Start Here
A short map of the main themes on this site.
This site is about the practical side of ML and AI systems: how models, agents, training pipelines, evaluations, and product interfaces behave when they leave the paper and enter a real engineering loop.
The fastest entry point is to pick a path below, then follow the linked topic hub when you want the full archive.
Reading Paths
LLM Agents
Tool use, agent runtime design, evaluation, context, and production patterns for systems that act across tools and environments.
- Scaling RL for White-Collar Work: The Environment Foundry
A practical framework for turning common white-collar workflows into RL environments: spreadsheets, CRM tasks, customer support, web research, dashboards, and other software-mediated work.
- How to Arbitrarily Increase the Difficulty of Agent Evaluation Sets
A practical framework for making agent benchmarks harder in a controlled way: treat difficulty as trajectory-graph complexity, not prompt wording. Covers deterministic scoring, capability facets, harness effects, and systematic data generation.
- Improving LLM Internationalization: Bridging the Gap in Tool Use and Agency
LLMs achieve 57% tool-calling accuracy in English but only 34% across 52 languages — and 6.8% for the worst. This post covers the full playbook for closing the multilingual gap: training-time techniques, agentic architecture patterns, failure mode analysis, and RL-based approaches for i18n.
Evaluation
Practical approaches to measuring model and agent capability with deterministic checks, rubrics, trajectories, and verifiable outcomes.
- Why Embeddings Cannot Solve Eval-Set Contamination
A technical deep dive on why semantic embedding search is useful but insufficient for eval-set decontamination: leakage is about evaluation advantage, not just text similarity.
- Pretraining Contamination: Why Don't Train on the Test Set Became Hard
A practical introduction to LLM pretraining contamination: why benchmark leakage is not ordinary deduplication, how public evals leak into web-scale corpora, and how layered decontamination pipelines reduce risk.
- How to Arbitrarily Increase the Difficulty of Agent Evaluation Sets
A practical framework for making agent benchmarks harder in a controlled way: treat difficulty as trajectory-graph complexity, not prompt wording. Covers deterministic scoring, capability facets, harness effects, and systematic data generation.
Post-Training
SFT, RLHF, preference optimization, instruction following, reasoning traces, and data pipelines for shaping model behavior after pretraining.
- Scaling RL for White-Collar Work: The Environment Foundry
A practical framework for turning common white-collar workflows into RL environments: spreadsheets, CRM tasks, customer support, web research, dashboards, and other software-mediated work.
- The Mercor Breach: What 4TB of Stolen Data Reveals About How Frontier AI Labs Actually Train Models
A $10B AI data vendor was breached, exposing 84 Airtable workspaces of training data for OpenAI, Anthropic, Apple, Amazon, and Meta. This post analyzes what the public reporting reveals about each lab's evaluation methodology — rubric design, RLHF pipelines, and quality control — and what it means for the industry.
- The Unverifiable Reward Problem: The Real Frontier of RL for LLMs
Deep research on tasks with unverifiable rewards in RL — the key bottleneck for scaling RL beyond math and code. Covers JEPO, NRT, RLNVR, self-play methods, GenRM, Constitutional AI, reward hacking mitigation, and more.
RLHF and Preference Optimization
Engineering notes and research synthesis on PPO, DPO, GRPO, reward modeling, preference data, and model behavior optimization.
- Scaling RL for White-Collar Work: The Environment Foundry
A practical framework for turning common white-collar workflows into RL environments: spreadsheets, CRM tasks, customer support, web research, dashboards, and other software-mediated work.
- The Mercor Breach: What 4TB of Stolen Data Reveals About How Frontier AI Labs Actually Train Models
A $10B AI data vendor was breached, exposing 84 Airtable workspaces of training data for OpenAI, Anthropic, Apple, Amazon, and Meta. This post analyzes what the public reporting reveals about each lab's evaluation methodology — rubric design, RLHF pipelines, and quality control — and what it means for the industry.
- The Unverifiable Reward Problem: The Real Frontier of RL for LLMs
Deep research on tasks with unverifiable rewards in RL — the key bottleneck for scaling RL beyond math and code. Covers JEPO, NRT, RLNVR, self-play methods, GenRM, Constitutional AI, reward hacking mitigation, and more.
Generative UI
How AI systems can produce, steer, and execute user interfaces with structured representations and practical product constraints.
- Ad Formats in LLM Products: What's Live vs. What's Research
A survey of advertising formats in LLM products—separating what's deployed in production from what remains in research.
- Generative UI Doesn't Move the Needle—Steering Does
After shipping multiple generative UI features, I've concluded that the sophistication of AI-generated interfaces often doesn't translate to user benefit—but steering does.
- UI Representation and Action Execution for Generative UI
Exploring structured UI representation using JSON Schema, and how to implement action handlers for generative UI systems.