社区长文：TapeOut 与数字电路的 AI 交叉分析Community long-form: TapeOut at the AI–digital circuit intersection
Onchaingamblr 发布编号 1–11 的长文，分析 TapeOut 在模型训练与执行之间的位置。文中引用 NeurIPS 2025 论文《Mind the Gap》：CIFAR-10 上软硬电路差距接近 3%，实验差距降低 98%。并指出 Microduck #281 用 39 个 NAND 门处理 14 位虚拟深度与目标方向，#282 用 22 个 NAND 加 3 个 LATCH 增加固定转向。V2 元件方面，第三方可开发确定性、纯计算、无管理员的固定规格元件，工厂用 BEM 取得制造能力，用户用 BNB 铸造。作者还建议首个项目输入不超过 16 位、输出不超过 4 位。Onchaingamblr published an 1–11 thread analyzing where TapeOut sits between model training and execution. It cites the NeurIPS 2025 paper "Mind the Gap": the soft–hard circuit gap on CIFAR-10 is close to 3%, and the experimental gap was reduced by 98%. It notes Microduck #281 uses 39 NAND gates to handle 14-bit virtual depth and target direction, while #282 uses 22 NAND gates plus 3 LATCHes to add fixed steering. On V2 components, third parties can build deterministic, pure-computation, admin-free fixed-spec components, factories obtain manufacturing capability with BEM, and users mint with BNB. The author also advises keeping a first project to no more than 16 input bits and 4 output bits.
来源 · 8 条Sources · 8
- @Onchaingamblr
Preface Over the past few years, the AI industry has been constantly adding. More parameters, longer contexts, more GPUs, larger data centers. Models are increasingly becoming black boxes capable of containing half of the internet, and we have also become increasingly accustomed to using scale to
- @Onchaingamblr
BNN makes a neural network more like a digital circuit; DLGN directly makes the digital circuit the model. But here we must immediately draw a system boundary. A small logic core does not equal a complete AI system. A 20-gate controller usually only means that its logic core is very small. It doe
- @Onchaingamblr
Later research exposed the most practical difficulty: A soft model performing well does not mean the discretized hard circuit will also perform well. NeurIPS 2025’s Mind the Gap observed a soft-hard gap close to 3% on CIFAR-10, and reduced the experimental gap by 98% and improved convergence speed
- @Onchaingamblr
3. Why Tapeout fits exactly between training and execution Large models are good at probability. They can give you an answer that is probably correct based on enormous amounts of experience, and sometimes hallucinate while being extremely confident. Blockchain and digital circuits are good at someth
- @Onchaingamblr
5. Microduck goes one step further: Circuits can be combined, but memory isn’t truly on-chain yet Microduck’s structure is more interesting. #281 uses 39 NAND gates to process 14-bit virtual depth and target direction and output steering. #282 uses 22 NAND gates + 3 LATCHes to add a fixed steerin
- @Onchaingamblr
7. How developers should get started For the first project, don’t touch image classification, and definitely don’t touch LLMs. Choose a task with no more than 16 input bits, no more than 4 output bits, and one whose entire input space can be fully verified. Step 1: Define the inputs before writin
- @Onchaingamblr
9. What should V2 components actually sell? According to the official V2 preview, third parties can develop deterministic, pure-computation, administrator-free, non-agent fixed-specification components. Factory uses BEM to obtain component manufacturing capabilities. Users use BNB to mint. A por
- @Onchaingamblr
11. How I would rather imagine Tapeout’s future Future large models will probably become larger and smarter, and increasingly resemble utilities such as electricity and water, provided by a small number of infrastructure providers. But not every action needs to ask a large model again. Once an ex