09: IP-Oh
Your AI vendors, now central to your org, are about to change in ways you don't control. How do you plan?
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
Two big things happened this week. First, Anthropic filed to go public on June 1, and OpenAI is weeks behind it. Ahead of both, the Wall Street Journal reported the two are cutting token prices to win customers before the IPO. So AI is cheaper right now, for now.
Second, on Friday night, the US government ordered Anthropic to block foreign access to its two best models. To comply, Anthropic shut off Fable 5 and Mythos 5 for everyone. A model thousands of companies were working with was taken away suddenly. Not because Anthropic decided to. Because the government did.
Put those events together. The company that runs your AI is about to answer to public shareholders, and shareholders want growth. The low prices right now are temporary. Once these companies have to show profit, expect prices to rise, contracts to get longer, and models you depend on to be retired on the vendor’s timeline, not yours. Or, as Friday showed, a model can simply stop working overnight for reasons outside your control. A few years ago these were research labs, and their goals and yours mostly matched. As public companies, they won’t.
So how do you make plans on such shaky ground? You can’t know, so don’t try. A forecast commits you to one future. Prepare for several instead.
What to say to your CEO this week: “We depend on one or two AI vendors, and this week showed they’re about to change in ways we don’t control. They’re going public. They’re getting regulated. One was shut off Friday night. I don’t want to bet the company on guessing how this plays out. Let’s name three things: the big bet we’re making on purpose, the protection we have against the worst case, and the cheap moves we should make no matter what.”
This week’s move: prepare for more than one future
Scenario planning has one rule people skip. Don’t just list what might happen. Sort your responses into three buckets.
No-regret moves. Worth doing in every version of the future, so start now. Find out your real cost to switch vendors for each workflow, the actual number. Keep your prompts, your tests, and your training data in a form you can move, so you’re not stuck with one provider. And measure the work you actually finished, not the hours the tool says it saved.
Protection against the worst case. Cheap insurance. Run at least one real workflow on a second provider or an open model, in production, not as a plan on a slide. Then switching is something you’ve already proven you can do. Friday is the reason: a model you depend on can stop working with no notice, and you want a fallback ready before that happens, not after.
Big bets. Where you spend real money to improve your own position. Commit to one vendor now, while you still have leverage, and lock in the terms. Or build your own capability on open models so a price increase can’t trap you. Or add so much of your own value on top, your data, your workflow, your judgment, that you can swap the model underneath without much pain. You can’t do all three. Choose one. That choice is your strategy.
Say this part out loud in the room. Most companies made a few no-regret moves, called it a strategy, and made one large bet without consciously meaning to: full dependence on a vendor whose goals no longer match theirs.
Top stories
The labs are going public, and cutting prices to get there. Anthropic filed for an IPO on June 1. OpenAI is reportedly aiming to list as soon as September. Ahead of both, the Wall Street Journal reported on June 10 that OpenAI is weighing steep token price cuts to compete with Anthropic’s Fable 5 and Claude Code. For buyers, AI gets cheaper now. The open question is who covers the gap between what companies will pay and what these models cost to run, once public shareholders are watching. [CryptoBriefing]
The government pulled a frontier model offline overnight. On Friday night, the US Commerce Department barred foreign nationals from Anthropic’s two best models. To comply, Anthropic shut off Fable 5 and Mythos 5 for every customer. Its other models stayed up. It is reportedly the first time a deployed frontier model has been taken offline by government order rather than the company’s own decision. Anthropic says it is complying but disagrees, and wants access restored. If you built on those models, you learned something Friday: you don’t fully control whether they keep working. [Bloomberg]
The people who sold the future spent the week revising it. In two weeks, all three frontier-lab heads softened their predictions about jobs. Sam Altman said he was “delighted to be wrong” that entry-level white-collar jobs would be gone by now. Microsoft’s Mustafa Suleyman, on the Decoder podcast, said his February “twelve to eighteen months” line meant tasks, not jobs. Dario Amodei walked back his half-of-entry-level number. Worth remembering the next time a vendor gives you a confident timeline to plan around. [Business Chief]
CIOs are being told to expect higher, more locked-in pricing. A June CIO analysis asked analysts and IT leaders what the IPOs mean for budgets. The expectation: a move from flexible per-use pricing toward metered, locked-in contracts, with extra charges for features included today. Richard Amos, CIO at the integrator Blue Mantis, put it simply. Before an IPO, a company chases growth. Once public, shareholders want steady quarterly results, and pricing follows. He expects companies to start managing AI spending the way they already manage cloud spending. [CIO]
A third of the time AI saves goes back into fixing AI. A Workday study of 3,200 companies found that 85% of workers say AI saves them one to seven hours a week. But about 37% of that time goes back into correcting what the tool got wrong. Only 14% come out ahead. That is why you should measure the work people actually finished, not the time the tool claims to save. The time saved and the time kept are different numbers. [Workday]
Last time around
September 1972. In a London conference room, a French oil executive named Pierre Wack gave Shell’s directors a long presentation about the future of oil. He laid out several possible futures, and in one of them, the price of crude rose sharply. Wack was clear that he was not predicting it. He was trying to get Shell’s leaders ready to act quickly if it happened.
In October 1973, it happened. OPEC’s embargo sent crude from about $2.50 a barrel to $11. Most oil companies were caught unprepared, and some failed. Shell’s leaders had already thought through this exact situation, so they moved quickly, cutting costs and adjusting while competitors were still reacting.
Be careful with the legend. Shell is often said to have risen from seventh to second among the major oil companies because of this, and researchers disagree about whether the planning actually caused it. So take the smaller, safer claim. The point of the exercise was never a more accurate forecast. It was a faster response, because Shell’s leaders had already considered more than one future before this one arrived. That is the whole idea, and it costs you one afternoon of thinking about things you’d rather not. [strategy+business]
From the frontier
In June, a Cambridge team reported that an AI-designed coronavirus vaccine passed its first human trial. Thirty-nine volunteers. The vaccine was safe, well tolerated, and produced immune responses not just to COVID but to SARS and to bat viruses that haven’t reached humans yet. Note who did what. The humans set the hard question, what a vaccine against a whole family of viruses would need to target, and ran the trial. The AI designed the antigen, searching through more viral features than a person could check by hand. It’s a safety trial, not a cure, and it’s years away from real use. But it’s the first vaccine designed entirely by AI to reach a human trial, and that’s the kind of result this newsletter’s name is about. [ScienceDaily]
Potpourri
From someone doing it. Robin Sloan, the novelist who writes using these tools, asked a question worth borrowing. AI, he writes, makes it easier to fool yourself, so you have to be honest about whether you’re thinking with it or just letting it do the hard part for you. The same question works for a whole company. You can hand off the work. You can’t hand off the thinking, because the thinking is the part that’s still yours when the tool you were using changes. [Robin Sloan]