I Tested Codex on a Real Hardware Project

Summary

The video tests OpenAI Codex, running with GPT-6.1, on a hardware-engineering task rather than a conventional software project. Codex was asked to create three versions of a four-stage CMOS ring counter, simulate them in LTspice, verify their behavior, find their maximum passing clock frequency, compare performance, and document the results.

A ring counter passes a single active state through four outputs in sequence as the clock ticks. The video uses that predictable output pattern to check whether the simulated circuits work, rather than judging the designs by whether their files merely look plausible.

Experiment and results

Codex built three implementations: static CMOS, transmission gate, and TSPC. It ran circuit simulations and checked waveforms for the expected sequence. To measure speed, it tested clock frequencies—first broadly, then in progressively narrower increments—until it identified a verified passing boundary for each design.

Design Highest verified frequency Transistor count Static CMOS 118 MHz 144 Transmission gate 124 MHz 88 TSPC 135 MHz 60

At the shared 10 MHz comparison point, TSPC also had the lowest measured supply power among the three. Thus, in this particular experiment, TSPC had the fewest transistors, the highest verified frequency, and the lowest measured power. The video does not give specific power figures.

Key analysis and limitations

The main point is the workflow Codex completed: translating a requirement into project files, using an engineering simulator, checking waveforms, testing for failures, comparing measurements, generating graphs, and documenting limitations. The creator argues that coding is becoming just one part of broader agent workflows involving tools and independently checkable results.

The results are simulation-based, not evidence of a fabricated or production-ready chip. The project used simplified educational CMOS models, and its frequency limits apply only to the defined test setup—not universally to those circuit designs. The experiment therefore does not establish that Codex can replace semiconductor engineers.

Guide or tutorial

The video gives a brief, non-specialist explanation of how a ring counter works and how circuit simulation and waveform checks help verify a design. It is primarily an experiment and analysis, not a detailed circuit-design tutorial.

Main speakers and sources

• Jayanth Murala — narrator and experimenter.
• OpenAI Codex with GPT-6.1 — AI agent used to create and test the designs.
• LTspice — circuit-simulation tool used for testing and waveform inspection.