He's Building For AI. But Doesn't Trust It With His Own Code.
Weights & Biases16 June 2026Watch on YouTube
Part of series
Ep. 1 · Database Own 15b
View the seriesWhat you'll learn
- ClickHouse currently uses AI for 50% of its codebase, expected to rise to 80% within six months, but all code goes through human review before shipping.
- For critical infrastructure where customers depend on the entire data operation, blind trust in AI-generated code is not an option.
- Practical balance between AI efficiency and quality assurance is essential at production-system codebase scale.
Frequently asked questions
How much of ClickHouse's code is AI-written today?
Why does Aaron Katz require human review of all AI-generated code?
What are the practical challenges of AI code in critical systems?
How does ClickHouse balance AI efficiency with quality assurance?
Topics
Read next
AI-code wordt intern omarmd, maar het beheer blijft ver achter
Uit het rapport 'State of AI-Powered Software Development' van Black Duck blijkt dat 97 procent van de grote bedrijven AI-codeertools gebruikt, maar slechts een derde heeft daarvoor governance ingericht. Teams die dat wel doen, melden een hogere efficiëntie.
Zelfcorrigerende AI-agents als bouwstijl in de Nederlandse startup-scene
Meerdere Nederlandse startups bouwen AI-systemen die zichzelf bijsturen op basis van feedback en meetdata, van dynamische prijsmodellen tot vroege detectie van industriële storingen. Een vergelijking laat zien hoe reinforcement en self-correction zich vertalen naar concrete producten.
Description from the channel
50% of ClickHouse's code is AI-written today, and it will rise to 80% in six months. But Aaron Katz won't let a single line ship without human review. When your customers are betting their entire data infrastructure on your service, blind trust in AI-generated code is not an option. How much of your codebase are you comfortable handing to AI?