No Priors Ep. 10 | With Copilot's Chief Architect and founder of Minion.AI Alex Graveley

Summary

• GitHub Copilot advanced from passing fewer than 10% of in-the-wild tests to more than 60%, showing how Alex Graveley’s team improved a weak model through evaluation, prompting, code data, historical versions, diffs, and scale. Alex emphasized that the early results were not a reason to conclude the approach was impossible: “the crazy thing is that these things work at all,” and they can improve as they scale.
• Copilot’s key product breakthroughs were ghost-text block completion and extreme latency reduction, rather than a conventional question-and-answer interface. Programmers could judge a completed block quickly and press Tab; Alex recalled a “crazy” statistic that every additional 10 milliseconds meant roughly 1% fewer accepted completions. Poor results in India traced to requests traveling through Europe to OpenAI’s Texas data center.
• The economics improved faster than expected: estimated inference cost fell from $30 per user per month to $10 and then lower, while weekly-cohort retention remained above 50% months after launch. Alex pushed for low pricing and market capture. Elad framed 40% of a user’s code as not much at that price, while Alex said some “whales” have Copilot write 80% of their code.
• The next platform shift Alex described is from AI that answers questions to agents that take controlled actions. Minion targets scheduling, travel, taxes, and identifying neglected contacts—“Copilot applied to everyday activities”—with Alex saying he thinks “file my taxes” can become feasible within the next few years.
• Real-world action can provide an observable feedback loop. Code is unusually valuable because it can be run; similarly, web actions reveal success, failure, and annoyance. Alex also sees people’s clicks on the web as a gigantic, currently “unowned” dataset.
• Alex sees near-term AI risk primarily in fraud and malicious use, not machines autonomously killing humanity. He forecasts “many years of extreme discomfort” involving impersonation and stolen savings, and argues for anti-abuse systems and legislation rather than expecting OpenAI to solve everything. He cited spam blockers, account filtering, and fingerprinting; Elad noted that agents can absorb some messaging costs, while Emil pointed to traceable bank accounts and compared the threat to a better-resourced or more distributed North Korea.
• Alex described AI-native product intuition as bimodal: very senior builders and very young or naive people often see the potential, while the experiential middle may reject it. The reusable traits are experimentation, metric selection, intuition, and tenacity after failed attempts. He is interested in what emergent properties may appear if “these things” become 10 times bigger, comparing the exploration to building the Large Hadron Collider.

Deep dive

Not yet available upstream; scheduled sync will retry.