Eric Ma

Your AI classifier is confidently wrong, and you can't tell when. Fine-tuned open models beat closed APIs. 20 min of fine-tuning: accuracy 0.26 to 0.94. Trustworthy triage instead of confident nonsense? Read it, then tell me: would you ship this? ericmjl.github.io/blog/2026/10... #opensource #ml

A judgment language where if-conditions are plain English, billed per call? I swapped the hosted judge for an open-weight model on my machine: one file changed, $0.00 a call, deterministic answers. buff.ly/o3okrVk

If you have shipped AI into a real drug discovery workflow, this one is for you. SLAS Discovery wants practitioner case studies: what happened, what changed, what you'd do differently. Edited by Raminderpal Singh and Srijit Seal. Due Dec 15: www.slas-discovery.org/content/call...

Multiplayer AI agent = a coding agent in a container. A websocket, a subprocess, and one markdown knowledge base. Shared grounding, bounded budgets, public answers: all plain code. ericmjl.github.io/blog/2026/9/...

Went in as a GxP rookie, left with assumptions dismantled. Regulators state principles; companies write the procedures. AI validation = validating a measurement system. "Validation isn't an event. Validation is data." ericmjl.github.io/blog/2026/9/...

A model that fails for reasons you understand can teach more than a leaderboard win. A SLAS Discovery special issue welcomes negative results and failed AI deployments. Edited by Raminderpal Singh and Srijit Seal. Due Dec 15: www.slas-discovery.org/content/call...

Team stuck going in circles on a decision? Hand out 1,000 points of funny money and let everyone bet across the options. My 70-point bet vs. a teammate's 900 showed where conviction actually lived, more than hours of arguing did. ericmjl.github.io/blog/2026/9/...

Ask AI all day, remember nothing? That's cognitive debt. Answers aren't understanding. The fix is learning how to learn with AI as the accelerator, and one focused week is enough to acquire those skills. What would you learn first? ericmjl.github.io/blog/2026/8/... #learning #ai

Meta and X have been quietly downranking external links for ages, but we're now entering a new era of open hostility toward the open web. It's a great time to build your own website, free from any billionaires demanding a tithe for things that have always been free.

But, in many cases, whether measles, COVID, or climate change, bad science has serious negative consequences. In this case, it is the partial reversal of one of the great public health achievements of the last century. 🧪 #episky #medsky

What did your last AI deployment actually change? Hit rate? Cycle time? The decision made differently? A SLAS Discovery special issue asks it of real deployments in drug discovery. Edited by Raminderpal Singh and Srijit Seal. Due Dec 15: www.slas-discovery.org/content/call...

Quantum computing isn't magic, it's a sampler for spaces too big to enumerate. I mapped quantum amplitudes to Bayesian posteriors after months of 1-on-1 tutoring. Interactive diagrams included. buff.ly/SRHm9eQ

What's the last thing you asked AI to explain? Could you reproduce that from memory right now? If not, you recognized it. You didn't learn it. Here's the fix. buff.ly/GhLwDKt #ai #learning

The best AI stories in drug discovery never make it into print. A new SLAS Discovery special issue collects them: real case studies, measurable outcomes, failures welcome. Edited by Raminderpal Singh and Srijit Seal. Due Dec 15: www.slas-discovery.org/content/call...

🧪⚛️ Two terrible lab safety incidents at universities this week (HF at Hokkaido, dimethylmercury at MIT). Please be careful out there! nanoscale.blogspot.com/2026/08/lab-...

Agents write the code now; understanding it is the bottleneck. Approve a PR you can't explain and you witnessed it, not reviewed it. I mapped the 2026 zeitgeist, with a playable quiz inside: buff.ly/46T6a5g

New episode is out! A very practical one, where we dive into the new research by Adrian Seyboldt, Eliot Carlson and Bob Carpenter about how to make your MCMC samplers go faster -- lots of gems in there, make sure to tune in!

