The Turing Award winner who wrote the RL textbook just opened a neolab betting against the LLM playbook.
In July 2026, Richard Sutton — 2024 ACM Turing Award laureate, co-author of Reinforcement Learning: An Introduction, and the researcher most associated with temporal-difference learning — co-founded Oak Lab in Toronto with Khurram Javed. The pair left John Carmack's Keen Technologies to pursue what Sutton calls "fundamentally new ideas" for AGI: agents that learn from runtime experience, not curated static datasets.
A July 13 post from @MTSlive (~94K views) amplified the launch; @oaklab_ai published its first research note the same day — Learning from experience instead of curated datasets. explainx.ai maps what OaK is, why Sutton broke from Keen, and how this fits the neolab wave chasing continual learning while frontier labs double down on RLHF-scale alignment.
TL;DR — what people are asking
| Question | Answer |
|---|---|
| Who? | Richard Sutton + Khurram Javed — ex-Keen Technologies (Carmack) |
| Where? | Toronto, Canada — not Alberta (Amii/DeepMind Edmonton hub) |
| When? | July 2026 launch; site updated Jul 13 |
| Thesis? | RL from experience · world models · continual learning — not static PT |
| Architecture? | OaK — model-based RL; FC-STOMP abstraction loop |
| Moonshot? | ~1T params, real-time learn+plan, ~20W |
| vs LLMs? | Sutton: DL is "weak and inefficient" — needs reworking, not tweaks |
| Acquihire jokes? | Social thread bets Anthropic — no deal announced |
From Keen to Oak — same religion, different church
Sutton's departure statement (via X, reported by The Decoder and Office Chai):
| Shared with Keen | Oak-specific |
|---|---|
| Intelligence from runtime experience | Reject current deep learning as sufficient foundation |
| Reinforcement learning central | New algorithms — not scaling existing DL stacks |
| AGI ambition | Alberta Plan / OaK roadmap vs Carmack's engineering sprint |
Keen Technologies (Dallas) — Carmack's AGI bet, backers include Tobi Lütke — pursues human-level capability with heavy engineering. Sutton and Javed wanted to rethink learning machinery itself, not only ship faster on today's architectures.
explainx.ai read: This is the second high-profile RL-neolab fork narrative in 2026 — talent leaving well-funded AGI shops when algorithmic philosophy diverges, not only compensation.
OaK architecture — what Sutton has been pitching
Oak Lab is named for the OaK architecture — Sutton's NeurIPS 2025 invited talk and MIT CSAIL Dertouzos lecture (May 2026) outline:
As AI has become a huge industry, to an extent it has lost its way. What is needed to get us back on track to true intelligence? Most of all we need agents that learn continually from their first-person experience.
Three structural bets
- Everything learns continually — no frozen pretrained backbone with a thin finetune head
- Per-weight step sizes — meta-learned via online cross-validation (Sutton's step-size optimization line)
- FC-STOMP abstraction loop — grow structure over time:
- Feature construction
- SubTask posed from the feature
- Option learned to solve it
- Model of the option (world model chunk)
- Planning on that model
Prior research on oaklab.ai
| Paper / theme | Year | Relevance |
|---|---|---|
| The OaK Architecture: A Vision of SuperIntelligence from Experience | 2025 | Manifesto |
| The Alberta Plan for AI Research | 2023 | Multi-year roadmap |
| Horde — sensorimotor knowledge | 2011 | Many parallel value functions |
| SwiftTD, step-size optimization, columnar-constructive networks | 2023–24 | Continual / real-time learning machinery |
| Learning from experience instead of curated datasets | Jul 13, 2026 | First Oak Lab publication |
| Event-driven NNs, batch-size-one learning | Coming soon | Listed on site |
Thread replies praised Sutton's NeurIPS OAK talk — the lab productizes a ** decade-long Alberta Plan**, not a weekend manifesto.
The anti-LLM pitch — imitation vs evaluation
Sutton's June 2026 public framing (via press summaries): generative AI imitates but cannot evaluate its own outputs — blocking discovery.
| LLM industry loop (2026) | Oak Lab counter |
|---|---|
| Pretrain on curated web/code | Learn online from agent's own stream |
| RLHF/DPO on human/AI preferences | Reward from environment + internal critique |
| Frozen weights + periodic retrain | Continual weight + step-size updates |
| World as text | World models + planning |
That does not mean Oak ignores OpenAI's beneficial-trait RL results — it reframes them as patching imitation systems rather than building experience-native agents.
