AI Safety & Alignment

Excited to announce: I'm joining Anthropic as Head of AI Safety Research. After 8 years at DeepMind, this feels like the right moment to focus entirely on alignment. The problems are getting harder, but the community is getting stronger. Grateful for everyone who supported this journey. Let's build safe AI together. 🙏

Unpopular opinion: The best MLOps is the MLOps you don't need. Before building a complex ML pipeline, ask: 1. Can a simpler model solve this? 2. Do you actually need real-time inference? 3. Is batch processing good enough? 4. Can you use a managed API instead? 90% of the time, the answer to at least one of these is 'yes'. Stop over-engineering.

If you're an AI researcher feeling burned out, you're not alone. The pace of this field is unsustainable. New papers every day, pressure to publish, constant paradigm shifts. Here's what's helping me: • Blocking 2 hours daily for deep reading (no Slack) • Saying no to 80% of speaking invitations • Accepting that you can't read everything • Finding 2-3 people you trust for paper summaries Your m...see more

Fascinating result from our experiments on in-context learning: We found that the order of few-shot examples matters dramatically — sometimes more than the examples themselves. Optimal ordering improved accuracy by 15-30% across 12 benchmarks. We're calling it 'positional priming' and working on a paper. Has anyone else observed this?

Just finished interviewing 50+ AI engineers about their tech stack. Consensus stack for 2026: • Models: GPT-4o + Claude 3.5 Sonnet + Llama 4 (open) • Frameworks: LangChain or LlamaIndex • Vector DB: Pgvector (startups) / Pinecone (enterprise) • Orchestration: Temporal or Inngest • Monitoring: W&B + LangSmith • Deployment: Replicate, Modal, or self-hosted The stack is maturing fast.

A practical guide to model monitoring in production: 1. Track output distribution shifts (not just accuracy) 2. Monitor latency at p50, p95, and p99 3. Set up automatic fallbacks to simpler models 4. Log all inputs/outputs (with PII handling) 5. Create canary deployments for model updates Most teams skip monitoring until something breaks. Don't be most teams.