AI Enters a New Era: DeepMind Researchers Predict Shift to “Experience”-Based Learning

There are so many AI research papers published these days that it’s hard for any one to really stand out. But one recent paper has sparked a lot of discussion within the tech industry. A lot of people are calling it incredibly inspiring, and it's set the AI community abuzz.

A Shift in AI Development

Written by Google DeepMind’s VP of Reinforcement Learning, David Silver, and renowned computer scientist Richard Sutton, the paper boldly announces a new era for artificial intelligence. The authors identify two previous eras. The first, characterized by Google’s AlphaGo, a model that learned to master the game of Go in 2015, relied on digital simulations. The second, the era we’re currently in, is defined by OpenAI’s ChatGPT and its ability to generate human-like text. Now, Silver and Sutton propose we're entering a new period: "the Era of Experience".

Many see this as Google trying to address a major challenge for AI – the limited amount of training data available – while also charting a course away from the winning formula pioneered by OpenAI.

The Three Eras of AI

The Simulation Era

During the mid-2010s, researchers used digital simulations to help AI models learn by repeatedly playing games—chess, poker, Atari, even racing games—and receiving rewards for positive results. They were essentially incentivized to find strategies. Google’s AlphaGo emerged from this approach as did AlphaZero, which even redefined strategy in games like chess and Go.

However, this method was limited to specific problems with clear rewards, and wasn’t readily adaptable to real-world, open-ended scenarios.

The Human Data Era

A 2017 Google research paper, "Attention Is All You Need," shifted the focus to training AI models on vast amounts of data created by humans, readily available online. This led to the development of ChatGPT and other powerful generative AI tools. This approach essentially enabled AI to “pay attention” to human-created content and learn to perform various tasks like humans.

While impactful, this era has hit a snag due to the diminishing supply of high-quality human data, and the rising cost of acquiring it. The authors suggest something valuable was lost in the transition away from reinforcement learning.

The primary issue is that current AI relies heavily on human judgment. The models can’t really exceed what humans already know. They're limited by what we've already thought or written.

The Era of Experience

Silver and Sutton advocate for a new approach where AI agents actively interact with the real world to gather their own data and learn independently. They envision a shift towards AI systems that continually evolve based on their own experiences, removing the reliance on fixed datasets created by humans. The authors argue that “experience will become the dominant medium of improvement and ultimately dwarf the scale of human data used in today’s systems.”

Key Concepts: How This New Era Will Work

Streams

Current language models focus on short, isolated interactions. Information doesn't carry over between “episodes.” The new approach proposes “streams” – continuous lines of experience where AI can learn and adapt over time, much like humans. Imagine an AI language tutor that builds a long-term understanding of a student’s strengths and weaknesses, tailoring lessons over weeks or months.

Actions and Observations

AI agents will increasingly act autonomously in the world, beyond simply responding to human input. They’ll use machine-friendly interfaces like APIs, but also interact with the world through more traditional methods. This allows for exploration and discovery of strategies beyond human intuition.

Rewards

Instead of relying solely on human-defined rewards, AI agents will be able to create their own reward functions based on measurable signals from their environment – everything from heart rate and test scores, to economic indicators. This ground-level reward based learning can help advance AI’s understanding in ways previously impossible.

Planning and Reasoning

Current AI reasoning methods often mimic human thought processes, but the authors suggest this is inefficient and limited. They propose that AI should develop its own, potentially non-human, methods of thinking and planning. One possibility is creating "world models" – simulations that allow agents to predict the consequences of their actions.

Impact on the Cryptocurrency Market

These ideas are already having a real-world impact, notably in the cryptocurrency market. On April 16, 2025, following David Silver’s announcement, the AI token SingularityNET (AGIX) experienced a 5% price increase and a 20% surge in trading volume. Bitcoin (BTC) and Ethereum (ETH) also saw minor gains. This demonstrates a clear correlation between AI advancements and investor interest in AI-related cryptocurrencies. The market clearly believes these systems will lead to a better future.

Technical indicators – like the Relative Strength Index (RSI) and Moving Average Convergence Divergence (MACD) – also signaled a bullish trend for AGIX, supported by increased on-chain activity.

What the future holds

Silver and Sutton are confident that the technology needed to usher in the Era of Experience is already available – sufficient computing power, simulation environments, and reinforcement learning algorithms. They emphasize a willingness to embrace a new paradigm, moving away from the limitations of human-centric AI and towards systems that learn and think independently. The authors summarize and state that: “Experiential data will eclipse the scale and quality of human-generated data… This paradigm shift will unlock new capabilities.”

Ultimately, the goal is to create a more adaptable, intelligent, and autonomous AI that can solve problems in ways humans might not even imagine.