The Smartest Model in Your Stack Might Be the Smallest
Why the AI race is quietly moving to the retrieval layer — and a 350-million-parameter model just beat its bigger rivals there.
Read the briefingShowing 97-108 of 256 briefings
The Real Test for AI Agents Isn't Autonomy — It's Whether They Can Check Their Own Work
Cadence just put an autonomous AI engineer inside the chip-design loop and cut a five-week verification job to under a day. Why it works there — and stalls almost everywhere else — is the most useful lesson in enterprise AI right now.
Read the briefingTwo Roads, One Month: The EU Tightens Its AI Rulebook as Washington Moves to Tear Up the States'
In the span of a few June weeks, the world's two largest economies pulled in opposite directions on AI governance: Brussels racing to operationalize the most comprehensive AI law ever written, even as its own guidance slips, and Washington trying to stop its own states from writing rules at all.
Read the briefingThe Wire Became the Bottleneck — So AI Is Rebuilding It Out of Light
Copper is becoming the thing AI clusters trip over, so the next infrastructure race is moving light into the package: co-packaged optics and silicon photonics are turning into shipping interconnects, not just lab demos.
Read the briefingAI Designed the Molecule in Months — The Clinic Still Takes Years
AI can now design drug candidates in months, but as of mid-2026 zero AI-discovered drugs have FDA approval; the real test is still the clinic.
Read the briefingAn Open-Weights Model Just Caught the Frontier on Coding — at One-Sixth the Price
Z.ai's open-weights GLM-5.2 lands within a point or two of GPT-5.5 and Claude Opus 4.8 on coding and long-horizon agent tasks - at roughly one-sixth the price.
Read the briefingMicrosoft Put a Meter on Its AI. Then It Went Shopping for a Cheaper Engine.
Microsoft made Copilot Cowork usage-based on June 16, putting a per-task meter on agentic AI, then reports emerged it is weighing a cheaper DeepSeek V4 engine.
Read the briefingWhen the AI Starts Building the AI
Anthropic says Claude now writes 80%+ of its own code - and is floating a coordinated, verifiable slowdown on frontier AI.
Read the briefingReasoning Models Were Built to Think Longer. 2026 Is Teaching Them When to Stop.
The industry sold a simple promise: more thinking equals a smarter answer. A wave of 2026 research is rewriting that into something more useful - and a system that taught itself the lesson for about $40.
Read the briefingThe Chip Stopped Being the Bottleneck — Now It's Power and Memory
As of mid-2026, AI's binding constraint is no longer the GPU but the two things it can't run without — grid power and high-bandwidth memory. Both move on the timescale of substations and packaging lines, not software, and the squeeze is now showing up on consumer RAM price tags.
Read the briefingWorld Models Grew Up: AI Stopped Generating Scenes and Started Predicting Actions
NVIDIA's Cosmos 3, DeepMind's Project Genie, Waymo's driving simulator, and World Labs' spatial tools point to the same 2026 shift: models that predict actions a machine can execute, not just scenes a human can watch.
Read the briefingThe Hardest Part of an AI Agent Isn't the Agent
The bottleneck in agentic AI has shifted from model capability to operations. In 2026 vendors converged on the 'agent gateway' control plane, backed by MCP and A2A interop standards, to drag agents from impressive demos into auditable production.
Read the briefing