10 posts.
Passwords ask people to carry a secret from one website to another. Passkeys change the design: the device proves it holds the right cryptographic key without handing a reusable secret to a look-alike site.
Language models learn by guessing the next token, billions of times over. The surprise is not that it works - it is how much of intelligence lives inside that one narrow task.
Claude Opus 5.5, GPT 5.6, Gemini 3.7, Grok 4.6, DeepSeek V4, and two dozen more. Every major model of 2026 placed on one tier list, from S-tier flagships to F-tier forgettables.
As of 2 August 2026, the AI Office can fine providers up to 35 million euros or 7 percent of global turnover for prohibited practices. Article 50 transparency duties are live. Here is what that means for anyone shipping AI in Europe.
MCP turned the m-times-n integration problem into m-plus-n by giving every AI agent one standard way to find and call tools. Two years after open-sourcing it, it is the default way agents touch the outside world.
A 70B model is 140 GB at FP16 and about 35 GB at 4-bit — the difference between an eight-GPU rack and one consumer card. How the math works, why it barely hurts quality, and how to pick a format.
NYC Local Law 144 forced ~200 companies to publish bias audits for hiring tools. The results, the lawsuits, and what Colorado's broader law will do next.
Open 70B models now match last year's flagship APIs on many tasks at a tenth the cost — if you have the GPUs and the use case fits. A decision framework with real math.
California's SB 1047 veto, Colorado's algorithmic discrimination law, New York's automated employment decision tools rule — states are running the real policy experiments on AI. A federal framework now would freeze half-baked approaches. The laboratories of democracy are working; Congress should watch, not preempt.
NVIDIA's new flagship model isn't chasing benchmark leaderboards — it's built for the actual work AI does now: multi-step reasoning, tool use, and autonomous operation. Here's what changes when the training objective shifts from 'answer well' to 'act reliably.'