Leopold: From OpenAI Safety Advocate to AI Hedge Fund

This can be called a “dual reversal drama” in the AI industry and investment circle: A 23-year-old German youth, who was once a follower of OpenAI’s core figure Ilya Sutskever, was dismissed for insisting on reminding the company to pay attention to AGI safety risks.

However, after leaving OpenAI, he used a 165-page paper as an “investment guide” and achieved a 47% return on investment in six months without any financial background, raising $1.5 billion in funds, far surpassing many seasoned Wall Street investors. What enabled this young man named Leopold Aschenbrenner to accomplish all this? What predictions about the future of AI are hidden in his controversial paper?

Entry into OpenAI

First, we need to talk about Leopold’s experience entering OpenAI, as all his later choices are closely related to the “ideals and conflicts” in this experience.

Back in 2023, a 21-year-old Leopold, with the ideal of “protecting AGI safety,” stepped through the doors of OpenAI’s headquarters in San Francisco, joining the “Superalignment” team led by Ilya Sutskever and Jan Leike.

The core goal of this team was very clear: to solve the problem of “superintelligence alignment” within four years, ensuring that future AI, which would be smarter than humans, could always follow human interests and values without risk of loss of control.

For Leopold, this was not just a job; it was a path to realizing his ideal of being an “AGI guardian.” Therefore, he soon threw himself into his work and, due to his deep understanding of AI safety and alignment with Ilya’s philosophy, became one of the core members around Ilya.

Clash of Ideals and Reality

However, this ideal was soon met with a clash of reality. As OpenAI, under the leadership of Sam Altman, accelerated its product iterations, from GPT-3.5 to GPT-4 and various commercial applications, the company’s focus gradually shifted towards “rapid implementation,” while Leopold’s concerns about “safety measures” were slowly being pushed aside. What worried him the most was discovering significant weak points in the company’s safety systems and even an incident of hacking that was not publicly disclosed. This made him even more anxious about the potential risk of the company’s key algorithms being stolen by competitors.

As someone who had “AGI safety” ingrained in his bones, Leopold could not sit idly by. He quickly wrote a very stern “safety memo,” bluntly stating that OpenAI’s safety measures were “extremely inadequate,” and even described some aspects as “virtually non-existent.” More crucially, instead of following the usual process to submit it to his superiors, he bypassed the hierarchy and directly handed the memo to the OpenAI board of directors. This action directly sparked tensions between the company’s upper management and the board, making Leopold a “special element.”

To make matters worse, during the later ordeal involving Sam Altman’s temporary dismissal and reinstatement, most employees signed a petition supporting Altman’s reinstatement. However, Leopold did not sign, which the company interpreted as a sign of “insufficient loyalty.” HR soon approached him with a warning. If this was merely a matter of “taking sides in the workplace,” perhaps the situation could have been salvaged. But the next incident completely led Leopold and OpenAI toward a rupture.

Parting ways with OpenAI

In early 2024, Leopold organized a “brainstorming document” on AGI safety measures, which consisted of analyses and thoughts based on publicly available information and contained no internal secrets of OpenAI. He even invited three external scholars to provide feedback. However, the company deemed this document a “leak of internal information” and initiated an investigation based on it. During the investigation, his previous “bypassing of hierarchy to submit the memo to the board” came to light. Ultimately, in April 2024, OpenAI formally dismissed Leopold for “leaking information and not cooperating with the investigation.”

Interestingly, shortly after Leopold was fired, his former direct supervisors, Ilya Sutskever and Jan Leike, also left OpenAI. While there is no direct evidence linking their departures to Leopold, those familiar with the internal workings of OpenAI know that it is likely related to the “conflict between safety philosophy and commercialization pace.” Leopold’s experience may have simply been a “microcosm” of this conflict.

A New Venture

After being dismissed, Leopold did not fall silent; instead, he compiled all his thoughts and observations into a 165-page paper titled “SITUATIONAL AWARENESS: The Decade Ahead.” This paper was published in June 2024, just two months after his dismissal.

In the introduction, Leopold specifically noted, “This paper is dedicated to Ilya Sutskever,” which made it clear to the public that he identified himself as a “follower of Ilya.” The core argument of this paper not only established his status as an “AI trend forecaster,” but also became his “core weapon” for entering the investment circle later.

So, what exactly did this 165-page paper contain?

The core prediction is that Leopold forecasts “around 2027, humanity may welcome AGI, and by 2030, we will move toward superintelligence.” He also strongly urged that the world needs to unite to establish an “AI Manhattan Project” to address the safety risks that AGI may bring. This prediction is backed by a large amount of data and logical deduction.