For builders, the actionable split:
- Ship products today → LLM + RLHF oversight stack
- Research AGI foundations → ask whether online RL + world models escape dataset ceiling and catastrophic forgetting
20 watts, trillion parameters — literal or lodestar?
Press reports Sutton's long-term goal: an agent with ~1 trillion parameters that learns and plans in real time on ~20 watts.
| Context | Scale |
|---|---|
| Human brain | ~20W — often cited in AGI rhetoric |
| Single H100 | ~700W — runs models orders smaller than 1T interactively |
| Frontier LLM training | Megawatts across clusters |
Plain read: The number is a north star for efficiency + continual learning, not a 2026 product spec. Oak Lab is an algorithm lab first — hardware co-design may follow (Keen DNA) but is not the launch headline.
Bitter Lesson tension — thread FAQ answered
Neolab skeptics on X asked: Doesn't Sutton's own Bitter Lesson say scale + general methods beat hand-crafted algorithms?
| Bitter Lesson core | Oak Lab response (implicit) |
|---|---|
| General methods win | Agree — but static backprop on fixed datasets may not be the winning general method for continual agents |
| Compute + data scale | Oak bets experience streams are the data — efficiency matters (20W) |
| Human knowledge hurts | OaK builds abstractions online (FC-STOMP) instead of hand-engineering features |
Sutton is not anti-scale — he is anti-"scale the wrong paradigm forever." Oak Lab is a neolab bet that 2020s DL is the new 2010s chess engines: impressive, industry-defining, not the final AGI substrate.
Neolab landscape — where Oak sits in July 2026
| Lab / thread | Focus |
|---|---|
| Oak Lab | Continual RL, OaK, experience > datasets |
| Keen | Carmack engineering AGI (Sutton alumni) |
| Frontier LLM labs | Scale PT + alignment RL |
| Physical AI | LeCun world models / Moravec |
| Agent harnesses | Claude Code, OpenClaw — tools atop LLMs, not new learning laws |
Social acquihire jokes (Anthropic buys Oak in months) reflect RL talent scarcity — Sutton is the canonical RL citation — not a disclosed deal.
Toronto vs Alberta: Riley RostovM and others noted Oak is not in Edmonton (Amii / historical Sutton base) or Vancouver — but Canada retains the research gravity Sutton built at University of Alberta for decades.
Who should care
| Audience | Takeaway |
|---|---|
| RL researchers | Watch oaklab.ai for batch-size-one / event-driven posts |
| LLM product teams | Oak does not invalidate your stack — it questions 10-year substrate |
| Investors | Neolab thesis diversity — not every AGI bet is more transformers |
| Policy / safety | Continual agents raise new monitoring problems — static eval harnesses may not transfer |
Summary
Richard Sutton co-founded Oak Lab in Toronto (July 2026) with Khurram Javed after leaving Keen Technologies — pursuing new AGI algorithms built on reinforcement learning from experience, the OaK architecture, and continual world models, not curated dataset scaling. The lab's first note dropped July 13; Sutton's moonshot remains a ~1T-parameter, ~20W, real-time planning agent. Social hype (~94K views) frames a neolab moment; the substance is a direct challenge to LLM pretrain + RLHF as the final path to intelligence — from the researcher who literally wrote the book on RL.
Related on explainx.ai
Update — July 14, 2026: Same week, Demis Hassabis published a frontier AI governance essay betting AGI arrives in a few years via scaling — the opposite timeline thesis from Oak's algorithm-first path.
- Scalable oversight — RLHF & Constitutional AI
- OpenAI beneficial trait RL — good generalizes OOD
- AI alignment introduction — outer vs inner goals
- What is fine-tuning — RLHF stage explained
- Yann LeCun — LLMs, physical agents, continual learning
- History of AI — DeepMind RL lineage
- Teaching Claude why — alignment vs capability
Sources: Oak Lab · NeurIPS 2025 — OaK invited talk · MIT CSAIL — OaK lecture May 2026 · The Decoder — Jul 13, 2026 · Office Chai · @MTSlive Jul 13, 2026
Oak Lab claims and Sutton quotes reflect July 2026 public announcements. AGI timelines and hardware targets are aspirational — not verified product commitments.