Anthropic Scientist Warns: If AGI Arrives Within Three Years, Ph.D. and Tenure Plans May Already Be Obsolete
A talk given in June by Jascha Sohl-Dickstein, a scientist at Anthropic, has recently appeared again on the social platform X and started a discussion. The talk took place at a conference on physics and AI at Stanford University. Sohl-Dickstein was one of the early researchers of diffusion models, a type of AI model. His main point is that AI could change how scientific research works within just a few years. If that happens, the traditional academic career path that young researchers have followed for decades may no longer provide the security they expect.
The talk was called "Advice for Young Researchers in the First and Last Days of the Anthropocene." Sohl-Dickstein supports his argument with several estimates. Over the past ten years, the computing power used to train the largest AI models has grown about 10 million times. He roughly estimates that this amount of computing is getting closer to the amount of computation the human brain carries out over a lifetime. He predicts that if this growth continues, the computing power used in 2030 could be about 10,000 times the level in 2025.
| Metric | Sohl-Dickstein's assessment |
|---|---|
| Growth in training compute of the largest models over the past decade | About 10 million-fold |
| Compute in 2030 relative to 2025 | About 10,000-fold |
| Length of software tasks a model can complete independently | About 10-fold growth in under a year |
| Progress in mathematics | Some Erdős problems began to be solved over the past roughly seven months |
Note: These figures are based on estimates and sources Sohl-Dickstein cited in his talk. The estimates of increases in computing power are approximate.
Sohl-Dickstein believes that one of the biggest risks in a world where technology grows exponentially is spending two years working on what seems like an important result, only to find that someone else can get a better result by giving the latest foundation model a single prompt after it becomes available. Because of this, he advises young researchers to choose projects with clear goals, work with other people and move quickly. Their research needs to keep up with the growing abilities of AI models.
This way of thinking leads him to question the value of the traditional academic career path. The talk was prompted by a conversation with a very capable graduate student. The student believed that artificial general intelligence (AGI), or AI with broad abilities comparable to human intelligence, could arrive within three years. Yet just minutes later, the student started planning to earn a Ph.D., complete a postdoctoral position and then apply for a faculty job. Sohl-Dickstein sees a clear contradiction: if you believe a world-changing technology is about to arrive, your other decisions should reflect that belief.
He openly admits that he does not know what academic research will look like in ten or twenty years. However, he is fairly sure that most jobs will look very different from the way they do today. People usually compare risks with possible rewards and choose what seems like the safest path. But if AI really could cause major changes across society, everyone already faces serious risks from changes to the whole system. Following the traditional safe path may not protect people from those risks. Instead, it may cause them to miss out on bigger opportunities. His advice is to do interesting work rather than simply choose what looks safest.
The line from the talk that people quote most often is: "Writing is thinking. If you outsource your writing, you are effectively outsourcing your thinking." He encourages researchers to use AI to find weaknesses in their arguments, raise objections and check their reasoning. But he does not recommend asking a model to write the first draft. Otherwise, researchers might end up with a paper that looks complete before they truly understand their own work.
When talking about tools, Sohl-Dickstein says he uses Claude every day. He compares it to a graduate student who is extremely good at carrying out tasks and willing to work hard, but who has clear weaknesses in research intuition and judgment. He now almost never writes code himself and rarely even reads code, but he still uses code every day to do research. He emphasizes that using more advanced tools means people do not have to understand every detail of how a computer works. However, they must be able to question the results an AI model produces instead of simply accepting its explanations.
Referring to Richard Sutton's 2019 essay "The Bitter Lesson," Sohl-Dickstein explains that, over time, general methods that make full use of computing power tend to outperform methods designed for specific problems using human knowledge. This lesson has two sides. If research can continue to improve as computing power and intelligence increase, its value could be much greater than people expect. But if a problem is likely to be solved simply by making AI models bigger and more powerful, researchers should seriously consider whether their work will still have its own value by the time they finish it. He also expects robots to enter laboratories and carry out real-world physical experiments within the next 10 to 15 years.
Sohl-Dickstein identifies several research areas worth investing in. One is using AI to advance science, including the discovery of new materials, the study of protein structures, weather prediction and fusion reactors. Another is studying AI models themselves, using ideas from fields such as interpretability, statistical physics, psychology and economics to understand how they work inside. Other important areas include AI safety, predicting future AI capabilities, fairness in technology, access to resources and policy and governance. He points out that AI safety still has many important problems that have not been clearly defined. He believes the importance of this field is much greater than the amount of research effort currently devoted to it.
Sohl-Dickstein also challenges the idea that "AI development has already surpassed the reach of individuals." Based on current trends, computing power in 2030 could be 10,000 times higher than in 2025. If these figures were shown on a straight-line chart, today's level of computing power would look almost like zero by comparison. This suggests that exponential growth is still at an early stage, when small decisions can have very large effects over time. He gives several examples: the standardization of email protocols, the Haber-Bosch process that made it possible to produce ammonia on a large scale, and the early medical and legal organizations in the United States. None of these developments was inevitable. Choices about technology and how institutions were organized early on helped shape the world we live in today.
The renewed interest in this talk also reflects an internal employment report that Anthropic recently made public. The report shows that junior employees made up 21% of the company's technical workforce in early 2025, but that share fell to 14%. Junior employees also made up 27% of technical hires at first, but that share dropped to 9%. Meanwhile, the company grew to three times its original size over the same period. In other words, the company expanded overall, but the entry-level route into technical jobs became narrower. The report also says that, by August of this year, AI agents working interactively under direct employee supervision were already working six times as many hours as human employees. The average number of code submissions made by each engineer also increased fivefold in six months.
Together, Sohl-Dickstein's warning and Anthropic's report point to the same problem. As AI models take over more entry-level tasks, newcomers have fewer chances to develop their own judgment. Yet judgment is exactly what people need to review AI-generated work and decide which research directions to pursue. For young people hoping to enter research or the technology industry, this means they may need to rethink how far ahead they plan their careers. It does not mean they should simply give up on further education. Instead, when choosing a research topic and career path, they should start by asking which skills and abilities are likely to become more valuable as computing power grows.