YouTube - Julian Schrittwieser (Anthropic) - Underestimated AI Exponential Trajectory - RL Driven Agents to Surpass Experts by 2027, Emphasizing Alignment and Non-Zero-Sum Growth

The Underestimated Exponential Leap of AI: Agents May Match or Surpass Human Experts by 2027

In this era of unprecedented technological velocity, we often default to linear thinking, failing to fully grasp the true nature of exponential growth. When it comes to the current rapid advancements in AI, this cognitive bias may lead to a severe underestimation of the technology’s genuine potential and impact. Julian Schrittwieser, a researcher at Anthropic and a core contributor to DeepMind’s landmark projects, including AlphaGo and MuZero, posits unequivocally that the progress in AI capability is following a stable exponential trajectory—a trend he believes is widely overlooked by both the public and investors. He boldly predicts that by 2026 to 2027, AI Agents will possess capabilities rivaling or even surpassing those of human experts, heralding a massive wave of non-zero-sum growth across knowledge work, scientific discovery, and global economic productivity.

The Overlooked Exponential Curve: AI’s "COVID Moment"

With his characteristic insight, Julian Schrittwieser likens the current pace of AI development to the early stages of the COVID pandemic. Initially, people dismissed isolated cases, but once the mathematical reality of viral spread—where cases doubled every one to two weeks—was understood, the inevitability of a large-scale outbreak became clear. He argues we are currently experiencing a similar "COVID Moment" in AI: while the surface may appear calm, progress within frontier labs has never slowed.

The key metric is astonishing: AI capability continues to improve steadily, with the models becoming able to independently complete a task twice as long as the previous benchmark approximately every three to four months. If this trend continues to extrapolate, top-tier AI models will be capable of working completely autonomously for an entire day, or even longer, within the next one to two years. This exponential increase in long-horizon autonomous task capability carries immense potential for restructuring the vast landscape of knowledge work within the global economy. Julian Schrittwieser suggests that the current market valuations and revenue projections for frontier labs like OpenAI, Anthropic, and Google may be overly conservative. He observes a unique duality: while the broader AI ecosystem might contain speculative bubbles, these core frontier labs are positioned on an exceptionally robust growth track.

The 2026-2027 AI Capability Outlook: Task Length and Expert Performance

Julian Schrittwieser bases his specific predictions for future AI capabilities on linear extrapolation from benchmarks like Meta Eval, emphasizing that statistical models are often more reliable than individual intuition. He argues that “Task Length” is the core metric for measuring the delegatability of an LLM/Agent. Even the smartest model’s real-world value diminishes significantly if it requires frequent user intervention, such as every 10 minutes.

Based on this, he offers three key forecasts:

• Mid-2026: AI Agents will be capable of working autonomously for an entire day, completing large-scale tasks such as developing a full application's functionality or drafting a comprehensive research report.
• End of 2026: At least one AI model is projected to reach the level of an industry expert across numerous professions.
• 2027: AI models will frequently outperform human experts on a wide array of tasks.

These predictions are not baseless speculation. OpenAI’s GDP Eval, a benchmark designed to measure AI progress across multiple professions by collecting real tasks from domain experts, confirms the rapid approach of AI toward producing significant economic impact. Julian Schrittwieser points out that the “messiness” of a task is highly correlated with its “time length”; a task requiring 8 hours or 16 hours of human effort inherently involves real-world complexity. Therefore, the Agent’s ability to genuinely handle this “messiness” will be the critical factor determining its practical utility.

AI Exponential Growth

Figure 1: Julian Schrittwieser discussing AI's exponential growth during a YouTube interview. Image Source: YouTube

The DeepMind Legacy: Reinforcement Learning (RL) as the Agent’s Key

Julian Schrittwieser’s profound insights are informed by his invaluable experience at DeepMind. He details the series of breakthroughs in the RL domain that laid the foundation for the rise of modern AI Agents:

• AlphaGo (2016): The first AI to defeat a top human Go player, merging deep neural networks with Monte Carlo Tree Search, marking a major AI milestone.
• AlphaGo Zero (circa 2017): Stripped of all human Go knowledge, it trained entirely through self-play, rediscovering all Go strategies from scratch, demonstrating an immense capacity for tabula rasa learning.
• AlphaZero (circa 2018): The algorithm was generalized to successfully compete in three different games—Chess, Go, and Shogi—proving the versatility and generalization capability of RL algorithms.
• MuZero (co-authored by Julian Schrittwieser): Solved the problem of environments that cannot be perfectly simulated. Its key innovation was training the model to predict the future state of the environment, meaning the model could learn what would happen after each action without needing a complete world model, crucial for handling robotics or other complex real-world tasks.

