Late last month, Julian Schrittwieser, a core contributor to the legendary DeepMind projects (AlphaGo Zero, MuZero) and a current top researcher at Anthropic, published a blog post titled "Failing to Understand the Exponential, Again," which quickly went viral in the AI community.
The reason is simple: this blog post hit upon a critical question: Our understanding of AI's growth rate may have been consistently flawed.
In a recent in-depth interview, Julian systematically elaborated on the true trajectory of frontier AI development, specific predictions for the next year or two, and the key evolutionary path from AlphaGo to AGI.
Following this top researcher's perspective, we will discover astonishing truths.
Portal: https://www.julian.ac/blog/2025/09/27/failing-to-understand-the-exponential-again/
I. Why Does the Public Repeatedly Misjudge AI?
Julian points out that humans are consistently slow in grasping "exponential growth." This cognitive illusion appeared early in the COVID-19 pandemic—even though data was exploding, society still felt it was far away. The same is true for AI today.
Julian summarizes two major cognitive biases of the public:
1. Ignoring the Exponential Curve.
Most people still view AI's progress from a linear perspective, mistakenly believing it is "slowing down" or will halt. However, in data from frontier labs, AI capabilities are growing extremely steadily and exponentially, completely contrary to the "bubble theory" narrative prevalent in mainstream discourse.
2. Making Permanent Judgments based on Current Flaws.
Seeing AI occasionally make mistakes now (e.g., writing programs, designing websites), they conclude it will never reach human levels. This leads to the conclusion that AI will never be able to perform these tasks at a human level, or will only have a small impact.
Yet, a few years ago, getting AI to do these things was pure science fiction. Some people, seeing little difference in dialogue between two consecutive model generations, assert that AI has peaked and the era of scaling is over.
Although discussions about an artificial intelligence bubble are rampant on social media, Julian sees a different picture in frontier labs:
"We haven't seen a slowdown in progress; instead, we've seen very steady progress that has continued for years."
This progress can be precisely depicted with data.
II. AI's Long-Task Capability Doubles Every 7 Months
According to the research report "Measuring AI Ability to Complete Long Tasks" by the independent organization METR (Model Evaluation and Testbed for Robustness),
Frontier models show a stable doubling curve in their performance on task duration.
• In mid-2024, Claude Sonnet 3.5 could independently complete tasks for only about 30 minutes;
• By 2025, Sonnet 3.7 could work autonomously for about 1 hour continuously.
Extrapolating this trend, AI's long-task capability doubles every 7 months. The latest generation models like GPT-5, Grok 4, and Opus 4.1 have already surpassed this trend prediction, capable of handling tasks longer than 2 hours.
Julian points out:
"This means models can already operate independently on medium-sized projects without human real-time intervention."
The improvement in these capabilities is not "linear addition" but exponential multiplication.
If we continue to extrapolate along METR's trend line, in just over a year, AI will be able to work continuously for a full workday.
Some might object: Is it appropriate to generalize from software engineering tasks to a broader economic scope?
III. AI Is Approaching Human Experts, Delegability Is Expanding
Fortunately, we have another study to reference: OpenAI's GDPval assessment—which measures model performance across 44 occupations in 9 industries.
In this assessment:
• Each task was designed by industry experts with an average of 14 years of experience,
• Models and human solutions were compared in a blind evaluation.
The results show: GPT-5 and Claude Opus 4.1 are already close to the average level of human experts.
These cross-industry data illustrate:
AI's "delegability" is rapidly expanding.
If a model requires human feedback every 10 minutes, its efficiency would be greatly limited; but when a model can work continuously for several hours and match experts at a professional level, it is no longer just a tool copy, but a "virtual collaboration partner" that can be delegated tasks in batches and managed as part of a team.
Julian believes that even conservative trend extrapolation is sufficient to conclude that 2026 will be a key year for AI's widespread integration into the economy.
"Based on simple linear fitting extrapolation of data and past trends, AI is predicted to achieve round-the-clock autonomous work within the next one to two years, and reach or even surpass expert levels in specialized fields."
He provides three key predictions for time nodes:
• By mid-2026, models will be able to work autonomously for a full day (8 working hours).
• Before the end of 2026, at least one model will match human expert performance across multiple industries.
