This report asks whether an independent hardware developer who cannot afford frontier model API costs can use low-token-cost models in an agentic harness to design, simulate, and document a simple electronic circuit. I asked two low-cost LLMs, one locally hosted and one remote, to develop a hardware random bit generator where randomness is extracted from avalanche noise in a reversed-biased semiconductor junction, displaying one random bit on an LED each time a button is pressed. The models were prompted to simulate their circuit designs with ngspice and to generate schematics and parts lists. Both models failed to produce a working circuit design. Schematic drawing failed entirely despite remediation attempts with more capable models. I conclude that low-token-cost agentic AI does not offer meaningful assistance to the frugal hardware hacker. This reinforces a broader concern that capable agentic AI tools are out of reach for independent developers who have to buy tokens with grocery money.
Leave a Reply