Eric J Ma's Website

When my team was stuck talking past each other with no slam dunk option in sight, I had everyone bet 1,000 points of funny money across the options, and the allocations revealed the team's collective belief distribution. It beats voting because it captures preference strength and confidence separately. Curious how a 70-point bet and a 900-point bet unblocked the discussion?

In this post, I share why Dan Chen and I are running the Learn Anything with AI retreat: a five-day, in-person gathering built on the idea that AI can either hand you answers you'll forget or accelerate genuine understanding. Drawing on my own journey from lab science to computing, I explore learning science, the cognitive debt of skipping deliberate practice, and why teaching what you learn makes it stick. The retreat week is for acquiring the skills of using AI to learn anything; mastering new fields is what those skills then deliver. What could a year of practicing those skills do?

In this post, I share what I've learned about quantum computing from the perspective of a probabilistic Bayesian. I strip away the physics to reveal a familiar picture: qubits as amplitude vectors whose squared magnitudes are probabilities, gates that shape distributions through interference, and measurement as sampling. The parallel to Bayesian inference is striking. Our priors and likelihoods shape posteriors; quantum gates shape amplitudes. What if the quantum computer isn't a magic box, but a sampler purpose-built for the combinatorial spaces we already struggle to navigate classically?

In this post, I share what I learned from a summer spent reading the research on learning with AI. The biggest takeaway? AI doesn't help or harm learning on its own; it simply amplifies the cognitive habits you bring to it. If you passively ask it for answers, it breeds over-reliance. But if you use it to retrieve, defend, and test your own thinking, it becomes an infinitely patient tutor. I map the evidence, share practical prompt swaps, and even quiz you along the way. So I'm curious: the next time you open a chatbot, who will be doing the thinking?

I map a growing consensus: now that AI agents write so much of our code, producing it is cheap, but *understanding* it is the real bottleneck. I share my own struggles reviewing forty-file diffs and explore how cognitive debt quietly accrues when we approve polished PRs we can't actually explain. Drawing from thinkers like Geoffrey Litt and Simon Willison, I highlight practical techniques to fight back, like literate diffs, interactive micro-worlds, and quizzes. (I even quiz you on this very post!) How are you keeping your mental model alive when your agent types faster than you can think?

Bayesian modeling with heavy-tailed likelihoods and hierarchical structure automates away manual data analysis tasks like hand-flagging outliers. By using Student-t likelihoods and partial pooling, your pipeline handles messy data points and weird curves in a principled, reproducible way. No clicking through software required. What could your team do with all those afternoons back?

Pair-programming with a coding agent in a marimo notebook competition gave me new ideas for using AI to help me learn ML papers. By building interactive artifacts instead of just reading, I grasped concepts like BitNet's ternary weights viscerally. The agent handled the typing, but my curiosity stayed in the driver's seat. What could you learn if jargon stopped getting in the way?

In this blog post, I share my thoughts from SciPy 2026 on how open source projects should handle AI-assisted contributions. Drawing from a community discussion and a new AI policy template, I distinguish between developers using AI as a power tool versus mass-produced slop. I encourage maintainers to pick a policy tier, write it down, enforce it, and engage with the humans behind the PRs. What AI policy fits your project?

In this blog post, I share how my idea of using xarray for unified lab data evolved into an agent skill for Ian Hunt-Isaak's repo. I explain why xr.DataTree is the key to storing multi-assay data on incompatible grids in a single file, how custom indexes let you query experiments by their actual physical logic, and why Bayesian estimates belong right next to raw data. Ultimately, I argue this pattern scales a single data scientist's judgment across many campaigns. If you are juggling multi-assay biological data, what are your biggest coordinate-matching headaches?

In this blog post, I share my experiments running Qwen3.6 on Modal with Ollama, vLLM, and SGLang. Frustrated by agonizing cold starts, I benchmarked the engines on identical hardware. Spoiler: vLLM blew Ollama out of the water. By leveraging Modal's GPU snapshots, vLLM not only generated 60% more tokens per second but also restored from a cold start significantly faster. I walk through the exact configuration, three stubborn bugs I had to fix to get snapshots working, and why vLLM's single-process design is the secret. SGLang hit a dtype bug and could not run yet. Have you ever fought with serverless GPU cold starts?