no equivalent for.
- ModelsRTL generators, PPA predictors, floorplanners, verification agents
- DatasetsLayout databases, netlists, testbenches, waveform corpora, expert corrections
- SpacesRunnable demos: a floorplan viewer, a DRC-violation browser
- IP blocksHardware blocks with their collateral, versioned and licensable
Your published work builds your public profile — every artifact, credit and signed release under your own name at /u/you.