Key Dimensions of Analysis

We can understand his arguments from three dimensions: “The speed of evolution of AI model capabilities,” “the explosive growth of computational power investment,” and “breakthroughs in algorithm efficiency and unlocking gains.”

First, looking at the speed of evolution of AI model capabilities, Leopold made a very intuitive analogy in the paper. If we compare the capabilities of AI models to human intelligence levels, he pointed out that the GPT-2 model released in 2019 had capabilities roughly equivalent to that of “preschool children.” At that time, GPT-2 could only generate a somewhat incoherent paragraph, and in text summarization tasks, the results it generated were only marginally better than simply “randomly selecting three sentences from the original text.” Even basic tasks like counting from 1 to 5 could present logical breaks.

By the time GPT-4 was released in 2023, its model capabilities had reached the level of a “smart high school student.” In the sciences, GPT-4 could think deeply and reason through complex mathematical problems, capable of handling advanced programming tasks from writing complex code to iterative debugging. In the humanities, it could maintain logical and content consistency across tens of thousands of words of text and even engage in deep philosophical and historical discussions with humans.

More crucially, in almost all standardized tests, including AP (Advanced Placement) and the SAT (Scholastic Assessment Test), GPT-4’s scores surpassed the vast majority of high school students.

In just four years from 2019 to 2023, the GPT series of models had jumped from “preschool” to the “high school” level, whereas it would take humans at least 12 years to complete this phase of intellectual development. Leopold calculates that the speed of evolution of AI model capabilities is more than three times that of human natural intelligence development. If this speed continues, by 2027, just four years after the release of GPT-4, the model capabilities reaching “AGI level” seems to have logical support.

Economic Considerations

Next, looking at the explosive growth of computational power investment, Leopold compared the differences in computational power growth between the “Moore’s Law era” and the “AI era.” In the heyday of Moore’s Law, computational power only grew by 1 to 1.5 orders of magnitude every ten years; however, during the AI era, the annual growth rate of AI hardware in terms of computational power has reached 0.6 orders of magnitude, meaning about a fourfold increase per year, which is more than five times faster than before.

Using the GPT series models as an example makes this even more intuitive: from GPT-2 to GPT-3, the training equipment transitioned from “small experimental devices” to “large data centers,” with computational power increasing by 2 orders of magnitude (i.e., 100 times) in a year. The training power of GPT-4 achieved significant improvements on this basis. Moreover, from OpenAI’s continuous stockpiling of GPU chips, it appears that this speed of computational growth is not a short-term burst but will gradually evolve into a long-term trend.

Leopold made an even bolder prediction in his paper: by the end of this decade, before 2030, trillions of dollars will be invested globally in GPU, data center, and power construction, with the U.S. needing to raise power production by several tens of percentage points to support the development demands of AI.

He also derived predictions from the trend of AI companies’ revenue. With the rapid growth of revenue from AI products, tech giants like Google and Microsoft are projected to reach around $100 billion in AI-related annual revenue around 2026. This revenue growth will further stimulate capital investment, projecting that by 2027, global annual AI investment could exceed $1 trillion. If we extend the timeline further to 2028, the cost of training a single AI model’s training cluster could reach the level of $100 billion, surpassing the International Space Station’s estimated $150 billion cost. By the end of this century, the cost of an ultra-large training cluster could even reach $1 trillion, at which point the power needed for AI is expected to increase from less than 5% of total U.S. power generation to 20%. While these predictions may sound exaggerated, they are not entirely without basis given the rapid growth trend in AI computational investments in recent years.

Breakthroughs in Algorithms

The third dimension, which Leopold believes is the easiest to underestimate, is breakthroughs in algorithm efficiency and the resulting “unhobbling gains.” First, regarding algorithm efficiency, Leopold used two specific cases in the paper to demonstrate the speed of algorithm progress. The first case is the MATH benchmark test, a standard test measuring AI’s mathematical reasoning abilities. From the earlier Minerva model to the later Gemini 1.5 Flash model, the reasoning efficiency improved nearly three orders of magnitude, which is a 1000-fold increase, under the condition of reaching 50% accuracy. Although reasoning efficiency does not equate to training efficiency, this trend is sufficient to show that major improvements at the algorithmic level are feasible and happening rapidly.

The second case is the ImageNet image recognition task. A review of published algorithm research from 2012 to 2021 found that to train models of the same performance, computational costs decreased at an almost consistent pace of about 0.5 orders of magnitude per year, meaning costs decreased by approximately two-thirds each year. Moreover, this downward trend applies to all mainstream model architectures. Although teams developing large models rarely publish data on algorithm efficiency, estimates from EpochAI suggest that from 2012 to 2023, large models achieved about 0.5 orders of magnitude increase in algorithm efficiency annually, implying that over eight years, algorithm efficiency total improvements were about 10,000 times. This data is enough to break the stereotype that “AI progress relies entirely on computational power.”

