All content
Takeaways
Short takes on what's new in AI and data, and what it actually means for your work.
Your AI evaluation might be biased against the model you're replacing
Property Finder found 96.8% of its eval labels came from its old model. After rebuilding ground truth, a different model won at 86% lower cost.
AI access is a different governance problem from web access
Cloudflare now separates Search, Training and Agent crawlers. The question shifts from 'can AI access this?' to 'why is it accessing it?'
Which parts of your role are moving from human-led to AI-led?
Anthropic says Claude leads 26% of its AI R&D work. The better question about AI and jobs isn't replacement, it's which tasks shift.
As code gets cheaper, verification gets more valuable
OpenAI's report on researchers using coding agents: even successful long tasks needed human steering. Knowing what good looks like is the skill.
If AI builds the pipelines, what's left for data engineers?
Zeit AI is building an 'autonomous data engineer'. Connecting systems may get easy; trustworthy definitions, lineage and edge cases won't.
The analytics interface is changing
Self-service analytics is moving from dashboards to natural language. The warehouse tells you what happened; business context explains why.
Prompting isn't the interesting skill with coding agents
Context, decomposition, architecture, validation and judgement matter more. Two habits that made Claude far more useful in my data work.
If your semantic layer disappeared tomorrow, could you rebuild it?
AI makes writing SQL easier, but the hard part was always knowing what the business means. What should data teams actually preserve?
One useful takeaway, straight to your inbox
Join the newsletter for new takeaways and the tools worth your time. No hype.