Reminder that correlation is not causation, but also that dismissing every correlation with 'correlation is not causation' is not wisdom — it's intellectual laziness. Use causal inference methods.

Just published our new paper on constitutional AI evaluation frameworks. We found some surprising results about how different alignment techniques interact. Link in comments.

We're hiring senior engineers for our fraud detection team! If you're passionate about ML applied to financial crime prevention, DM me. Remote-friendly. 🎯

The best product managers I've worked with share one trait: they're genuinely curious about their customers' problems. Not the features, not the solutions — the problems. This is harder than it sounds.

Year 2 of our startup: - 5 team members → 23 - $0 ARR → $2.1M - 0 customers → 340 Biggest lesson: focus on 10 users who love you, not 1000 who think you're okay.

The gap between 'I can train a model' and 'I can deploy a model to production' is vastly underestimated by the ML community. Infrastructure engineering is not beneath you — it's the bottleneck.

Unpopular opinion: most product roadmaps are just wishlists with dates on them. A real roadmap should be 70% problems to solve, 20% experiments to run, and 10% features to build.

TIL that PostgreSQL's JSONB indexing is way more powerful than I thought. GIN indexes with jsonb_path_ops can be 3x smaller and faster than default GIN for containment queries. Sometimes the boring database does everything you need.

Hot take: most microservices architectures would be better as monoliths. I've spent 7 years migrating services TO microservices and 3 years migrating some BACK. The right architecture depends on your team size and deployment cadence, not industry trends.

Excited to share that our team just shipped a new distributed caching layer that reduced p99 latency by 40%! 🚀 Huge shoutout to the infra team for their support during the migration.

Gave a talk at BSides Austin this weekend on detecting browser automation. The behavioral signals are fascinating — turns out mouse movement patterns are incredibly hard to fake convincingly. Even Bézier curves have telltale smoothness that real hands don't.

Engineering career advice nobody gives you: your first 2 years at a company are spent building context. The real impact starts in year 3. Job hopping every 18 months means you never get past the context-building phase.