• By the end of 2027, models will frequently surpass experts in many tasks.
These are not fantasies, but results of data extrapolation. Julian says this may be more reliable than many "expert judgments."
IV. After the AI Productivity Revolution, There Is an AI Creativity Revolution
Not only that, AI's value is no longer just "efficient work." Its creativity is also rewriting the pace of scientific exploration.
Julian's understanding of AI creativity stems from his personal experience at DeepMind. In 2016, AlphaGo competed against top Go players and made the "move 37," which shocked the industry. That move was highly unconventional, surprising even professional Go players.
The significance of this move was immense—it proved that AI is not just mechanically calculating the optimal path, but can also make genuinely novel and creative decisions.
Current language models also possess this creativity. They can generate an unlimited amount of novel content, such as new code or new paper snippets.
The real challenge is not "novelty," but "useful novelty."
To achieve this, the task must be difficult and interesting enough, and AI must be able to judge the quality of ideas. It needs to forge new paths while ensuring these paths have practical value.
Today, this creativity is already being applied to scientific discovery.
AlphaCode can find new programs, AlphaTensor can discover new algorithms. Google DeepMind and Yell are also making new discoveries in the biomedical field using AI.
Julian predicts:
"Perhaps next year, we will see discoveries made independently by AI that are significant enough to shake the scientific community."
Regarding current controversies, he remains very optimistic:
Some results are still controversial, but the process is always advancing, and when the evidence is clear enough, the controversy will naturally disappear.
Even more exciting are Nobel Prize-level breakthroughs.
Julian even predicts that by 2027 or 2028, AI models will be smart enough to independently make scientific breakthroughs worthy of a Nobel Prize.
In the future, AI may even challenge for the Fields Medal in mathematics. It will help us unlock the mysteries of the universe and improve human living standards.
In other words, after the AI productivity revolution, an "AI creativity revolution" is brewing.
V. AGI's Path and Challenges: The Combination of Pre-training and Reinforcement Learning
When discussing AGI, Julian's judgment is extremely clear:
"No new mysterious technology is needed; the paradigm of 'pre-training + Transformer + reinforcement learning' is sufficient to achieve human-level intelligent systems."
The effectiveness of this paradigm is vividly demonstrated in the evolution of the AlphaGo series:
• AlphaGo defeated top Go players relying on deep networks and self-play
• AlphaGo Zero took this process to another level, completely removing reliance on human knowledge, starting from scratch with self-play, surpassing the original in days
• AlphaZero generalized the logic to chess and Shogi, achieving cross-game versatility
• Mu0 then extended the framework to real-world reinforcement learning problems
In his view, AGI will not be a sudden singularity but a smooth curve.
The increasing technical difficulty is real, but as long as productivity gains offset the increase in research costs, progress will not stagnate.
As for whether pre-training will be abandoned in the future in favor of only reinforcement learning, Julian's answer is "most likely not."
"Pre-training brings some interesting safety perspectives. Creating an agent with values similar to ours"—pre-training is not only efficient but also helps AI align with human values.
Perhaps some will train "AI from scratch" for scientific interest, but on a practical level, "pre-training + reinforcement learning" remains mainstream.
Ultimately, AI's goal is not "superintelligence," but to solve global problems such as climate change, healthcare, and education.
"We need to ensure that artificial intelligence serves humanity, not opposes it"—this is the most core principle in technological evolution.
VI. Conclusion
Julian's sharing actually helps us correct a cognitive bias: Don't use "current AI" to judge "future AI."
Because it is growing exponentially.
What seems "impossible" today might become "routine" next year.
From working autonomously for a full day, to catching up with or even surpassing experts, and then helping humanity win Nobel Prizes—AI's evolutionary speed is faster than we imagine.
As Julian says:
"Artificial intelligence is a tool—a powerful tool that can help us solve problems and achieve things once thought impossible."
But it is ultimately a tool. Whether it can be used responsibly and ensure it benefits all humanity is in our hands.
How good the future is depends on whether we can use this tool well and ensure it stands with humanity. We have a huge opportunity ahead, and how to make the most of it depends on us.
The next two to three years will be a critical period for AI to change the world. We should remain attentive, and also rational—neither underestimating its potential nor ignoring its challenges.