Episode Details

Building an AI computer from scratch forces a question: why not just rent compute from a cloud provider? This episode reveals the answer—a research finding that reshapes everything. VEC2 text can reconstruct up to 92% of your original input from numerical embeddings alone, proving your data isn't as hidden as you thought. Privacy just became non-negotiable. // 00:00 - Why rent when you can buy? // 02:30 - Embeddings decoded: How text becomes coordinates // 07:15 - The VEC2 text bombshell: Your embeddings leak secrets // 13:45 - Why this changes the entire AI equation --- Sources & further reading: • ComputerBuild repo — decisions/decision-log.md, D8 (embedding inversion, single security boundary, manually-unlocked volume) • ComputerBuild repo — workloads/workload-map.md, section E (threat model, security architecture, autonomy ramp) • Morris, Zhao, Chiu, Shmatikov et al. — vec2text and embedding-inversion research (arXiv 2406.10280, 2411.05034, 2507.07700): https://arxiv.org/abs/2406.10280 • ComputerBuild repo — decisions/decision-log.md, D4 and P16 (keeping the home server separate; plaintext transcripts still exposed) • ComputerBuild repo — docs/build-recommendation.md, section 5 (disk-encryption performance, and the warning against quoting the worst case) This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.