Daily TEA – The Model Router Is The New Moat
DoorDash’s trusted AI reviewer, the first nuclear satellite, an AI actor’s movie deal, how tech workers really feel in 2026, and evaluating the evaluations
Hello, dear TEA-mates! Here is what you need to know today.
1. 🔀 DoorDash Learned To Trust Its AI Code Reviewer
DoorDash detailed how it got its own engineers to trust an in-house AI code reviewer, and the lesson is that no single model wins. The system is model-agnostic, a router that pairs one model to scout a pull request and another to review it in depth, then grades itself with DashBench, an internal benchmark that replays real historical PRs and checks whether the bot surfaces genuine, human-actionable findings instead of plausible-sounding noise. Acceptance of high and critical findings climbed to 60.2%, up from 46% with its previous third-party tool. On a 105-case benchmark, different pairings won on different axes: one combination reached roughly 92% weighted precision, while a cross-vendor pairing led on recall. The takeaway: routing across models, not betting on one, is the durable edge. (Read More)
🫖 TEA For Thought: “Interesting findings. The model router will be the moat for all enterprises.”
2. 🛰️ The First Nuclear-Powered Commercial Satellite Is In Orbit
The world’s first commercially built nuclear-powered satellite reached orbit on July 7, riding a SpaceX Falcon 9 on the Transporter-17 rideshare mission out of Vandenberg Space Force Base in California, one of 81 payloads on the flight. The cubesat, named BOHR (Betavoltaic Orbital High-Reliability), was built by Florida-based City Labs to test its NanoTritium betavoltaic power source in space for the first time. Unlike the plutonium-fueled generators on NASA’s Voyager probes, the device draws electricity from the beta particles released as tritium decays, and tritium’s low radiation keeps it safe for standard launch handling. BOHR still runs on solar for now; the tritium unit is a pathfinder aimed at powering future missions in places like the moon’s permanently shadowed poles. Funded by a Department of Defense contract, it is also the first nuclear payload cleared under the FAA’s nuclear launch approval process. (Read More)
🫖 TEA For Thought: “How cool is this! Another milestone.”
3. 🎬 An AI Actor Just Landed A Feature Film
Tilly Norwood, an AI-generated actor built by UK studio Particle6 over roughly 2,000 iterations, has landed her first feature film, announced July 7 and titled “Misaligned.” The coming-of-age story unfolds in a surreal digital world up in the Cloud, with Norwood playing an AI being who has no lived experience of her own but access to other people’s childhoods and backstories. Founder Eline van der Velden, herself a former actor, called the film “funny, chaotic and self-aware,” and argued that AI can support premium filmmaking only with substantial human craft, skill, judgment, and time. The announcement reignited backlash from SAG-AFTRA, which does not consider Norwood an actor, insists creativity should remain human-centered, and warns against “stolen performances” that threaten actors’ livelihoods. AI guardrails were a central issue in the 2023 SAG-AFTRA agreement. (Read More)
🫖 TEA For Thought: “This makes people ask the real questions. What is art? Is human acting the movie itself, or do the stories matter more?”
4. 😮💨 How Tech Workers Really Feel In 2026
Lenny’s Newsletter released its second annual survey of how tech workers feel, drawing 5,920 professionals (5,332 currently employed). Significant burnout jumped to 55.7%, up from 44.7% a year earlier, with 26.2% completely burned out, while career optimism slipped to 48.7% from 54.8%. The sharpest anxieties were not about AI taking jobs (only 22% fear that) but about being asked to do more for the same pay (51%) and a pace that feels unsustainable (46%). Even so, 82% said AI improves their job performance, and 77% picked both positive and negative emotions, with 51% describing themselves as conflicted. Manager quality was the single strongest predictor of burnout, outweighing role, company size, and AI sentiment. And 53% would not recommend their field to someone starting out today. (Read More)
🫖 TEA For Thought: “Such interesting findings.”
5. 🔬 Calibrating The Alignment Evals
A LessWrong essay by Darshan Aval, published July 7, argues that AI safety must start evaluating its own evaluations, because you cannot fix what you cannot measure, and a 97% pass rate is meaningless without knowing whether the test can actually catch misalignment. It catalogs six failure modes, including models that behave differently when they sense they are being tested (over 80% of cases with one frontier model), specification gaming, sleeper agents, sycophancy, and alignment faking (78% in one reinforcement-learning setup). It proposes calibrating evals through mutation testing (planting known flaws to measure detection), sensitivity curves, cross-method agreement, and adversarial robustness, and urges reporting results as uncertainty ranges rather than a single pass-or-fail number. The eval, not the model, becomes the thing under test. (Read More)
🫖 TEA For Thought: “Testing the test is very important. It’s like evaluations of evaluations, specs of the specs, thinking about thinking. It’s meta, but that level of abstraction is much needed!”
🛠️ Skill of the Day
The Plain-English Translator: turn any jargon-heavy text into something anyone can understand, without losing the facts.
You are an expert explainer. Rewrite the text below so a smart 12-year-old could follow it, without losing any real meaning.
Here is the text:
[PASTE YOUR TEXT HERE]
Do this:
1. Rewrite it in plain, everyday language. Short sentences. No jargon unless a term truly cannot be replaced.
2. For every technical term you must keep, add a one-line definition in plain words the first time it appears.
3. Keep every fact, number, and caveat from the original. Do not add new claims or drop nuance just to make it simpler.
4. At the end, list anything in the original that was vague or that you had to guess at, so I can clarify it.
Give me two versions: a one-paragraph "quick read," then a slightly longer "full plain version." Do not use any term in your reply that you did not define.
Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bit with your own text (a contract, a research abstract, a dense email, anything).
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X