The concept of “unhobbling gains” proposed by Leopold is even more interesting. Simply put, AI models’ inherent capabilities are sometimes ‘bound.’ Through simple algorithmic improvements or technical adjustments, these potential capabilities can be unlocked, yielding returns far greater than typical algorithm optimization. Though it is fundamentally a form of algorithm improvement, unlike traditional optimization, “unhobbling gains” can leap outside existing training paradigms to deliver a step-level increase in model capabilities and practical value. Leopold presented four typical examples in the paper:

First is RLHF (Reinforcement Learning from Human Feedback). By allowing humans to score and provide feedback on outcomes generated by AI, and then using this feedback to train the model, a small model can achieve performance “hundreds of times greater than its original model.” For example, a 10-billion-parameter model optimized through RLHF may perform comparably to a 1-trillion-parameter unoptimized model on certain tasks.

Second is CoT (Chain of Thought). This involves the AI verbally elaborating on its thought process step-by-step when solving problems rather than giving direct answers. This simple prompting method can lead to over ten times “effective computational power improvement” for reasoning tasks, allowing the model to tackle more complex problems with the same amount of computational power.

Third is increasing context length. By extending the amount of text AI can process from early 2048 tokens to over 1 million tokens, new applications were unlocked, enabling AI to process an entire book at once, analyze lengthy codebases, summarize hours of meeting transcripts, and more. This capability expansion is not merely “quantitative” but represents a “qualitative” change.

Fourth is post-training. After a model’s official release, continual input of new data and optimization of training strategies allows for secondary training. This has been a major factor behind continually improving the performance of GPT-4 after its release. For instance, the mathematical reasoning ability of GPT-4 improved nearly 30% compared to its initial release six months later, thanks to post-training. According to studies by EpochAI, such “unhobbling gains” techniques generally bring about 5 to 30 times “equivalent computational power gains” for AI models. Another research institution, METR, found that in agent tasks, the basic performance of the GPT-4 model was only 5%, but after post-training, it could improve to 20%. Combining tools and framework optimization, performance could leap to nearly 40%, an improvement level unattainable by simply increasing computational power.

Future Projections

When the factors of growth in computational power, improved algorithm efficiency, and unhobbling gains converge, Leopold draws his most core conclusion: by 2027, AI will be able to automate all cognitive tasks, or all tasks that can be conducted remotely.

In his paper, he made a more specific deduction that from GPT-2 to GPT-4, the model roughly experienced an effective computational expansion of 4.5 to 6 orders of magnitude, and transitioning from the base model to a functional chatbot corresponds to the acquisition of about 2 orders of magnitude of “unhobbling gains.”

At this rate, in the next four years, meaning by 2027, improvements in computational efficiency will significantly accelerate the model iteration speed. If GPT-4 took three months to train, then by 2027, leading AI laboratories will be able to train a model of GPT-4 caliber in “one minute.” Secondly, “unhobbling gains” will transform AI from a purely “chatting tool” into an “agent capable of independently completing complex tasks.” It will be able to use computers, manage long-term memory, engage in several hours or even days of continuous thought on an issue, and adapt quickly to new companies’ workflows and business logic, just like human employees.

However, Leopold also emphasized in the paper that this prediction comes with a “large margin of error.” If “unhobbling gains” gradually stagnate or algorithmic progress does not solve the “data exhaustion” problem, the arrival of AGI would be delayed. Conversely, if “unhobbling gains” can release greater model potential or if new breakthroughs emerge, AGI may arrive even earlier than 2027.

Transition to Investment

This forward-looking paper not only gained Leopold more attention in the field of AI safety but also unexpectedly became his stepping stone to cross into the investment circle. In the second half of 2024, Leopold established his own hedge fund, with the core investment logic being what he proposed in the paper—“situational awareness.” In simple terms, this means “accurately predicting the AI industry’s technology trends and market demands to strategically position related enterprises.”

Though he had no prior experience in the financial industry, his position as a KOL in AI and the thorough data support from his paper allowed him to successfully persuade a large group of top investors. According to public information, his fund supporters include Patrick Collison and John Collison, founders of the payment company Stripe, who are well-known entrepreneurs in the tech circle with a profound understanding of AI trends.

There are also recent hires at Meta (previously Facebook) for AI business, such as Daniel Gross and Nat Friedman, the former being a renowned AI investor and the latter a former GitHub CEO, also deeply entrenched in the tech field. The endorsement from these supporters not only brought funding to the fund but also industry resources and credibility. The fund’s performance has not disappointed either.