Clicking through curves to flag outliers by hand? Deleting points is statistically broken. I fit 36 plates in 4 minutes flat. How would you feel about automation instead of manual clicking? Share with your team! ericmjl.github.io/blog/2026/7/... #bayesian #assays

Reading ML papers feels like drowning in jargon. Pair-coding notebooks with an agent taught me more in an evening than a week of reading. How would you feel about playing with concepts instead of fighting syntax? ericmjl.github.io/blog/2026/7/... #ml #ai

Drowning in AI PRs? Stop arguing with bots. Pick a policy. A clear AI_POLICY.md ends the slop. How would you feel about real dialogue instead of bot relay? ericmjl.github.io/blog/2026/7/... #opensource #ai

Spending weeks merging assay files? One file can hold every assay. I built a skill that auto-assembles linked data packages. One curator manages many campaigns. How would you feel about shipping instead of wrangling? ericmjl.github.io/blog/2026/7/... #xarray #bioinformatics

Garfield was not a cartoon. He was a PROPHET. #Caturday

What if your LLM woke in seconds, not minutes? Tested Ollama vs vLLM on Modal (Qwen3.6-27B). vLLM wins on cold starts and throughput. Post: setup, 3 snapshot bugs, CUDA graphs. How do you handle cold starts? buff.ly/e8LX5OE #llm #vllm

Exactly the type of old school data science project that I love, with the help of new school tools www.iandmacomber.com/blog/gopro-g...

formats evolve over decades, apps over months, intelligence is evolving every day in terms of pace layers "file over app" defines priorities: invest deeply in proven formats, and don't expect anything at the intelligence layer to last very long

Y saltó Don Gato. 😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹 And Top Cat jumped in. 😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹😹

Does your dog bite? No, he blesses. Original post

I wonder if this will make people reconsider their stance or become even more defensive.

Does AI widen the expert-novice gap? Live-demoed a Bayesian model with an agent. It amplified my judgment. For experts, a microscope. For novices, fast wrong answers. 'Vibe zones' and trust scaling with verification. What's your experience? #ai #datascience buff.ly/BVb49jC

The Norwegian team has been a delight to watch and they represent all that is positive about football.

If you're not prepared to get your kids vaccinated, don't have children.

What if your agent learned from its mistakes without you telling it to? I built a plugin that watches sessions, extracts lessons, and patches skills. 922 observations, 200+ reviews in 10 days. It patched a skill on its own. No babysitting. #opencode #aiagents ericmjl.github.io/blog/2026/6/...

Stuck switching branches without losing half-finished work? I use git worktrees now. Multiple doors into the same repo: no syncing, no duplication, separate folders sharing one brain. One brain, many desks. What's your go-to trick? #git #programmingtips buff.ly/B8p0Xt6

Building voice-first AI apps is a wild ride. Learned a ton (and hit walls) making projects where you talk and the AI listens. Sharing my biggest lessons and surprises. buff.ly/IJmtC8l #voiceai #aidev

What is an #agentharness, really? I tried defining this term that keeps popping up across AI. Broke down tools, controls, and action space with real examples. Your take? buff.ly/Eu5qr1G #ai

📚 Learn from experts during two full days of tutorials, covering everything from Python fundamentals to advanced topics such as GPU-accelerated computing, machine learning, and building AI-powered coding assistants using retrieval-augmented generation (RAG).

💛 It's where maintainers meet first-time contributors. Where researchers discover tools that transform their work. Where students find mentors. Where hallway conversations become collaborations, papers, startups, and open-source projects used by millions around the world.

🎉 Less than one month until #SciPy2026! 👉️ ti.to/scipy/scipy2... When people ask what SciPy is, it's tempting to say it's a conference about Python for scientific computing. But SciPy has always been much more than that.

Eeyore is now immortalized on a license plate.

At some point, you have to flip the switch from "it's dangerous to let everyone have this for offense" to "it's dangerous not to let everyone have this for defense." The jailbreak is credible but IMO the decision wasn't substantive. Anthropic created glasswing to prepare for this exact need.

Saving this to my zotero so I can reference it across my work and organize my library aka using the devil machine for lazy liars according to bsky historians

TOMORROW!!! #NoKingsDay 🇺🇸

GLM 5.2 Is Out https://digg.com/tech/ii9xibgn (https://news.ycombinator.com/item?id=48518684)

Wow, technical wizardry. My brain kinda broke seeing lua code/C++ being shipped as a Python package...

🎉 Scikit-learn 1.9 released: ■Solid improvements to many existing estimators: faster, more stable, handling missing values, adding GPU support… ■Also, enhanced estimator displays in notebooks, ■And callbacks that enable progress bars or monitoring of convergence blog.scikit-learn.org/updates/rele...

A SECOND person died of exposure last winter after DHS dumped them in the middle of nowhere, miles from their home, without notifying their family.

This is so Artistic 😊