Experience-first AGI: Carmack's Keen, Sutton's Oak Lab, the continual-learning bet

status: watch, opened 2026-09-13, last moved 2026-09-20. AI agi models

• The thread: the anti-LLM-orthodoxy research cluster — intelligence as continual learning from live experience, not frozen pretrained models. John Carmack's Keen Technologies and Richard Sutton's Oak Lab (spun out of Keen Jul 2026) are the named poles; adjacent: Graepel/Hafner ventures (deepmind-exodus), LeCun's AMI Labs (world-models thread), Silver's Ineffable Intelligence.
• 2026-09-13 (dig): Carmack's public record is thin but real. Upper Bound 2025 (May 2025; his first slide deck in 35 years): thesis — "LLMs can't be the whole answer," "LLMs can know everything without learning anything"; transformers don't learn like children; several fundamental problems have "not even a line of sight to an answer." Concrete artifact: a robot playing a real Atari 2600+ via camera + servo joystick ("the rootler") — point: reality is not a turn-based game; ~180ms action latency breaks state-of-the-art action-conditioned world models (BBF collapses with a few frames of delay). Real-world mess cataloged: scan-out, glare, worn servos, score-OCR reward detection broken by a moving tablecloth. Unsolved problems he's attacking on Atari suites: continual learning (sequential games destroy each other — catastrophic forgetting), transfer learning ("almost completely failed" for everything — OpenAI's Sonic challenge produced zero transfer; Gato negative transfer), sparse rewards/intrinsic motivation (curiosity, boredom, rage-quitting as generality signals), exploration beyond epsilon-greedy, latency-aware action conditioning (feed your own past actions as observations). Admitted missteps: hand-rolled CUDA instead of PyTorch, cloud avoidance, Sega Master System as environment. Promised open-sourcing of the rig + framework + a community benchmark (sequential Atari, no task IDs, sticky actions, no dedicated eval phase); not yet public. Replay-buffer experiment (28TB NVMe) promised "within weeks" of Jul 2025 — no public result.
• 2026-09-13: Sutton's departure is the cluster's status update — Oak Lab pitches real-time continual learning without replay (NetworkIDBD selective credit assignment; agent learns from one step at a time, no stored data) and the stated holy grail: a trillion-parameter agent that learns and plans in real time on 20 watts (brain power budget). OaK (Options and Knowledge): world model learned from interaction + planning. Sutton: current deep learning is "weak and inefficient... need not more tweaks, but fundamentally new ideas." He praised Carmack and Keen on the way out — "a slightly different path toward understanding intelligence." Same Sutton quoted in our Sep 12 ai-social record as the anti-doom voice. Carmack stays active on X but non-hype (Sep 11, 2026: hand-coding moving from "-jitsu to -do" — AI absorbing programming craft; Musk replied re technology in war). Full post context (GoKawiil capture): Carmack was reading Musashi's Book of Five Rings — the intro on swordsmanship's evolution from battlefield necessity to sport triggered the analogy; cites hand-assembling opcodes as the already-lost art, blesses the retro-computing scene as hobby, and lands on "Musashi would probably have been pretty enthusiastic about assault rifles." Musk's reply (Sep 11, one line — paraphrase only, no verbatim in any of Benzinga/IBTimes/ FrontierNews captures): technology is the "most important advantage" in war — an extension of the metaphor, not a rebuttal: technology (rifles) superseded a lifetime of swordsmanship; AI supersedes coding craft the same way. Musk escalating to the national-competition frame is its own tell. Keen: research-only, no product/revenue/new funding as of mid-2026 (7wData/PitchBook); keenagi.com is a bare landing page.
• Watch for: Keen's promised open-source Atari rig + benchmark (the cluster's first community artifact if it lands); Oak Lab's NetworkIDBD papers and any 20-watt demonstration; whether any experience-first lab shows a continual-learning result that beats frozen-model fine-tuning on transfer; Keen's replay-buffer vs one-step-learning resolution; Upper Bound 2027 (May 18–21, Edmonton) for the next round of public reveals.
• 2026-09-20 (backfill note — Carmack-timeline re-poll surfaced his Jul–Aug history; content dates to summer): Khurram Javed's own Jul 13 account confirms the split — "Last month, @RichardSSutton and I left Keen to do our own thing" (ALI/"Animal-Like Intelligence" framing in his posts; June-vs-Jul 14 date tension noted). Adjacent backfill: Sequoia's Training Data episode with Sutton + Javed (Aug 19); Carmack advising Keen researchers on compute discipline (Aug 12). Thread substance unchanged; identity confirmed by the principals.