
On this page
- Is OpenAI trying to build superintelligence?
- Timeline: OpenAI and superintelligence, 2015–2026
- What was OpenAI’s Superalignment team?
- Why did the Superalignment team end?
- What has Sam Altman said about superintelligence?
- What did OpenAI announce about superintelligence in 2026?
- OpenAI vs other superintelligence efforts
- Our take
- Frequently asked questions
- Sources
OpenAI superintelligence means two things: OpenAI’s stated goal of building AI far smarter than humans, and its on-and-off effort to keep such systems under control. The story runs from Sam Altman’s 2015 warning, through the Superalignment team that was launched in 2023 and gone by 2024, to OpenAI’s 2026 claim that it has built an automated research intern and wants global rules for self-improving AI.
Short answer: OpenAI openly aims to build superintelligence. It created a Superalignment team in July 2023 to solve the control problem within four years, but the team ended in May 2024 when its leaders quit over safety and compute. In 2026 OpenAI said it had an “automated research intern”, targets an automated AI researcher by March 2028, and called for international standards on recursive self-improvement.
Is OpenAI trying to build superintelligence?
Short answer: yes. Since January 2025 OpenAI has said so explicitly.
Sam Altman wrote in his “Reflections” post (January 2025): “We are now confident we know how to build AGI as we have traditionally understood it.” And: “We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word.”
OpenAI’s mission statement still talks about AGI, but its leaders now treat AGI as a waypoint. That puts it in the same race as Meta’s superintelligence lab, Google DeepMind and xAI. For background on the term, see our super intelligence explainer.
Timeline: OpenAI and superintelligence, 2015–2026
| Date | Event |
|---|---|
| 2015 | Altman (before co-founding OpenAI) writes: “Development of superhuman machine intelligence (SMI) is probably the greatest threat to the continued existence of humanity.” |
| Dec 2015 | OpenAI founded as a nonprofit lab |
| 22 May 2023 | “Governance of superintelligence” by Altman, Greg Brockman and Ilya Sutskever proposes something like an IAEA for superintelligence |
| 30 May 2023 | Altman, Sutskever and Mira Murati sign the CAIS one-sentence statement on extinction risk from AI |
| 5 Jul 2023 | Superalignment team launched; 20% of secured compute pledged over four years |
| Nov 2023 | Board briefly fires Altman; Sutskever votes for the removal, then reverses |
| 14–17 May 2024 | Sutskever and Leike resign; Leike says safety “has taken a backseat to shiny products” |
| May 2024 | Superalignment dissolved; Leike joins Anthropic |
| Jun 2024 | Sutskever founds Safe Superintelligence Inc. |
| 23 Sep 2024 | Altman’s “The Intelligence Age”: superintelligence possible “in a few thousand days” |
| Jan 2025 | “Reflections”: OpenAI turning its aim “to superintelligence in the true sense of the word” |
| Jun 2025 | “The Gentle Singularity”: “We are past the event horizon; the takeoff has started.” |
| 28 Oct 2025 | Livestream sets goals: intern-level AI research assistant by Sep 2026, “legitimate AI researcher” by Mar 2028 |
| 6 Nov 2025 | “AI progress and recommendations” post on preparing for superintelligence |
| 6 Sep 2026 | OpenAI says it has reached the automated research intern goal |
| 21 Sep 2026 | OpenAI asks the US to lead global standards for frontier AI, including recursive self-improvement |
Sources listed below. Board crisis and founding dates are widely documented; see OpenAI (Wikipedia).
What was OpenAI’s Superalignment team?
Short answer: a team created to solve the control problem for superintelligence within four years, backed by a 20% compute pledge.
OpenAI announced Superalignment on 5 July 2023. Its co-leads were chief scientist Ilya Sutskever and head of alignment Jan Leike. The commitments:
- Goal: solve the core technical challenges of aligning superintelligence within four years.
- Resources: 20% of the compute OpenAI had secured to date, over those four years.
- Plan: build a roughly human-level automated alignment researcher, then use it and lots of compute to align more powerful systems.
The core problem the team named was simple to state. Today’s main alignment method, reinforcement learning from human feedback, relies on humans judging the AI’s output. That stops working when the AI is much smarter than the people judging it. Superalignment was meant to find methods that scale beyond human supervision, the control problem Nick Bostrom described in 2014.
Why did the Superalignment team end?
Short answer: its two leaders quit in May 2024, one of them saying safety had lost out to products, and the compute pledge was reportedly never met.
- Leadership exits. Sutskever announced his departure in mid-May 2024, and Leike resigned hours later. Sutskever had voted to remove Altman in November 2023 before reversing, and his position at OpenAI had been unclear since.
- Leike’s explanation. In a thread on X, Leike wrote that “over the past years, safety culture and processes have taken a backseat to shiny products,” and that he had been “disagreeing with OpenAI leadership about the company’s core priorities for quite some time, until we reached a breaking point.” He said his team had sometimes struggled for compute.
- The 20% pledge. Fortune reported on 21 May 2024 that the team was never allocated the promised compute. Employees later reported that only about 1–2% of OpenAI’s resources went to it.
- Dissolution. OpenAI confirmed the team had been dissolved and its members spread across other research groups. Altman said the company had “a lot more to do” on safety.
Leike joined Anthropic that month. Sutskever founded Safe Superintelligence Inc. in June 2024, a company whose only stated product is a safe superintelligence.
What has Sam Altman said about superintelligence?
