Daily TEA – Night Shift Agents, Improv Training Data, AI as Utility and More
agentic night shift, improv actors for AI, intelligence as utility, battlefield AI data, the last mile stays human
Hello, dear TEA-mates! Here is what you need to know today.
1. 🌙 Night Shift: Spec by Day, Let Agents Execute by Night
Jamon Holmgren designed “Night Shift” — a human-agent workflow that eliminates babysitting and delivers 5x faster output with better quality. The core principle: human time and energy are expensive; agent tokens are cheap. Day shift (human): interface with people, gather requirements, write detailed spec documents. No agent running, no interruptions, sustainable pace. Human clocks out, locks computer. Night shift (agent): autonomous execution with zero human steering. The agent loads a process guide and documentation router, picks tasks (bugs first, then specs), develops a testing plan, runs sub-agents as review personas — Designer, Architect, Domain Expert, Code Expert, Performance Expert, Human Advocate — implements against strict linting, type checking, and testing, then commits with detailed messages and loops until all specs are complete. Morning: human reviews the changelog and each commit. The critical feedback loop: don’t just fix bugs — analyze why the agent erred, improve docs and validation so it makes better decisions next time. Results after one month: far less babysitting, more deep thinking, zero context switching, and compound daily improvement. (Read More)
🫖 TEA For Thought: This is the dream workflow. You think deeply during the day, write specs that capture your intent, and wake up to finished work. The key insight: the feedback loop compounds. Every morning you’re not just reviewing code, you’re training your agent to be better tomorrow.
2. 🎭 AI Companies Are Hiring Improv Actors for Emotion Training Data
Handshake, a training data provider for OpenAI and other labs, posted job listings for improv actors, sketch comedians, and performers to generate training data on “human tone and emotions.” The role: get matched with other performers over video, receive a light scenario, and improvise. Requirements: authentic emotional expression, ability to recognize and shift emotions, and “interactions that feel grounded, human, and fun.” Pay averages $74 per hour, though workers report rates declining post-signup. Context: Handshake demand tripled in summer 2025, reaching a $150M+ run rate by November. AI labs are chasing multimodal models, text plus image plus video plus voice, and the latest inflection point is voice models that need realistic emotional inflection and conversational grounding. The sessions are unscripted and open-ended, essentially training LLMs on human conversational dynamics. Reddit reactions ranged from “dystopian” to “live improv will resurge as people want real, unpolished entertainment.” (Read More)
🫖 TEA For Thought: World models are the next frontier of AI, understanding physical reality, social dynamics, emotional nuance. Everyone is rushing to gather the data to get there. It might still be 3 to 5 years out, but in the era of AI, “soon” keeps arriving sooner than people thought.
3. ⚡ Sam Altman: Intelligence Will Be a Utility Like Electricity and Water
Sam Altman, speaking at BlackRock’s Infrastructure Summit on March 12, articulated OpenAI’s endgame: intelligence as a utility, metered by token usage like electricity or water. “Our business and every other model provider’s is going to look like selling tokens.” The constraint: compute capacity determines access. If OpenAI cannot build enough capacity, either they cannot sell it or prices skyrocket — pushing AI toward the wealthy or forcing governments to allocate scarce compute. (Read More)
🫖 TEA For Thought: Intelligence is becoming a commodity. The entire competitive advantage shifts from model capability to infrastructure scale and energy availability. The question is no longer “who has the best model” but “who can deliver the most tokens.”
4. 🛡️ Ukraine Opens Battlefield AI Data to Allies in World-First Move
Ukraine announced a world-first: opening real battlefield AI training data to allied nations and defense companies. The data flows through an AI platform inside Ukraine’s Ministry of Defense Center for Innovation. The core resource: millions of annotated frames from over 5 million drones across tens of thousands of missions, covering hundreds of weapon types, unit formations, and targeting tactics，continuously updated in real-time from an active high-intensity conflict now in its fifth year. Partners gain access to labeled photo and video from live combat, the most operationally rich dataset globally. Security follows NIST standards with Big Four annual audits; the platform isolates training data from sensitive military databases. The win-win framing: partners get training data they could never replicate in a lab; Ukraine gets faster autonomous capability development. US Defense Secretary Hegseth called in January for unrestricted military AI integration. Ukraine’s Deputy Defense Minister put it plainly: “The highest risk is the absence of information.” (Read More)
🫖 TEA For Thought: The future of warfare belongs to autonomous systems. Over a dozen countries are in active conflict right now, and militaries globally are accelerating autonomous defense investment. Warfare is becoming a technology battle first, kinetic second.
5. 🧩 AI Won’t Replace Innovation — The Last Mile Remains Human
Aspiring for Intelligence counters the displacement narrative: AI will not replace innovation, but reshape it. Historical precedent is clear — the printing press seemed to threaten scribes but created publishing, journalism, and education; Ford’s assembly line seemed to displace workers but created the automotive ecosystem; the Green Revolution reduced US agriculture from 40% to 2% of the workforce while expanding food production. AI follows the same trajectory: intelligence becomes abundant, new industries emerge. Four barriers prevent frontier labs from capturing all value. First, AI Harnesses — turning models into reliable systems requires routing, sub-agents, context retrieval, guardrails, and evaluation loops. Second, Product Taste — raw capability is not a product; UX, defaults, and workflows differentiate. Third, The Last Mile — reliability in messy real-world systems, integrations, edge cases, iteration. Startups focused on one problem for one segment outcompete general platforms. Fourth, Behavioral Change — users need education, trust, and participation. The outlook: an ecosystem, not a monopoly. Frontier labs build infrastructure; thousands of startups will turn intelligence into real-world solutions. (Read More)
🫖 TEA For Thought: There is still hope, folks. Delivering the last mile for the customer — that’s where the real value lives. Intelligence is becoming infrastructure. What you build on top of it is what matters.
Prompt Tip of the Day
Get richer analysis by making AI argue with itself using Perspective Prompting. Instead of asking for one answer, force the model to examine the problem from multiple expert viewpoints before synthesizing.
“Analyze [your topic] from three different expert perspectives: (1) A skeptic who sees the biggest risks and flaws, (2) An optimist who sees the greatest opportunities, (3) A pragmatist focused on immediate next steps. For each perspective, give 2-3 key points. Then synthesize: where do all three agree? Where do they disagree? What’s the smartest path forward given all three views?”
This works because a single perspective creates blind spots. Forcing three viewpoints surfaces trade-offs and nuances that a one-shot answer misses — especially useful for strategic decisions, product direction, or evaluating new technology.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X