Justin
@justinhjohnson
Executive Director @ AstraZeneca | Nexus of Data, Science, Tech | Global Business Leader | Top Data Science Voice | #datascience #AI #buildinpublic #indiehacker Blog rundatarun.io BuildInPublic jandsgroupllc.com
"A CPU-only PyTorch wheel installs without a single error and runs twenty times slower in total silence."
Everyone in AI is talking about "the loop" right now. Anthropic shipped it as a feature. The slogan going around is that the winners will not have the smartest model, they will have the best loop.
There are roughly 200,000 eye specialists on Earth, and well over a billion people living with diabetes and high blood pressure, the diseases that take sight before anyone catches them. The math does not work.
People keep asking how I stay current on AI. For a long time I answered badly. I said I read a lot, which is true and completely useless to the person asking.
"The first time you see an engineer build something in 45 minutes that would have taken a week a year ago, but then see it not ship for another 6 weeks, you will be radicalized."
NVIDIA let us borrow some iron: eight H100 GPUs on a grant clock that doesn't stop. So the test isn't how fast the node is, it's how full we keep it.
The new AIXplore is live. I rebuilt it from the ground up into a lab notebook rather than a blog: interactive widgets you can drive, Tufte-style margin notes for citations and asides, and side quests that open the deep technical detours without cluttering the main read.
New AIXplore: reasoning post-training has split into two camps, and the split is the reward, not the architecture.
I'm on sabbatical. For most people that means rest. For me it means building full-time for the first time in years.
A post claimed autonomous AI research running on 800 markdown skills. We run about fifty. My first reflex was to scroll past, because in agent systems a big skill count usually means someone confused inventory for capability.
Inside one week, Perplexity and Anthropic shipped the same architectural bet from different directions, and it's the one worth internalizing now.
A year ago I wrote up Amazon Kiro vs Claude Code as two opposite bets: Kiro forced process on you, Claude Code stayed out of your way. You picked the side you believed.
There's a small robot on my desk. You say "Hey, Reachy" and it turns its head, thinks for a second, and answers out loud.
Busy week away from the usual headliners. Google shipped DiffusionGemma (text generation by diffusion, ~4x faster) and Gemma 4, Microsoft revealed its MAI frontier models, and Bezos raised $12B for an 'artificial general engineer.' All in today's wire.
I asked Claude to build a new capability for ARIA, the autonomous research agent I keep running. A way to train a model across many clinic sites without the patient data ever leaving the building.
I built a small thing this week that I thought was clever. One of my automated routines now reads back over its own run, asks what it just learned that its instructions didn't already cover, and edits its own checklist so the next run is a little sharper.
🧵 4/4 This matters for production AI systems where reliability is non-negotiable. Read the full announcement: bit.ly/42SiOHO @OpenRouterAI @AnthropicAI @OpenAI #AIInfrastructure #LLMOps
This pattern—from notes to knowledge—is a fantastic starter project for anyone AI-curious. It takes a bit of elbow grease, but the payoff is huge. I wrote a full guide on how you can build your own. 👇 bit.ly/41J3lcl 4/4
Action item: Re-evaluate your hosting strategy. Optimize for speed, cost, and quality. Independent benchmarks are key to cutting through the noise. Full data: bit.ly/414UUIj
New benchmarks reveal hosting impacts AI model performance more than expected. gpt-oss-120B varies up to 10% on reasoning tasks (GPQA Diamond x16, AIME 2025 x32) across providers. Top: @PascalAI (78.8%), @TogetherAI (76.5%), @FireworksAI (74.3%). Bottom: @Azure (70.6%), @awscloud (70.3%). 🧵 1/4
The pace of AI innovation is relentless, from agent frameworks to healthcare breakthroughs. This week's research highlights a push for more powerful, reliable systems. Read the full recap: bit.ly/3J8nYIP #AI #LLM #AIResearch 🧵 4/4
Bottom line: We're moving from Q&A to execution. Prompts → Protocols Questions → Workflows Descriptions → Prescriptions Master context engineering once. Deploy anywhere. This is different thinking than we've done before. Full post: bit.ly/4lk9jaM 🧵 4/4
Explore how claude-code-router can transform your development workflow. Learn more here: bit.ly/45oRQb0 @AnthropicAI @GitHub #AICoding #DeveloperTools #Infrastructure 🧵 4/4
To realize this, Meta has formed Superintelligence Labs, recruiting top talent like former OpenAI leaders and investing billions in massive compute infrastructure to lead the AI race. More: nyti.ms/4ldwt2D 🧵 3/4
Open-source AI changes everything. As Hugging Face's Florent Daudens notes: when AI tools aren't "locked behind trillion-dollar data centers" but distributed through open models, "capital concentration becomes much harder to maintain." bit.ly/4obNFbm 🧵 3/4