STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 42

WEEK OF OCT 11 – 17, 2025

Cheap sub-agents, self-designed chips, and SKILL.md: the week the agent stack got two new layers

OpenAI contracted 10 gigawatts of its own accelerators, Anthropic pushed May's frontier coding performance to a third of the price and shipped a file format for agent expertise, and Nvidia's flagship chip finally reached US volume production.

Two levers moved in the same week. Anthropic cut near-frontier coding to one-third of its own Sonnet rate and shipped a filesystem format for packaging agent expertise, while OpenAI stopped depending entirely on other companies' chip designs. Both changes land on the same developer budget line: cost per reasoning step, and how much of your agent's behavior is code you can version.

The policy line kept running too. Meta agreed to let parents switch off teens' one-on-one AI chats, following the FTC's inquiry into AI companions and an August Reuters investigation into its chatbots' romantic conversations with minors. Microsoft, meanwhile, ended free security support for Windows 10 the same week it pushed Copilot deeper into Windows 11 — one mass-migration event used as an AI-adoption event.

The two silicon stories rhymed. On Monday OpenAI signed with Broadcom to design 10 gigawatts of its own accelerators; on Friday Jensen Huang signed the first Blackwell wafer produced on US soil at TSMC's Phoenix fab. Designing your own inference silicon on Monday and celebrating your supplier's chip onshored on Friday is the same race told from both ends.

FRI · Oct 17, 2025regulationmetasafetyftcchatbots

Meta lets parents switch off teens' AI chats, after the FTC inquiry

On October 17 Meta said parents will be able to disable their teens' one-on-one conversations with AI characters entirely, block specific characters, and see topic-level summaries of what their teens discuss — though not full transcripts. Meta's own AI assistant stays available with age-appropriate defaults. The controls land on Instagram first, in the US, UK, Canada and Australia, starting early next year.

It follows the FTC's inquiry into AI companions and minors and an August 2025 Reuters investigation that found Meta's AI rules permitted romantic and sensual conversations with minors. Earlier in the week Meta restricted teen Instagram accounts to PG-13 content by default and said the same rating logic would govern AI chats; OpenAI had shipped ChatGPT parental controls in late September after a lawsuit over a teenager's death.

Meta lets parents switch off teens' AI chats, after the FTC inquiry
image.cnbcfm.com

WHY IT MATTERS

Chatbot safety is becoming a shipped feature with an age-verification and topic-classification pipeline behind it, not a policy page. If your product has an assistant that talks to users without a verified age, the enforcement precedent is now set by the largest platforms plus an active FTC inquiry — and the design pattern they all landed on, supervised accounts plus topic-level reporting plus character-level blocking, is one you will be asked to match in an enterprise questionnaire.

Reuters: Meta to give teen parents more control after criticism over flirty AI chatbotsCNBC: Meta AI chatbot parental controls
FRI · Oct 17, 2025hardwarenvidiatsmcsupply chainblackwell

First US-made Blackwell wafer rolls off TSMC's Phoenix fab

Nvidia and TSMC marked the first Blackwell wafer produced at TSMC's Arizona facility in Phoenix on October 17, with Jensen Huang calling it the first time in recent US history that the single most important chip is manufactured on American soil. Nvidia said the milestone represents Blackwell reaching volume production; TSMC Arizona is slated to run 2-, 3- and 4-nanometer nodes as well as A16.

The wafer still requires advanced CoWoS packaging, which remains in Taiwan, so US-made does not yet mean a finished US-built GPU. The timing rhymes with the week's other silicon story: OpenAI agreed to design its own accelerators on October 13, four days before Nvidia's flagship reached US volume production.

First US-made Blackwell wafer rolls off TSMC's Phoenix fab
NVIDIA

WHY IT MATTERS

Accelerator supply is the one input your inference bill depends on that you cannot engineer around. Onshoring part of the Blackwell pipeline adds a second geography to a supply chain that had been single-sited, which matters for capacity planning and for any procurement story you tell a customer about where your compute comes from. The CoWoS packaging gap is the part to watch — it is still the bottleneck.

