Week in review: Sep 27 – Oct 4
The AI bubble's finances dominated the week — Anthropic's losses, IPO numbers, and hyperscaler concentration risk — alongside a steady stream of hands-on local-AI tooling: spec-driven coding with local models, heavyweight voice cloning, and Claude Code mods.
The week’s center of gravity was the money behind the AI boom. A leaked Anthropic S1 prospectus kicked things off on the Pivot to AI channel, and Better Offline — which supplied the bulk of the volume this week as its back catalog rolled through the exporter — made the same ledger from several angles: an $8 billion operating loss on $4.6 billion of 2025 revenue, $413 billion in non-cancellable compute commitments, and revenue concentration so tight that most hyperscaler AI growth traces back to OpenAI and Anthropic spending each other’s money. Whether or not you buy the “bubble bursting” framing, the arithmetic is now public and worth reading against the hype cycle.
The counter-theme was people actually building with these tools locally. No Place Like localhost continued its spec-driven vibe coding series, hitting and then solving a very practical failure mode — a local agent exhausting its output limit on a 400-line spec until the work was broken into stages. Prompt Engineering covered Claude Code “mods,” a new extension point that lives inside the agent rather than beside it, and BreezeBlue showed how demanding even a TTS voice clone can be on consumer GPUs.
A third thread ran through Doctorow’s “LLMs are real, AI is fake” essay and Gamers Nexus’s visit to a San Francisco hacker space: the argument for owning and understanding your technology, against a media narrative that conflates a specific tool with an omnipotent one.
Notable items
- Anthropic IPO filing: the disaster in numbers — a leaked draft S1 reveals a financially precarious picture that contradicts the company’s public claims
- Monologue: Anthropic Lost $8bn In 2025, Was A Worse Business Than OpenAI | Better Offline — the underlying 2025 financials: $2.75 spent to make a dollar
- Monologue: Anthropic’s $413bn Burden — non-cancellable compute commitments that depend on borrowing $50–100bn a year from a strained bond market
- The Hidden Recession Beneath The AI Bubble w/ Paul Kedrosky — how AI capex is distorting bond markets and why data center economics don’t add up
- LLMs are real, AI is fake — Cory Doctorow’s essay on separating the real tool from the mythical one
- Spec-driven vibe coding, part 2 followup — a local agent’s 64K output limit kills a one-shot 400-line spec implementation; the fix is staged delivery
- Claude Mods - The Biggest Claude Code Upgrade! — JavaScript/TypeScript functions that live inside Claude Code and intervene in the agent, unlike hooks or MCP servers
- AI voice cloning with BreezeBlue — an impressive local TTS engine that wants 12–24 GB of GPU for what it does
- It’s Time to Hack Your Devices and Own Them — Gamers Nexus visits Noisebridge as part of a series on taking ownership of personal technology
- FTC Chair Lina Khan: The Better Offline Interview — competitiveness in tech, controlling the cloud giants, and what might give hope heading into the year
A note on volume: of 61 items this week, the great majority came from a single source — Better Offline’s back catalog (including its CES 2025 series and 2024 finale) flushing through the attention feed at 25 entries a day — so breadth across sources was narrower than the raw count suggests.