Week in review: Oct 11 – Oct 18
The exporter backfilled six years of Practical AI and Pivot to AI archives, while the fresh items — Micron's memory windfall, the Linux kernel's AI-slop problem, and Google's TabFM — replayed the archive's recurring themes of bubble economics and agentic coding.
Honest volume note first: this week’s feed looks huge — roughly 110 items — but about 100 of them are archive backfill. The attention exporter walked the Practical AI catalog from 2020 forward and cleared a year of Pivot to AI episodes from late 2025, so the rollups read like a six-year retrospective. That turns out to be its own kind of story: bubble skepticism, agent security, and private/local AI recur almost unchanged from the 2020 episodes through the 2026 ones. If you want a single arc through the archive, “Are we in an AI bubble?” (Nov 2025) sits neatly between the 2024 Gartner-hype-cycle episode and the finance press finally saying the words.
The genuinely fresh items converged on AI’s economics meeting reality. Gamers Nexus broke down Micron’s “infinite money glitch” — $54.2B in quarterly revenue riding the memory shortage created by AI demand — while Pivot to AI covered a Steam voice-AI game demo going broke on just 430 players. The sharpest cautionary tale was the Linux kernel drowning in low-quality AI-assisted patches, to the point of dropping the requirement to label chatbot-assisted commits. South Korea’s $850M failed AI textbook program, backfilled from the archive, rhymes with all of it: money spent faster than quality can be verified.
The agentic-coding thread ran through both fresh and backfilled items. Prompt Engineering showed a self-improving “software factory” where agents pick GitHub issues, fix them, and an independent verifier gates the merge; Practical AI’s Steve Klabnik episode on building his new language Rue with Claude was the week’s most honest take on what agentic coding does and doesn’t do; and journalist Evan Ratliff’s experiment staffing a real startup almost entirely with AI agents rounds out the spectrum from tooling to thought experiment. On the data side, Google’s TabFM — a zero-shot foundation model for tabular data wired into an LLM-agent pipeline with Neon Postgres — is the “GPT-3 moment for spreadsheets” claim worth stress-testing.
Notable items
- Google’s GPT-3 Moment for Databases & Spreadsheets (TabFM) — a zero-shot foundation model for tabular data, combined with an agent and Postgres to skip per-task classical ML training.
- Build a Self-Improving Software Factory — GitHub issues in, agent fixes, independent verification, human merge gate.
- Micron’s Infinite Money Glitch with Memory — how AI-driven demand and tight supply turned the memory shortage into record margins.
- Linux kernel goes full AI slop, hides chatbot commits — maintainers overwhelmed by low-quality AI patches, and the labeling requirement removed.
- AI game demo is going broke on just 430 players — a voice-controlled AI game whose costs dwarf its audience.
- Humility in the Age of Agentic Coding — Rust contributor Steve Klabnik on building a new language with AI tools, from skeptic’s perspective.
- Inside an AI-Run Company — Evan Ratliff’s immersive-journalism startup staffed almost entirely by AI agents.
- Are we in an AI bubble? — Practical AI’s dot-com parallels framing for the whole week’s economics items.
- 20251019 - South Korea blows $850m on failed AI school textbooks — “poor quality, hastily put together”: procurement-scale AI failure.
- AI at the Edge is a different operating environment — Qualcomm’s Edge Impulse on small models, cascades, and the latency/power/privacy constraints that change the whole design.