AI tools at Oxide: useful edges, useless epicenters

The history of servers, the cloud, and what’s next – with Oxide

Topic 12 of 15

Published Dec 17, 2025 · Clip 1:02:53–1:17:36 (14:43)

Brian describes how Oxide uses AI: no part of the stack is vibe-coded, but LLMs are used for tedious tasks, test generation, and unmatched document comprehension in their writing-intensive culture (e.g., generating an RFD glossary). He recounts a prominent AI doomer's reaction video to his talk 'Intelligence Is Not Enough' that fast-forwarded through his concrete hardware examples. He tells the story of the first Oxide bring-up where the CPU wouldn't come out of reset—a problem no LLM could solve until a desperate engineer inspected the CPU-to-voltage-regulator protocol and found missing acknowledgement packets due to a Renesas firmware bug—arguing that building a board is not an IQ test and that teams with diverse approaches are essential. He cites Richard Sutton on conflating LLMs with AI—prompts are not goals—and stresses engineers remain accountable for their work, recommending people run LLMs locally to see their limitations, and finds LLMs more valuable in the small (Rust idiom checks, editing) than in the large.