The MLnotes Newsletter
Articles about The MLnotes Newsletter
Stop Feeding Your Agent Raw Files: How Spotify Cut Coding Agent Token Costs by 90%
Why 80% of agent compute is wasted on file I/O, the 3-layer shunting architecture, and how to delegate boilerplate without losing reasoning quality
The Autoresearch Loop: How Instacart Uses AI Agents to Beat Production ML Baselines
Why manual hyperparameter tuning and architecture sweeps are dead, and the 3-stage agentic modeling loop that cut delivery error by 4.8%.
How Netflix Uses Multimodal AI to Personalize Every Movie Poster (And How to Build It)
From blind ID collaborative filtering to CLIP embeddings: how a single two-tower model serves 260M+ users with zero-shot cold start
Tests Enforce Journeys. Agents Verify Goals: How Slack Uses AI for End-to-End Testing
Why replacing Playwright scripts with LLMs in CI/CD is an expensive mistake, what 200+ test runs revealed about MCP vs CLI, and where agentic testing actually belongs
Stop Dumping Chat History into Context
The 4 architectural decisions behind production agent memory
The Anatomy of an AI On-Call Agent: Why the Next Incident Responder Won’t Be Human
How Ramp automated 50% of incident triage, why bloated 20,000-token system prompts fail, and the 3-loop architecture of autonomous on-call agents.
The AI Model That Won’t Talk to You: Why “System One” Models Are the Missing Piece for AI Workflows
Why autoregressive JSON causes friction in production, how non-generative models deliver sub-100ms typed decisions, and what our 50-sample benchmarks revealed.
My Benchmark Was Testing a Model Nobody Used
Four bugs, a language that doesn't write its own vowels, and $2,530 spent before I built the listening test I should have started with.