Week in review: Oct 18 – Oct 25
AI's money meeting reality — Anthropic's underwhelming quarter, CoreWeave cash-outs, and warehoused GPUs — alongside a fair-use reversal for AI training and a growing verification bottleneck in agentic coding.
Honest volume note first: the rollups carried roughly 65 items this week, but the bulk of the Oct 18–19 files is archive backfill — the exporter walked the Practical AI catalog back through its 2018–2019 seasons, so those days read like a time capsule of BERT-era NLP and GAN hype. The genuinely fresh items are a much smaller set, and they hang together tightly.
The dominant arc was AI’s economics meeting reality. Ed Zitron’s monologues bookended the week: Anthropic’s big Q3 revenue underwhelmed an already-nervous market, and Amazon is warehousing billions of dollars of GPUs NVIDIA forced into early delivery. David Gerard supplied the sharpest data point — CoreWeave’s founders have now cashed out over $3.3 billion post-IPO, whatever happens to the company in the bubble. Alongside the money story, the legal ground shifted: the US Third Circuit held that AI training is not fair use after all, and Meta drew fire for trying to grab news publishers’ content for training for free.
The engineering thread was about trust and verification. Dave Farley argued the industry has converged on the same conclusion — generating AI code is cheap, specifying and verifying intent is the bottleneck — while Google’s Harness² work lets agents rewrite their own operating harness per task without retraining. Security showed both sides of that coin: a single ASUS router auth-bypass potentially hits 25–35 million devices, and VulnCheck’s exploit team says many AI-generated proof-of-concept exploits simply don’t deliver.
Notable items
- Monologue: Anthropic Slows Down — Anthropic’s big Q3 revenues underwhelmed an already-nervous market, and the real AI numbers keep looking worse.
- CoreWeave founders cash out — the rest is mop-up — over $3.3B in post-IPO stock sales means the founders secured their wealth regardless of what the AI bubble does to the company.
- Monologue: Amazon Is Warehousing Billions In GPUs — NVIDIA created an artificial supply crisis by forcing hyperscalers to take early delivery of chips they can’t yet turn on.
- 20261016 - US Third Circuit: AI training not fair use after all — an appeals court reversed the fair-use assumption for training data, leaving Anthropic on tenterhooks.
- 20261020 - Facebook tries to grab all news publisher website content for training for free — Meta can’t simply write itself permission to take publishers’ work.
- The ASUS + LG Nightmare Scenario — a router auth-bypass discovered by Mr. Bruh could hand admin control of an estimated 25–35 million devices.
- Exploit Development, Agentic Vulnerability Discovery, and Fake PoCs — AI is accelerating vulnerability research, but many agent-generated proof-of-concept exploits fail to deliver.
- The Industry Is Catching Up: The Only Way to Generate AI Code That Works — Bob Martin, Kent Beck, and Anthropic all converge on the same point: precise, automated verification is the real bottleneck.
- Harness² Explained: Self-Improving Agents Without Training — a framework where an agent rewrites its own prompts, skills, and control loops per task without retraining the model.
- Google’s GPT-3 Moment for Time Series & Forecasting (TimesFM 3.0) — a forecasting foundation model trained on over a trillion time points — “autocomplete for charts” with uncertainty bounds.
From the blog
- Week in review: Oct 11 – Oct 18 — last week’s digest, covering Micron’s memory windfall and the Linux kernel’s AI-slop problem.