Fieldnotes is a weekly read for the people inside companies who shape how AI gets used. Learn more about Superadditive.
This week’s main thing
I sat around an actual campfire last weekend, talking with a group of fellow parents who live and work in SF, and of course the topic of AI came up. Afterward, when I was playing back the conversation in my head (because I’m cuckoo like that), I really didn’t find my own thinking on AI to be very sharp. Partly it’s because we darted from headline to headline (“did you hear about this?”), but partly because that’s how all of us are framing our thinking on AI these days. By collision. Like we’re falling through a tall tree, whacking ourselves on every branch, and using the lumps and scratches as the means of forming a coherent point of view.
This newsletter has been doing that since it started. Here’s what to know this week. Here’s what to do. Here’s me performing expertise.
Can we try something else this week?
I built a thing. First for me, and then for you. For you to slow down and have a conversation with your LLM, but really with yourself. Give it a solid 20-40 minutes, at least. I recommend flipping on voice transcription and talking out your answers versus typing them. Have a conversation with yourself.
Open these instructions and copy them into your LLM of choice.
I won’t share yet where this led me, but I found it deeply valuable. I want your journey to be untainted by my own opinions. Do come back to this post, please, and comment if it led you anywhere you found valuable.
Did you know: our word “academy” comes from an actual forest? Plato founded his school in a sacred olive grove near Athens that was known as Akadēmeia. We can learn a lot while in the trees.
Top stories
Courts stop warning and start removing. On June 8, a federal judge in northern Mississippi, Sharion Aycock, removed all four lawyers from a contract case after both sides filed briefs built on AI-fabricated citations. One attorney testified she had not known AI could invent case law. Days earlier the Ninth Circuit suspended two California immigration lawyers for six months over the same conduct, then faulted them for recasting the fabrications as typographical errors. Florida’s mandatory AI-verification rules took effect June 15. Damien Charlotin’s running tracker now logs more than 1,600 such cases. Source
Bezos says the worry is backwards. At VivaTech in Paris on June 17, Jeff Bezos told the audience that AI will create a labor shortage rather than mass joblessness, by letting people pursue more ideas faster. He is now co-CEO of Prometheus, an AI startup chasing an “artificial general engineer” and reportedly raising a $100 billion fund. He compared AI to trading a shovel for a bulldozer. For context, Challenger, Gray & Christmas attributed more than a quarter of April’s US job cuts to AI, with tech-sector layoffs up 66 percent year over year. Source
Colorado’s AI employment law goes live June 30. It is the first broad US state law governing AI in employment decisions. Companies that use AI to help make hiring, firing, or promotion calls must run annual impact assessments, tell candidates and employees when AI is involved, offer human review where feasible, and report any discovered algorithmic discrimination to the state attorney general within 90 days. The law’s future is contested. xAI has sued to block it and the Justice Department has intervened against state AI rules. Until a court says otherwise, the obligations stand. Source
HR wants more AI than its rulebook can hold. SHRM’s new State of AI in HR report finds 92 percent of CHROs expect AI to move deeper into the workforce this year, and 87 percent expect more of it inside HR itself. Yet 54 percent say their own AI policies are already too narrow and tied to today’s tools. Across the 19 states with AI employment laws on the books, 57 percent of HR professionals did not know those laws applied to them. Fewer than half of organizations plan to use AI in HR at all in 2026. Source
Last time around
02:10 on the morning of June 1, 2009. Air France Flight 447, an Airbus A330, is cruising at 35,000 feet over the mid-Atlantic on its way from Rio to Paris. Ice crystals clog the speed sensors, the autopilot clicks off, and three trained pilots are suddenly hand-flying an airplane they had mostly been managing. They never recognize that they have stalled it. The plane falls into the ocean. The aircraft was mechanically sound the whole way down.
Twelve years earlier, an American Airlines training captain named Warren Vanderburgh had already named the problem. Pilots, he warned in a 1997 lecture, had grown so dependent on the magenta course line on their cockpit displays that they were losing the judgment and the raw skill to fly the plane themselves. He called them children of the magenta line. The economics pulled the other way. Automation kept improving and kept cutting workload, so the warning mostly went unheeded until the accident record made it impossible to ignore.
The lesson was not that the automation was bad. It flew those routes more safely than people did, almost all of the time. The lesson was that the rare moment it handed control back was the exact moment the humans were least prepared to take it, because they had stopped practicing. Aviation’s answer was deliberate. Mandate manual flying. Step down a level of automation on purpose. Source
From the frontier
In Nature on June 20, a team led by Sangmin Lee at POSTECH in South Korea, working with the Nobel laureate David Baker at the University of Washington, reported a single engineered protein that folds itself into a virus-like shell. It forms both the five-sided and six-sided arrangements that real viruses use to build their outer casings, something earlier designs needed several separate components to manage. The humans set the problem, defining the geometry a useful shell would need and what would count as a real result. The AI generated the protein sequence that actually self-assembles into it. The shells are a promising container for delivering drugs and vaccines. This is the kind of thing the word superadditive was meant to point at. Source
Potpourri
From someone doing it. Healthcare IT News asked a set of hospital technology leaders how they are actually using AI right now. Dr. Ash Goel gave the line worth keeping. If the most powerful technology of the era only squeezes more margin out of a system already built to put throughput ahead of healing, he said, that is not innovation. It is, in his words, “automating the dysfunction.” It doubles as the cleanest one-sentence test for any rollout. Before you ask whether the tool works, ask what process you are about to make permanent. Source
Enjoyed this prompt - I especially enjoyed the ending where it gave me interesting things to read and go deeper on. Thanks for sharing