Impressive Returns

In the first half of 2025, after deducting various expenses, Leopold’s fund achieved a 47% return. What does this achievement represent? During the same period, the average return of the world’s top hedge funds was about 12%, and even those focused on technology averaged around 25%. Leopold’s 47% is nearly double the industry average. More importantly, his fund’s size grew from several tens of millions of dollars to $1.5 billion within just six months, making it “one of the fastest-growing hedge funds in history.”

Leopold’s investment capabilities were vividly demonstrated in his judgments regarding NVIDIA. As early as in an interview in 2024, he predicted that “NVIDIA’s data center revenue will grow from several billion per quarter to $25 billion per quarter.”

At that time, many financial analysts deemed this prediction overly aggressive, but reality surpassed his expectations. NVIDIA’s Q2 revenue from data centers in the fiscal year 2025 reached $26.2 billion, exceeding his prediction of $25 billion, while in the second quarter of fiscal year 2026, this figure approached $41.1 billion, almost 1.6 times his initial prediction. Such precise judgments of industry trends have encouraged more investors to believe that this “AI expert with no financial education” truly possesses a unique capacity for “situational awareness.”

Industry Dynamics

Leopold’s success is, in fact, a reflection of the current fervor in the AI industry. According to CNBC citing Pitchbook data, in the first half of 2025, global AI startups raised a total of $104.3 billion, while the total exit amount based on venture capital exceeded $36 billion. The largest investments included OpenAI’s $40 billion raised in March 2025, AI data labeling company Scale AI’s $14.3 billion investment from Meta, and Anthropic’s $3.5 billion raised. These figures show that the AI industry has transitioned from “concept hype” to a “capital-intensive development” phase. On one hand, major tech companies and investment institutions are continuously pouring money to drive technological breakthroughs; on the other hand, “industry insiders” like Leopold are siphoning off the wealth that belongs to them from secondary markets. This cycle of “technology driving capital, capital reinvigorating technology” has created a “crazy growth” situation in the entire AI industry.

Conclusion

Leopold’s story is the most vivid annotation of this madness: an AI researcher dismissed for adhering to safety principles became a new star in the investment circle through a paper. This kind of cross-border success is almost unimaginable in other industries, but it has truly happened in the AI era.

However, we should also view this story objectively. Leopold’s success has its uniqueness; he is situated at the core of AI technology, capable of grasping industry dynamics in real-time while also having the ability to translate technological trends into investment logic. This “dual advantage” is not something everyone possesses. Moreover, his fund’s remarkable performance lasted only six months. Whether it can maintain such high growth in the future remains to be seen.

It is undeniable that his experience serves as a reminder: in an era of rapid AI development, “understanding technological trends” is becoming a core competitive advantage. Whether in technological research and development, product implementation, or investment decision-making, those who can grasp the direction of AI’s development earlier and more accurately are more likely to seize opportunities in this era.

References

1. Introduction - SITUATIONAL AWARENESS: The Decade Ahead SITUATIONAL AWARENESS - The Decade Ahead. 2024-06
2. Ilya Sutskever and Jan Leike RESIGN from OpenAI - My in-depth analysis - end of an era! YouTube. 2024-05-17
3. Scalable Oversight in AI: Beyond Human Supervision Medium. 2024-05-20
4. Meet OpenAI’s All-Powerful Board That Pushed Out CEO Sam Altman Observer. 2023-11-21
5. Tolga Bilge X (Twitter).
6. I. From GPT-4 to AGI: Counting the OOMs - SITUATIONAL AWARENESS SITUATIONAL AWARENESS - The Decade Ahead. 2024-05-29.
7. Generative AI Market Growth is Booming with 27.02% Precedence Research. 2023-07-11
8. Improving your LLMs with RLHF on Amazon SageMaker Amazon. 2023-09-22
9. Prompt Engineering 101 — Revolutionizing Chain-of-Thought Reasoning with Auto-CoT Medium. 2024-05-08
10. LLM Context Window and Query Process ResearchGate.
11. New LLM Pre-training and Post-training Paradigms Ahead of AI. 2024-08-17
12. Measuring the impact of post-training enhancements METR’s Autonomy Evaluation Resources.
13. Leopold Aschenbrenner — 2027 AGI, China/US super-intelligence race, & the return of history YouTube. 2024-06-05
14. Stripe’s new $95B valuation makes it the most valuable Silicon Valley startup Payments Dive. 2021-03-15
15. AI Investors Are Wooing Startups With Massive Computing Clusters Forbes. 2024-02-14
16. AI Nationalization is Inevitable – Leopold Aschenbrenner YouTube. 2024-06-06