ML Practitioner Eugene Yan Reflects on a Peaceful Year of Steady Progress
The 2024 review from a well-known ML engineer highlights the value of sustainable, incremental improvement over flashy breakthroughs.
AI & ML
Frontier models, research results, agents, and developer tooling.
Lessons from 39 Hard-Won ML Systems and Scaling Insights
Eugene Yan distills practical lessons on building, scaling, and shipping machine learning systems.
Benchmarks Saturated, but Chatbots Lack a Sense of Purpose
As LLM benchmarks like MMLU, HumanEval, and MATH approach saturation, one article asks whether user experience is improving in step—and argues a key element is missing.
Writing Rules Are Paradoxical, and That's the Point
Eugene Yan's essay argues that writing advice is inherently contradictory, offering no single formula.
Math-Heavy ML Designs Now Yield Only Marginal Gains
A new essay argues that the past decade has seen mathematically principled architectures deliver only marginal improvements, signalling a shift in how machine learning research advances.
LLMs Move From Chatbots to Financial Market Analysis
A new essay examines how large language models, the technology behind ChatGPT, are being applied to financial markets amid an AI investment boom.
AI Needs Positive Visions Tied to Wellbeing, Essay Argues
A new essay argues that AI discourse should move beyond fear and hype toward concrete, wellbeing-centered visions of the future.
A Human Judge Reflects on W&B's LLM-as-a-Judge Hackathon
Eugene Yan's first-person account offers a practitioner's view of evaluating LLM-based judges at a Weights & Biases hackathon.
Task-Specific LLM Evals: What Works and What Doesn't
A practical guide to evaluating LLMs for classification, summarization, translation, copyright regurgitation, and toxicity, separating reliable methods from misleading ones.
What LLMs Could (and Probably Can't) Do for Self-Driving Cars
An analysis in The Gradient examines whether large language models can overcome autonomous driving's hardest problems, and why trust remains the central obstacle.
LLM Deployment Lessons from Netflix PRS 2024
A practitioner talk from Netflix's PRS 2024 highlights evals, scalability, and guardrails as the key hurdles in applying LLMs to recommendation experiences.
GTC 2025 Talk Shares Lessons for Building LLM Apps
Eugene Yan and Chip Huyen present their hard-won insights on building LLM-powered applications at NVIDIA's conference.
Text Embeddings Can Be Inverted, Sparking Security Review
A new method called Vec2text shows that embeddings can be accurately reverted to text, highlighting the need to update data security protocols.
AI Coach Uses Speech-to-Text, LLM, and Virtual Number
Eugene Yan's personal project demonstrates how voice-based AI coaching can be built from speech recognition, text-to-speech, an LLM, and a virtual phone number.
Mamba: A State Space Model Alternative to Transformers
Mamba, a novel AI model based on State Space Models, offers a more efficient approach to processing long sequences than Transformers.
Eugene Yan's Writing FAQ: A Personal Look at Technical Communication
The post explains how the author started writing, why he writes, and who he writes for.
A Practical Guide to Core Prompting Techniques for Reliable LLM Outputs
A single detailed reference breaks down the fundamental prompting methods—structured input/output, prefilling, few-shot examples, and chain-of-thought—that practitioners can use to improve model responses and curb hallucinations.
Why Machine Learning Models Fail in the Real World
A Gradient essay examines why models that seem strong in training often break down when applied to real-world data.