Astra's machine-checked proofs, DeepMind's whole-body robots, a rebuilt agent protocol, and the EU's transparency deadline — July 27 to August 2, 2026. Some weeks in AI are all launch trailer. This one was mostly receipts. Between July 27 and August 2, a model produced mathematics you can machine-verify, a robotics stack learned to use its legs, the protocol underneath most AI agents got rebuilt from the floor up, and the EU's transparency rules stopped being hypothetical. He

Anthropic, Google, and a Chinese open-weight lab all made frontier-grade AI cheaper in the same seven days — while the money behind it got a lot bigger. If you blinked this week, you probably missed a model launch. The back half of July has become the most crowded stretch the AI industry has ever seen, with new systems arriving every few days. But this week had a theme worth pausing on, and it wasn't raw capability. It was price. Three of the most powerful models on the plane

A record-breaking open model out of China, a very public scramble for chips, and even Google quietly hitting pause. This week's AI story wasn't a single jaw-dropping demo. It was about plumbing — who has the compute, who can afford it, and who is willing to admit their next model isn't ready. The releases that mattered said as much about electricity and capital as they did about clever architecture. Here are the developments from the past week that actually move the picture.

In a single week, OpenAI, xAI, and Meta all shipped new flagship models — and every one of them was really shipping an agent. For most of the last two years, frontier model launches arrived one at a time, spaced far enough apart that each got its own news cycle. This week they collided. Between July 8 and July 9, three of the largest AI labs pushed new flagship systems into public hands inside roughly forty-eight hours. Read past the version numbers and they were all shipping

This week the frontier labs kept shipping cheaper, sharper models — while Washington quietly moved to the center of who gets to use them. The last seven days were a study in contrast. Anthropic, OpenAI, and xAI all pushed new models out the door, each promising more capability for less money. But the more revealing story was not in the benchmarks — it was in the access. For the first time, the question of who gets a frontier model first is being answered as much in a governme

Google lost its stars, OpenAI taped out its first chip, and Mistral quietly turned document reading into an enterprise weapon — the week the AI race moved down the stack. Some weeks bring a shinier model that nudges a benchmark a few points higher. This past week brought something more structural. The action moved away from chatbots and one layer down the stack — to the people who build frontier models, the silicon that serves them, and the unglamorous tooling that feeds them

Most explanations of a convolutional neural network are diagrams and equations. I wanted the opposite — something you can watch. So I trained a small CNN to recognize handwritten digits, then rendered it in 3D so you can see every step as an image flows through it. Explore the interactive version → Filters: tiny pattern detectors A CNN learns small 3×3 filters — little grids of numbers, each one a pattern detector. One fires on horizontal edges, another on diagonals, another

Every explanation of the Transformer — the architecture behind ChatGPT, Claude, and Gemini — seems to come as one dense diagram and a page of equations. I spent years teaching before I moved into machine learning, and I always wanted the opposite: something you can watch. So I built it: a real, trained model rendered in 3D, with a short walkthrough of every moving part. Explore the interactive version → Two halves: encoder and decoder The original Transformer has two stacks.

Google ships small and open, Anthropic's flagship gets a government stop order, and the big models keep quietly marching into government and the enterprise. This past week in AI didn't turn on one giant model launch. It was a week of contrasts instead: open weights getting smaller and faster, a frontier lab running headfirst into export controls, and AI quietly clocking in for serious work in government and the enterprise. Here are the developments worth your attention, and w

Transformers power ChatGPT, but the attention mechanism inside them is far more general. Here is what a Transformer really is, why it works on images and proteins as well as text, and why a Transformer is really a graph neural network on a fully-connected graph.

Backpropagation explained step by step — the forward pass, the loss, the chain rule, and gradient descent — with clear diagrams and the key formulas.

RAG vs. long context isn't a rivalry — it's a decision. A practical heuristic for when to retrieve, when to just paste it, and why the best systems do both.

Bigger context windows aren't free. Why attention scales quadratically, the 'lost in the middle' problem, and how to be deliberate about what you feed an LLM.

Two fraud models with nearly identical ROC-AUC, one with 4x more precision in the real operating zone. Why PR-AUC, capture-at-FPR, and expected dollars are the only metrics that matter at low base rates.

Two years ago, the playbook for serious LLM work was simple: you wanted Llama 2 70B, or you used a frontier API. Anything in the 7B range was a toy. That playbook is broken. As of mid-2025, a competent 7B open-weight model outscores Llama 2 70B on standard benchmarks, and a 14B model now matches what Llama 3 70B did six months ago. I want to walk through what changed and what it means for anyone deploying models in production. The numbers Looking at MMLU 5-shot — the standard

You can date an ML engineer's career by their panic vocabulary. There is a stable trajectory from "model converged, time for coffee" all the way down to "the boss saw the AWS bill." Most days you'll move up and down it like Maslow on a bad Monday. I want to walk through the levels, because naming the panic is the first step to managing it. ZEN — "Training is running. Loss is dropping." The myth of stable training. You set up your run, you tail the logs, and the loss curve des

When you open Netflix, Amazon, or Spotify and the next thing it suggests feels suspiciously right, what's running under the hood isn't what most people imagine. For a decade, recommendation systems lived on matrix factorization and collaborative filtering. Quietly, that has changed. The same transformer architecture that revolutionized language models has taken over sequential recommendation, and the benchmark gap is no longer subtle. I want to walk through how we got here, w

Knowledge distillation has a well-known shape: a big teacher model trains a smaller student model of the same family. GPT-4 distills into GPT-4-mini. ResNet-152 distills into ResNet-18. Tree-based ensemble distills into a single shallow tree. Teacher and student speak the same architectural language. What happens when they don't? I built a systematic study of cross-paradigm knowledge distillation — bidirectional knowledge transfer between Random Forests and Deep Neural Networ

Every benchmark blog you read about transformers tells you they scale. The truth, when you actually run 118 of them end-to-end, is that 51% of models fail in the transition from 512 to 1024 tokens, and 0% of them work at 2048. That's not a 'newer models are getting better' story. That's a hard ceiling that hits production deployments far below the context-length numbers vendors advertise. I ran the systematic benchmark to map it. Full code and data: transformer-scalability-wa

Choosing which LLM to deploy in production is mostly vibes-based. You read benchmark leaderboards, pick the highest-scoring model that fits your budget, and ship. Most of those benchmarks tell you what the model can do, not how it processes the task internally. I ran a study to look inside. The full code and data are at llm-neural-activation-patterns — a systematic comparison of how six different LLM architectures process twelve cognitive task categories, measured across 144