Free home repairs to train robots, Cloudflare cuts AI review costs 68%, Meta's thin orchestrator, US sanctions ex-YC China, and Ramp seeks a bank charter
Hello, dear TEA-mates! Here is what you need to know today.
1. 🔧 AI Lab Offers Free Home Repairs to Train Robots
HUMXN, an AI data and training company, is paying for select plumbing, electrical, HVAC, air-duct cleaning, and carpet-cleaning appointments in exchange for recording the work to train AI robots. Through partner home-service companies in Minneapolis, Chicago, and Miami, homeowners get qualifying services at no cost while technicians use "wearable capture technology" that records each job from their point of view. HUMXN turns those recordings into datasets for AI and robotics companies building systems that can understand, and eventually perform, physical-world tasks. The company describes a three-sided model: homeowners get free service, home-service companies are paid for the work, and robotics developers get hard-to-source real-world data. Founder Thomas Unise said robots "need to understand how skilled people interact with the physical world." HUMXN says the city mix is deliberate, since climates, building types, and equipment differ, and it plans to expand into more markets and trades. (Read More)
🫖 TEA For Thought: "The era of uploading knowledge is over, and the era of uploading humans has just started!"
2. ☁️ Cloudflare Cuts Its AI Code Review Cost by 68%
Cloudflare says the average cost of its internal AI code review rose from $1.19 in April to about $2.46 by the end of August, driven mostly by model changes, including moving every reviewer onto Anthropic models in early August. Starting in mid-September, the team had its agents trace where the tokens were going. The average review now costs $0.77, down 68% from $2.43 at its peak, and the median review time fell from 4m 09s to 1m 32s. The biggest single fix was cache reuse: a shared prompt prefix lifted first-request cache hits from 14 to 26% up to 91 to 99.5%. Smaller changes, such as removing tools reviewers could not use, reading files in parallel, and skipping reviewers with nothing new on re-reviews, cut about $0.20 per review. Code changes saved $0.65 of the $1.67 drop, while model and effort changes saved $1.06, including moving to Opus 5.5 at medium effort and to Sonnet 5.5, which was 47% to 62% cheaper per run than Sonnet 5. (Read More)
🫖 TEA For Thought: "Second Cloudflare message today. Cloudflare's observability is the beginning of the change for optimization."
3. 🧭 Meta's Thin Orchestrator: The AI Agent That Does the Least Work
Meta engineers say their AI agents for onboarding data columns across Meta's recommendation systems became more reliable once the central agent was given less to do. Their first version, a single agent loaded with all the domain knowledge for 10+ data sources, started hallucinating in production, mixing up naming conventions and code patterns across subsystems. Another single-agent approach succeeded 70% to 80% of the time with a custom prompt per task, but only 10% to 20% without one. The fix was a 1,300-line deterministic orchestrator that does zero domain work and routes 20+ workflows to 14 specialized subagents, each with isolated context and knowledge files loaded on demand. Only one step of its fixed 7-step protocol uses LLM reasoning. New use cases, such as pipeline migration, are added as workflow templates that reuse existing agents, with no new agent code. Column configuration prep that took three to four days by hand now takes under an hour, and about two-thirds of changes land without significant human edits. (Read More)
🫖 TEA For Thought: "This is very, very helpful: 'Draw the deterministic/probabilistic boundary explicitly. For every component, decide: Is this a routing decision (deterministic) or a reasoning task (probabilistic)? The orchestrator should be on one side, the subagents on the other. Don't mix them.'"
4. 🚫 US Sanctions MiraclePlus, the Former Y Combinator China
The State Department on October 19 sanctioned five China-based entities tied to Linwei Ding, the former Google engineer convicted of stealing Google's AI infrastructure trade secrets. The list includes Shanghai Zhisuan Technology, the AI startup Ding founded in 2023, and MiraclePlus, the Beijing venture firm that gave Zhisuan seed money for an ownership stake, plus two MiraclePlus-managed funds and a fund subsidiary. MiraclePlus grew out of Y Combinator China, which YC opened in 2018 under former Microsoft and Baidu executive Qi Lu and shut down in November 2019. Prosecutors say Ding uploaded more than 1,000 confidential files, including TPU designs, GPU system architecture, and networking components, to a personal cloud account. A judge later acquitted him of economic espionage, and on September 1 he was sentenced to 12 months minus a day on trade secret theft counts. The sanctions, issued under the Protecting American Intellectual Property Act, freeze assets and limit US credit and investment, and mark the law's first use against China. (Read More)
🫖 TEA For Thought: "Big move, much needed!"
5. 🏦 Ramp Applies for a National Trust Bank Charter
Corporate spend management fintech Ramp applied to the Office of the Comptroller of the Currency (OCC) for a national trust bank charter on Friday, the company said Monday in posts on X and LinkedIn. If approved, the trust bank would offer fiduciary investment management and digital-asset custody, but would not take deposits or make commercial loans. Ramp said it would bring these services "under a consistent federal regulatory framework." The filing came a day after Ramp announced a $1.85 billion funding round led by Dragoneer Investment Group, valuing the company at $60 billion. Block, Avant, and Modern Treasury have also sought OCC charters over the past two months. The rush is contested: the Independent Community Bankers of America sued the OCC on October 2, noting the agency has approved or conditionally approved 21 national trust bank charters since the start of President Trump's second term, 13 of them for crypto companies. (Read More)
🫖 TEA For Thought: "Ramp seems to want to be the future financial layer, not just for human beings but for AI agents as well. After all, the future of the economy is a machine economy."
🛠️ Skill of the Day
The Rules vs. Judgment Split: Breaks a messy recurring task into steps and shows which ones need a fixed rule and which ones actually need AI judgment.
Act as a workflow designer. I want to hand part of a recurring task to an AI agent without it going off the rails.
The task: [DESCRIBE THE TASK, E.G. "TRIAGE MY INBOX EVERY MORNING"].
How I do it today: [LIST THE STEPS AS YOU REMEMBER THEM].
What a good result looks like: [ONE OR TWO LINES].
Then do the following:
1. Steps: rewrite my process as a numbered list of small, single-purpose steps.
2. Label each step RULE (same input always gets the same answer, like sorting, routing, or checking a field) or JUDGMENT (needs reading, weighing, or writing).
3. For each RULE step, write the rule in one plain sentence so a checklist or simple script could follow it.
4. For each JUDGMENT step, list the only information the AI needs to see, and nothing extra.
5. Checkpoints: mark where I should review the output before the next step runs.
Rules: never mix a RULE and a JUDGMENT in the same step. Use plain language. If a step is unclear, ask me instead of guessing.Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own task and steps.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: