Richard Sutton, the 2024 ACM Turing Award laureate and co-author of the foundational "Reinforcement Learning: An Introduction" textbook, has launched a direct challenge to how the AI industry approaches alignment and intelligence. In July 2026, Sutton co-founded Oak Lab in Toronto with Khurram Javed after departing John Carmack's Keen Technologies, signaling a fundamental disagreement about the path to artificial general intelligence (AGI). Rather than scaling large language models (LLMs) with reinforcement learning from human feedback (RLHF), Sutton's new lab is betting on agents that learn continuously from real-world experience.

Why Did One of AI's Most Respected Researchers Leave a Well-Funded Lab?

Keen Technologies, backed by Shopify founder Tobi Lütke, represents the engineering-first approach to AGI: build faster on today's architectures. Sutton and Javed wanted something different. They left to rethink the fundamental learning machinery itself, not merely accelerate existing deep learning stacks. This departure reflects a growing philosophical divide in the AI research community: some believe scaling transformers with RLHF is the path to AGI, while others argue the entire paradigm may be a dead end.

Sutton's public framing is blunt. In a June 2026 statement, he argued that generative AI imitates but cannot evaluate its own outputs, blocking genuine discovery. The current LLM industry loop relies on pretraining on curated web and code data, then applying RLHF or similar techniques based on human or AI preferences. Sutton sees this as patching imitation systems rather than building agents that truly learn from experience.

What Is the OaK Architecture, and How Does It Differ From RLHF-Based Alignment?

Oak Lab is named for the OaK architecture, which Sutton outlined in a NeurIPS 2025 invited talk and MIT CSAIL lecture in May 2026. The core principle is radical: everything learns continually from first-person experience, with no frozen pretrained backbone and thin fine-tuning head. Instead of static weights updated periodically, OaK uses per-weight step sizes that are meta-learned through online cross-validation.

The architecture includes a feature called FC-STOMP, an abstraction loop that grows structure over time. Here's how it works:

• Feature Construction: The system builds features from raw experience streams
• SubTask Identification: A subtask is posed based on the constructed features
• Option Learning: The agent learns an option (a reusable skill) to solve that subtask
• World Model Development: A model of the option is created, representing how the world responds to that skill
• Planning: The agent plans using the learned world model rather than relying on static text representations

This stands in sharp contrast to RLHF, which fine-tunes a frozen model using human preferences. RLHF is a post-hoc alignment technique applied to models trained on static datasets. OaK proposes that alignment should be built into the learning process itself, from the ground up.

What Is Sutton's Moonshot Goal, and Is It Realistic?

Press reports indicate Sutton's long-term vision: an agent with approximately one trillion parameters that learns and plans in real time while consuming only about 20 watts of power. For context, the human brain uses roughly 20 watts, while a single H100 GPU consumes around 700 watts and runs models orders of magnitude smaller than one trillion parameters interactively. Frontier LLM training consumes megawatts across entire data center clusters.

The 20-watt target is best understood as a north star for efficiency and continual learning, not a 2026 product specification. Oak Lab is an algorithm research lab first; hardware co-design may follow but is not the launch headline. The lab published its first research note on July 13, 2026, focused on learning from experience instead of curated datasets.

How Does This Challenge Sutton's Own "Bitter Lesson"?

Skeptics on social media quickly pointed out an apparent contradiction: Sutton's famous "Bitter Lesson" essay argues that general methods combined with scale and compute beat hand-crafted algorithms. Doesn't Oak Lab contradict that principle? The implicit response is nuanced: Sutton agrees that general methods win, but he questions whether static backpropagation on fixed datasets is the winning general method for continual agents. Oak Lab bets that experience streams are the data, and efficiency matters. Rather than hand-engineering features, OaK builds abstractions online through FC-STOMP.

Sutton is not anti-scale. He is anti-"scale the wrong paradigm forever." Oak Lab positions 2020s deep learning as analogous to 2010s chess engines: impressive, industry-defining, but not the final substrate for AGI.

What Do Industry Observers Make of This Move?

The launch generated significant social media attention, with a July 13 post from @MTSlive reaching approximately 94,000 views. Social media threads included acquihire jokes suggesting Anthropic might quickly acquire Oak Lab, reflecting the scarcity of top reinforcement learning talent. However, no deal has been announced.

Oak Lab's location in Toronto, rather than Edmonton (where Sutton built the University of Alberta's AI research gravity for decades) or Vancouver, was noted by observers. However, Canada retains significant research momentum in reinforcement learning and continual learning research.

For different audiences, the implications vary. Reinforcement learning researchers are watching oaklab.ai for publications on batch-size-one and event-driven neural networks. LLM product teams should note that Oak does not invalidate their current stacks but questions the 10-year substrate. Investors see neolab thesis diversity, recognizing that not every AGI bet is more transformers. Policy and safety researchers face a new challenge: continual agents that learn from real-world experience raise novel monitoring problems, and static evaluation harnesses may not transfer to such systems.

How Does This Fit Into Broader Alignment Research Trends?

Oak Lab's emergence represents the second high-profile reinforcement learning neolab fork in 2026, with talent leaving well-funded AGI shops when algorithmic philosophy diverges, not merely for compensation reasons. This signals a maturing field where researchers increasingly disagree on fundamental approaches.

The timing is notable: the same week Oak Lab launched, Demis Hassabis published a frontier AI governance essay betting that AGI arrives in a few years via scaling, the opposite timeline thesis from Oak's algorithm-first path. This divergence reflects genuine uncertainty about whether the current LLM plus RLHF paradigm is sufficient or whether new learning algorithms are necessary.

Sutton's departure from Keen Technologies and launch of Oak Lab underscore a critical debate in AI alignment: Is the path to safe, capable AGI through better alignment techniques applied to existing architectures (the RLHF approach), or through fundamentally new learning algorithms that align safety into the learning process itself? Oak Lab's bet is that the latter is not only possible but necessary.