These breakthroughs demonstrate that RL is not just a powerful tool for games but is fundamental to building Agents capable of understanding and interacting with complex environments.

Reinforcement Learning (RL) and the Rise of Modern AI Agents

Julian Schrittwieser emphasizes that games serve as an excellent "sandbox" for RL science, allowing for the rapid verification of algorithms and the construction of robust learning systems. He notes that modern LLMs also possess an "Implicit World Model." To predict the next most probable word or passage, the model must internally model the underlying state of the world, mirroring MuZero's approach of implicitly predicting action consequences rather than rebuilding the full world state.

RL is the key to unlocking Agent capabilities. An Agent is defined as an AI that can act independently without frequent user interaction. However, pre-training data often lacks examples of actions, how humans interact with the world, and crucially, how Agents fail and self-correct. RL provides the mechanism for Agents to interact directly with the environment, reinforcing correct actions and learning from mistakes, thereby increasing robustness.

RL is not without its challenges. It is significantly harder to debug than supervised learning and involves feedback loops that make isolating the source of problems difficult. Julian Schrittwieser points out that AlphaZero generated extremely high-quality training data through massive planning and search, making its RL training highly stable. In contrast, the RL training for modern LLMs is relatively less stable, requiring stability to be improved by investing more reasoning capability into the models to generate higher-quality training data.

The Spark of AI Creativity: From "Move 37" to the Nobel Prize

Julian Schrittwieser cites AlphaGo’s famous "Move 37" against Lee Sedol to illustrate AI’s creative capacity. This move was unexpected, unconventional, and astonished professional players, demonstrating that AI could engage not just in pure calculation but also in innovative thinking.

He confirms that modern LLMs clearly possess a novel creative capacity, for instance, in writing code or papers, because they are trained to generate a probability distribution, allowing for the sampling of infinite novel sequences. The true difficulty lies in generating things that are both novel and useful.

Regarding when AI might achieve a breakthrough worthy of a Nobel Prize, Julian Schrittwieser remains optimistic. He predicts that by 2027 or 2028, AI models will highly likely be smart and capable enough to independently make a discovery or insight of Nobel Prize caliber. His genuine excitement centers on AI's potential to accelerate human scientific progress, unlock the mysteries of the universe, and ultimately improve human living standards.

Proceeding with Prudence: Safety, Alignment, and Societal Impact

In the face of AI’s exponential development, safety and Alignment become paramount. Julian Schrittwieser discusses Goodhart's Law in the context of AI benchmarks: "Any measure that becomes a target ceases to be a good measure." He recommends creating new "held out benchmarks" or internal benchmarks periodically to truly gauge model performance, rather than over-optimizing for publicly available ones.

On the front of explainability, he highlights the importance of Mechanistic Interpretability, giving an example of how researchers at Anthropic identified and modified specific neurons within the Claude model to influence the model's “preferences.” He stresses that as models become smarter, understanding their internal thoughts, values, whether they are lying, and whether they genuinely follow instructions is critical for safety.

Anthropic places immense importance on safety and Alignment. Julian Schrittwieser states that before any model release, they conduct rigorous analysis and verification to ensure the model will not exhibit harmful autonomous behavior or be maliciously exploited. He explicitly notes that if they are uncertain about a model’s safety, they will delay its release, signifying that safety takes precedence over immediate financial returns. Alignment is viewed as a full-stack problem, involving data filtering, behavioral monitoring, and safety guidelines embedded in system prompts.

Regarding AI's impact on society and employment, Julian Schrittwieser sees AI and human intelligence as complementary rather than simple substitutes. AI is superior to humans in some tasks and inferior in others; therefore, people will incrementally use AI to boost their own productivity. He firmly believes that AI will drive massive wealth creation, making the social outcome a non-zero-sum game by "raising the floor" for everyone’s achievement. He asserts that the most important objective is to "grow the pie." If society can increase everyone's productivity by 10x, advancements in medicine, energy, and materials science would no longer be constrained by human intelligence. How to distribute the wealth generated by this productivity surge, he concludes, is a political and social problem, not a technical one.