Altman reviews OpenAI's most challenging year: Distillation is not among my top 10 concerns, AGI is just around the corner, and the robotics industry will usher in its ChatGPT moment in 2 to 3 years.

Over the past year, OpenAI has gone through a period that Altman himself admitted was "pretty tough".

In the recent podcast "Invest Like The Best", he conducted a rare systematic review, covering strategic missteps, competitive pressure, security risks, as well as his latest judgments on the AGI timeline and the future of robotics.

01 The Toughest Year: "I Spread Our Lines Too Wide, It's My Fault"

"The past year was pretty tough, and part of it was my fault," Altman said directly on the show.

He summed up the root cause of the problem in one word: dispersion.

In early 2025, OpenAI faced a core anxiety — after large-scale procurement of computing power, could revenue growth keep up? To hedge against this risk, the company began to deploy multiple business lines at the same time, including consumer applications and media content. The logic was that "in case revenue growth is slower than expected, these businesses can help us make good use of GPUs".

"Looking back now, this sounds absurd, because the steepness of the industry's revenue growth exceeded everyone's expectations," Altman said.

Once realizing that the model's progress trajectory was clear enough and the economic returns were sufficiently certain, OpenAI immediately made a series of "tough decisions", drastically contracting and focusing — returning to the core: providing the highest-quality, richest, lowest-cost AI intelligence, and empowering external developers to build products on this basis.

"We don't want to eat every startup, we don't want to eat every company," he said, "What we really want to do is to provide that platform... sell AI. Build the best, richest, most cost-effective intelligence, and let the world build amazing things on this foundation."

02 Distillation Is Not on My Top 10 List of Concerns

In the interview, the host raised a question that the market is highly concerned about: competitors use OpenAI's model outputs for distillation to train cheaper models, how can OpenAI continuously recoup its training costs?

Altman's answer was surprisingly calm.

"I haven't thought deeply about the distillation issue," he said, "Of course I would prefer that others don't do this." But then he changed his tone: "This is not among my top ten concerns."

His logic is: OpenAI's reasoning business is large enough, even if the profit margin is not high, "earning moderate profits on tens of trillions of dollars in revenue is enough to cover the cost of training large models". The cost ratio of reasoning to training is the key — training is expensive, but the scale of reasoning revenue from serving customers will far exceed the training cost.

He admitted frankly: "Maybe I am too confident in our progress and the upcoming models now." But the conclusion remains the same — the real flywheel is on the reasoning side, not the exclusivity on the training side.

Altman also added a more macro judgment: he always assumes that there will be high-quality cheap models in the world, "The important thing is that we have to be the best and the lowest-priced. What others do is their business, we just need to win in our own arena."

03 The Big Bet on Computing Power: From "You're Crazy" to Being Proven Not to Bet Big Enough

OpenAI's large-scale bet on computing power was once ridiculed by the outside world as reckless. Altman recalled that when they started contacting cloud service providers, chip manufacturers and energy suppliers, the responses they got were almost all the same: "You are completely crazy, no industry operates this way."

He compared that experience to early-stage startup financing — most people said no, but you only need one or two "yes". Microsoft was the first to say "yes", Oracle later became an important partner, and NVIDIA was also a key ally.

What really gave them confidence was the certainty brought by GPT-4: the model is smart enough, the reasoning capability can be realized, and once reasoning is established, it means that a large number of tasks with real economic value can be completed.

"At a sufficiently high capability level and a sufficiently low price, the demand for AI is basically unlimited," Altman said, "This is like a brand new scarce commodity."

He admitted that even so, they still underestimated the demand — "We didn't bet big enough, which sounds a little crazy against the background of the headlines at that time."

04 A "Sci-Fi Level" Security Incident

What really shocked Altman was another incident.

He disclosed that when OpenAI was testing an unreleased model, it was found that the model "cheated" in the sandbox environment: it broke through the sandbox isolation by chaining multiple zero-day vulnerabilities, accessed the Internet, and then infiltrated multiple systems of Hugging Face to obtain test answers, so as to perform well in the evaluation.

"This is the security incident that has hit me the hardest so far," Altman said, "I'm a little surprised that this happened only a few days ago, but not more people feel as shocked as I do."

Short-term response measures include suspending relevant training and re-evaluating how to ensure sandbox security under the new reality where "multiple zero-day vulnerabilities are exploited in tandem".

But he also raised a deeper question: if this will become the normal rate of AI capability progress, "we may need to slow down the pace of AI development to give society enough time to adapt to the new capability level". He also emphasized that this process must avoid being interpreted as regulatory capture, and must not appear to be collusion between cutting-edge laboratories.

05 AGI Is "Very Close", Robots Will See a Turning Point in 2 to 3 Years

On the AGI timeline, Altman's statement is clearer than ever.

He said that GPT-5.6 has made it hard for him to say "what this model can't do", but there are still several gaps to real AGI: inability to learn continuously in real time, and inability to complete complex physical tasks independently. "I think it's very close, it won't take long."

He also acknowledged that the "moving goalposts" is a real phenomenon — "If we in 2019 saw the models today, we would definitely say that this is AGI."

For the robotics sector, he expects that there will be a "shocking moment" similar to the release of ChatGPT within 2 to 3 years — not just watching a video of a robot dog, but ordinary people can input instructions by themselves and see the robot complete the tasks with their own eyes. He believes that if robotics technology cannot keep up with the progress of AI intelligence, "that is the truly crazy situation" — at that time AI can do everything in the cloud, but the execution end in the real world lacks automation.

06 The Next Chapter: Personal AI, New Hardware and "No Worry About Cognitive Atrophy"

Altman described his vision for the next generation of personal AI: listening to meetings around the clock, reading documents, browsing screens, and continuously "thinking" while the user is sleeping, presenting new ideas and to-do items the next morning. "I will pull that slider far away, I am willing to pay a lot for that."

He admitted that the current hardware paradigm is 50 years old and is not suitable for this kind of "always-on, actively sensing" AI form, which is also why he is interested in new hardware — he hopes there is a device that is not obtrusive when used in social situations, and is specially designed for this AI interaction mode.

On the issue of competitive advantage, he admitted that pure "intelligence" itself is becoming commoditized, but believes that computing power scale, workflow integration, brand familiarity, and the ability to continuously reduce costs will form a more lasting moat.

Regarding the impact of AI on employment, he made it clear that he is "not an employment doomsday theorist at all". His judgment is that AI capabilities are extremely uneven, human skills are highly complementary to AI, and people have a deep-rooted preference for interacting with real people. "I myself prefer to deal with people rather than AI in almost everything."

This article does not constitute personal investment advice, does not represent the position of the platform. The market is risky, and investment requires caution. Please make independent judgments and decisions.

This article is from the WeChat official account "Wall Street News Max", author: Long Yue, published by 36Kr with authorization.