A classifier is not a classifier is not a classifier
Four levels of yes/no decision problems—and how to choose between rules, classifiers, language models, and agents.
From the team
Blog
Notes on building reliable AI systems, running models at scale, and the experiments that shape our work.
Bridging the Subjectivity Gap: How Automated Prompt Optimization Helps Teams Build Expert-Aligned AI Functions
How reflective optimization aligned eleven models to Sutro’s lead-scoring judgment using only 30 annotations.
The Future (and Present) of AI is Synthetic Data
How language models are changing synthetic data—and what that unlocks for training, retrieval, simulation, and privacy.
The End of Moore’s Law for AI? Gemini Flash Offers a Warning
What Google’s Gemini Flash price increase says about the economics of inference and the limits of ever-cheaper intelligence.
(No) Need For Speed: Why Batch LLM Inference is Often the Smarter Choice
Why bulk AI workloads are often cleaner, less expensive, and faster to operate with batch APIs instead of synchronous inference.
Workhorse LLMs: Why Open Source Models Dominate Closed Source for Batch Tasks
A practical comparison of model performance, inference cost, and open-source alternatives for common AI workloads.
Generating 1 Million Synthetic Humans
A method for seeding diverse language-model outputs, with an accompanying open dataset of one million synthetic people.
Hacker News is Obsessed with Aviation
What we learned by using small language models to classify more than 40 million Hacker News posts and 10.7 billion tokens.
Model Security with Large-Scale Inference
A practical framework for probing open-source coding models for malicious behavior before bringing them inside your systems.