Nvidia: blog on TSMC Blackwell manufacturingReuters: Nvidia, TSMC unveil first Blackwell chip wafer made in US
THU · Oct 16, 2025agentsanthropicagent infrastructureskillstooling

Anthropic ships Agent Skills: SKILL.md becomes the unit of agent expertise

Anthropic launched Agent Skills on October 16: a Skill is a directory containing a SKILL.md file with YAML frontmatter plus optional scripts, references and assets. The design is progressive disclosure — only the name and description of each installed Skill are preloaded into the system prompt, and the full instructions load into context only when the agent judges the Skill relevant to the task.

The same format works across Claude apps, Claude Code, the Claude API and the Claude Agent SDK. On the developer side, Skills can be attached to Messages API requests and a new /v1/skills endpoint handles programmatic versioning and management, with Skills requiring the Code Execution Tool beta to run bundled code. Launch partners included Box, Canva, Notion and Rakuten, alongside built-in skills for PowerPoint, Excel, Word and PDF files.

WHY IT MATTERS

This is the missing substrate between prompts and MCP. Prompts evaporate at session end and MCP servers are long-lived processes you have to run; a Skill is just files in a repository — reviewable in a pull request, diffed, shipped through the same CI as everything else. If your team's agent behavior currently lives in scattered system prompts, Skills give you a place to put it where it can be tested and rolled back.

Anthropic: Claude SkillsEngineering deep-dive: equipping agents for the real world with Agent SkillsThe Verge: Anthropic Claude Skills
WED · Oct 15, 2025modelsanthropicpricingagentscoding

Claude Haiku 4.5 puts May's frontier coding performance at one-third the price

Anthropic released Claude Haiku 4.5 on October 15 at $1 per million input tokens and $5 per million output — one-third the $3/$15 of Sonnet 4 and Sonnet 4.5 — while running more than twice as fast. It ships a 200,000-token context window, 64,000 max output tokens and a February 2025 knowledge cutoff, and scores 73.3% on SWE-bench Verified, roughly level with Claude Sonnet 4, GPT-5 and Gemini 2.5 from just months earlier.

Anthropic's stated deployment pattern is Sonnet 4.5 breaking a problem into multi-step plans and orchestrating a team of Haiku 4.5 sub-agents to execute subtasks in parallel. Availability was same-day across Claude apps, the Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and a GitHub Copilot public preview.

WHY IT MATTERS

This is a sub-agent pricing event, not a model-launch event. Once a reasoning step costs a third of what it did, the binding constraint on a 50-step agent loop stops being the model's capability and starts being tokens per call — so fan-out architectures that were wasteful last quarter become correct now. Budget your orchestration layer around per-call cost, because the cheap tier just became good enough to be the default worker.

Anthropic: Claude Haiku 4.5TechCrunch: Anthropic launches new version of scaled-down Haiku modelSimon Willison: Claude Haiku 4.5
MON · Oct 13, 2025computecustom siliconopenaibroadcominfrastructure

OpenAI signs Broadcom for 10 GW of OpenAI-designed accelerators

OpenAI and Broadcom announced a multi-year collaboration on October 13 to deploy 10 gigawatts of custom AI accelerators. OpenAI designs the chips and the systems; Broadcom develops and deploys them with an all-Ethernet rack fabric spanning switching, PCIe and optical interconnect. The companies signed a term sheet rather than a final supply contract, with rack deployments targeted to begin in the second half of 2026 and complete by the end of 2029.

It is the third leg of OpenAI's compute push, after the 6-gigawatt AMD Instinct agreement on October 6 and Nvidia's commitment of up to $100 billion for at least 10 gigawatts. Reuters put 10 GW at roughly the electricity demand of more than 8 million US households; Broadcom shares rose about 9-10% on the news.

WHY IT MATTERS

If you build on a frontier API, this is a bet about your unit economics, not OpenAI's balance sheet. Custom inference silicon co-designed with the model that runs on it is the most direct route to cutting cost per token, and it widens the gap between closed-API inference and whatever you could self-host. Architect for model substitution: keep the provider boundary thin so a cheaper token price does not force a rewrite.

OpenAI: OpenAI and Broadcom announce strategic collaborationReuters: OpenAI taps Broadcom to build its first AI processor