On Tuesday, Sam Altman sent a memo to OpenAI staff. In it, he said three things that, taken together, tell you exactly where this is all heading.

One: OpenAI had finished pretraining its next major model, codenamed Spud.

Two: Sora, the AI video app, was being shut down immediately to free up compute for Spud.

Three: Altman was stepping back from direct oversight of the safety and security teams so he could focus on raising capital and building infrastructure at "unprecedented scale."

He added, at the end: "Things are moving faster than many of us expected."

The Sora Shutdown Is Not What It Looks Like

Sora launched as a standalone consumer app in September 2025. By January 2026, downloads had dropped 45 percent month-over-month. The video generation space had become impossible: Google has Veo, subsidized by ad revenue. Chinese competitors like Kling and Hailuo are faster and cheaper. xAI is dominating the leaderboards with it’s image and video models.

But Bill Peebles, who ran the Sora team, framed the pivot in terms that reframe the whole story. He said the research team would now focus on "systems that deeply understand the world by learning to simulate arbitrary environments at high fidelity." He called this "world models." And then he said: "The prize is automating the physical economy."

The video app was the surface-level product. The actual research goal was always to build a model that understands how the physical world works well enough to simulate it. That matters for robotics. That matters for manufacturing. That matters for anything that involves a machine operating in the real world.

OpenAI did not exit video. They exited the consumer app that funded the early research. The underlying capability is being pointed at something much bigger.

The Disney deal died in the process. Disney had committed to a $1 billion investment tied to a three-year licensing deal for their characters. That deal had never closed. It is now not proceeding. Disney's statement was diplomatic: "We respect OpenAI's decision to exit the video generation business." No money changed hands.

What Spud Is

OpenAI has not said much publicly. What Altman told staff: pretraining is done, the model is "very strong," it will ship "in a few weeks," and it "can really accelerate the economy."
That last phrase is doing a lot of work.

Altman does not use language like that for incremental upgrades. "Accelerate the economy" implies something categorically different from better autocomplete. It implies that the model can take on tasks that currently require people, and do them faster, cheaper, and at a scale that changes how companies operate.

The model will power OpenAI's upcoming desktop superapp, which is being built by merging ChatGPT, the Codex coding agent, and the Atlas browser into a single product. The org chart has already been updated: the product division was renamed "AGI Deployment." That is the first time OpenAI has made AGI a product category name.

Altman removed himself from the safety and security reporting chain, handing those teams to Mark Chen (research) and Greg Brockman (scaling). The official reason is bandwidth. He needs to focus on hardware, data centers, and capital. The timing, right before releasing what he is describing as a potentially transformative model, is worth noting.

Terence Tao Used ChatGPT to Prove a Theorem

While OpenAI was announcing all of this, the greatest living mathematician quietly published something that belongs in the same story.

Terence Tao holds the Fields Medal. He is widely considered the most technically gifted pure mathematician alive. In a recent paper, he wrote:

"After experimenting with AlphaEvolve, I was led to a way to prove the toy model's integral bound by splitting it into two lower bounds. ChatGPT proved the first, and I proved the second."

Read that again. He split a proof into two parts. He took one. AI took the other.
This is not a demo. This is published mathematical research. Tao is not a casual experimenter. He understood what he was doing, verified the AI's work, and attributed it in the paper. The proof stands.

AlphaEvolve is a system from Google DeepMind that uses Gemini to iteratively propose and test mathematical constructions. Tao used it to find the approach, then handed one piece of the proof to ChatGPT. The model in question is likely GPT-5.4 Pro, given the level of mathematical reasoning required, though Tao only said "ChatGPT" without specifying the version.

This is not the first time. Tao maintains a GitHub wiki tracking AI contributions to Erdos problems. As of March 2026, AI has contributed to roughly 100 problems on that list. In January, GPT-5.2 Pro solved three Erdos problems in seven days. Tao personally verified each proof.

Tao predicted in early 2025 that AI would become a "trustworthy co-author" by 2026. He made that prediction optimistically. It landed ahead of schedule.

Why These Two Stories Belong Together

OpenAI is killing products, restructuring toward "AGI Deployment," and preparing a model they say will accelerate the economy. The greatest mathematician in the world is splitting proofs with ChatGPT and publishing the results.

These are the same story told from two different directions.

From the top: a company betting its entire trajectory on one model, freeing compute from everything else to build it, pointing the org chart at AGI as a product category.

From the ground: a mathematician who could solve any problem in his field choosing to bring AI into his proofs not because it is faster, but because it finds things he does not see.
The question people are still treating as hypothetical, "when will AI do real scientific work," has been answered. It is doing it now. The question Altman is betting Spud on is: what happens when that capability arrives for everyone, in every field, at once?

He said things are moving faster than many of them expected.