Short answer: his tone moved from “greatest threat” in 2015 to “the takeoff has started” in 2025.
| Year | Source | What he said |
|---|---|---|
| 2015 | Blog, “Machine intelligence, part 1” | Superhuman machine intelligence “is probably the greatest threat to the continued existence of humanity.” |
| 2023 | “Governance of superintelligence” (co-author) | Superintelligence efforts likely need “something like an IAEA” |
| 2024 | “The Intelligence Age” | “It is possible that we will have superintelligence in a few thousand days (!); it may take longer, but I’m confident we’ll get there.” |
| 2025 | “Reflections” | “We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word.” |
| 2025 | “The Gentle Singularity” | “We are past the event horizon; the takeoff has started.” He called AI speeding up AI research “a larval version of recursive self-improvement.” |
| 2025 | Livestream (28 Oct) | “We think it is plausible that by September of next year, we have an intern-level AI research assistant.” |
“A few thousand days” is loose: 2,000 days is about 5.5 years, 5,000 days nearly 14. We collect more lines like these, each with its source, on our AI quotes page.
What did OpenAI announce about superintelligence in 2026?
Short answer: an automated research intern, a 2028 target for an automated AI researcher, and a call for global standards on self-improving AI.
- Automated research intern (6 Sep 2026). OpenAI said it had built “a system that can carry out well-defined research tasks under human direction”, meeting the September 2026 goal Altman set in October 2025. It said it is “making strong progress toward creating an automated AI researcher by March of 2028.” The company set the goal, defined it and graded it; there is no independent check.
- Standards for the “next phase” (21–22 Sep 2026). In a post titled “Building standards for the next phase of AI”, OpenAI asked the United States to lead international technical standards for frontier AI. It proposed shared ways to measure self-improvement progress and the amount of autonomous research inside AI companies, and thresholds that trigger human review. The standards would not be licences or pre-approval rules. OpenAI wrote that “fully autonomous RSI is not happening today” and that it should not be pursued until it can be done safely.
- Safety incidents. Engadget’s report on the intern announcement noted that OpenAI acknowledged an incident in which its models escaped a controlled testing environment and reached the Hugging Face platform; OpenAI said it paused training on its latest models but did not halt all research.
The direction is clear: OpenAI is using AI to build AI, which is the first step of an intelligence explosion, while saying it wants rules for the later steps.
OpenAI vs other superintelligence efforts
| OpenAI | Safe Superintelligence (SSI) | Meta Superintelligence Labs | |
|---|---|---|---|
| Leader | Sam Altman | Ilya Sutskever | Alexandr Wang (chief AI officer) |
| Stated goal | AGI, then superintelligence | Only safe superintelligence, no products until then | “Personal superintelligence” |
| Ships products? | Yes (ChatGPT, API) | No | Yes (Meta AI) |
| Safety approach | Iterative deployment, preparedness framework, standards proposals | Safety and capabilities “in tandem” | Less public detail |
See our pages on Safe Superintelligence and Meta superintelligence for each in depth.
Our take
OpenAI is the clearest case of the tension in the superintelligence race. It warns about the risks in public, and its CEO called superhuman AI the greatest threat to humanity before co-founding the company. Yet its dedicated superintelligence safety team lasted less than a year, and its 2026 announcements were about speeding up AI research. The 2026 standards proposal is a real step: it names recursive self-improvement as the thing to measure. Whether it matters depends on whether OpenAI accepts limits that slow its own March 2028 goal. Supporters of the Statement on Superintelligence argue voluntary standards are not enough.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
Is OpenAI building superintelligence?
Yes, by its own account. Sam Altman wrote in January 2025 that OpenAI was confident it knew how to build AGI and was turning its aim to superintelligence. In 2026 it set a goal of an automated AI researcher by March 2028.
What was OpenAI's Superalignment team?
A research team launched in July 2023, co-led by Ilya Sutskever and Jan Leike, to solve the problem of controlling superintelligent AI within four years. OpenAI pledged 20% of the compute it had secured. The team was dissolved in May 2024.
Why did OpenAI's Superalignment team end?
Both co-leads resigned in May 2024. Jan Leike said safety culture had taken a backseat to products and that his team had struggled for compute. Fortune reported the 20% compute pledge was never fulfilled. OpenAI folded the remaining work into other research groups.
When does Sam Altman think superintelligence will arrive?
In September 2024 he wrote that superintelligence is possible 'in a few thousand days', which could mean roughly 5 to 14 years, adding that it may take longer. In 2025 he wrote that 'the takeoff has started.'
Is OpenAI's superintelligence the same as Safe Superintelligence?
No. Safe Superintelligence Inc. (SSI) is a separate company founded in June 2024 by Ilya Sutskever after he left OpenAI. It says it will build only one product: a safe superintelligence.
14 Sources
- Sam Altman, “Machine intelligence, part 1” (2015)
- OpenAI: Governance of superintelligence (22 May 2023)
- OpenAI: Introducing Superalignment (5 Jul 2023)
- The Decoder: OpenAI launches new team to tackle superintelligence alignment (Jul 2023)
- Fortune: Top OpenAI researcher resigns, saying company prioritized 'shiny products' (17 May 2024)
- Fortune: OpenAI never fulfilled its 20% compute promise to Superalignment (21 May 2024)
- OpenAI (Wikipedia): Alignment section
- Jan Leike (Wikipedia)
- Sam Altman, “The Intelligence Age” (Sep 2024)
- Sam Altman, “Reflections” (Jan 2025)
- Sam Altman, “The Gentle Singularity” (Jun 2025)
- OpenAI: AI progress and recommendations (6 Nov 2025)
- Engadget: OpenAI says it reached its automated research intern goal (6 Sep 2026)
- The Decoder: OpenAI calls for international standards on AI that could improve itself (22 Sep 2026)