Altman says the safety climate is delaying OpenAI's IPO until next year

Somewhere there is a plan with a line on it saying a feature gets good when the next model ships. That line is a bet on another company's release schedule, and over the weekend the people who set those schedules argued in public for making them slower. Amodei's essay, published Saturday, asked the industry to slow "the pace at which we improve the capabilities of AI models" and warned about bioterrorism and "serious economic disruption" ^[1]. Altman and Elon Musk voiced support, according to Gizmodo ^[2]. Altman also said in a Saturday interview that OpenAI would pause more often during new model development ^[4].

Teams tell themselves they are waiting for a capability. What they usually do is ship a workaround, and the workaround turns out to be the product: a narrower feature with a person checking the output. That was the cheaper path when releases came fast, and it is the one that survives a vendor deciding to wait.

Altman said the current safety environment is factoring into OpenAI's decision to delay its long-awaited IPO until next year ^[3]. In the same interview he said: "If you're like 'Oh, OpenAI is not going to pause because they're worried about it costing their business money,' you like very deeply misunderstand us" ^[5]. Microsoft published a code of conduct on Monday promising to put limits on AI models ^[6].

The political pressure runs the other way. Trump attacked Amodei on Truth Social on Monday, calling public opposition to unrestrained AI development "a SICK conspiracy" that only benefits China ^[8], and wrote "WHOEVER WINS AI, WINS! We are leading China, and all others, and will continue to do so." ^[9] House Speaker Mike Johnson said on Sunday that Congress should refrain from AI regulation and that developers should bear responsibility for safety ^[10]. Earlier this month the administration closed the G20 meetings in North Carolina with the Carolina Principles, a document asking the biggest global economies to promise a more hands-off approach; Altman and Anthropic co-founder Tom Brown attended talks hosted by Commerce Secretary Howard Lutnick ^[11][12].

Amodei's essay limits how far it asks the slowdown to go. He argues development should not slow so much that the United States falls behind China, whose leadership he says would pose "grave danger" for the entire world ^[13]. The essay, the endorsements, Johnson's remarks, Microsoft's code and Trump's post all landed between Saturday and Monday ^[14]. In Gizmodo's account, the pause commitments carry no dates, no cadence, no deprecation terms ^[15].

Two questions sort a roadmap into four boxes: whether an item needs a capability today's model does not have, and whether it needs today's model to keep behaving exactly as it does. Items that need neither are yours to schedule. When only the future capability is missing, the date belongs to the vendor, and the feature needs a version that ships without it. Where the exposure is stability instead, the thing to ask for is a written deprecation window, plus an eval suite you run on every version bump. I would not put a customer-facing date on an item that needs both.

The step is smaller than the framework: for each line on the plan, name the model version that makes it work today. When nobody can name one, the line is a request to a vendor, and it belongs in a different column.