Richard Sutton on LLM Scaling: Bitter Lesson Pushback in November 2026

Richard Sutton — author of the 2019 Bitter Lesson and 2024 Turing Award co-laureate for reinforcement learning — posted on X around November 8, 2026 that “brute force computation is not enough for intelligence,” that “clever new methods are needed,” and that “the science of AI is far from finished.” The Decoder published a Sunday piece tying the post to Jerry Tworek’s critique of transformer-only scaling and to counter-quotes from Sam Altman and Dario Amodei. For builders, the fight is not academic nostalgia. It is about whether today’s recipe — massive pretraining on web text, then alignment passes — is a plateau or a runway.

explainx.ai already covered Sutton’s Oak Lab startup bet on algorithms that learn from experience. This article tracks the November public flare-up, not a new Oak release.

TL;DR

What Sutton actually argued in 2019

Reward-hacking shortcut under an obstacle, echoing Tworek’s claim that transformer scaling misread the Bitter Lesson

The Bitter Lesson is short and blunt. Sutton’s opening claim: over 70 years, general methods that leverage computation beat approaches that embed human domain knowledge — by a large margin. Chess, Go, speech, and vision each show researchers clinging to handcrafted structure until search and learning at scale win.

Two methods, in Sutton’s account, scale especially well: search and learning. That is why his 2026 X-line is not “algorithms do not matter.” It is closer to: the algorithms we are over-investing in today may be the wrong general methods.

In explainx.ai’s small-model economics thread, Hacker News commenters already fought over how to read the Bitter Lesson — scale vs cost vs task fit. November’s Sutton post reopens the same parse error at lab-founder scale.

What is new in November 2026

The Decoder’s November 8 story is a synthesis piece, not a paper drop. Its news value is timing:

1. Sutton’s X post restates that compute without new methods is insufficient.
2. Social uptake on X reportedly made the line a community flashpoint the same weekend.
3. Tworek is quoted framing misread Bitter Lesson theology as damaging — “reward-hacked the brains of a whole generation of scientists” in The Decoder’s paraphrase.
4. Jayden Teoh adds a moderation angle: promising ideas were dismissed for not looking “Bitter Lesson enough,” even when they later enabled cleaner general methods.
5. Frontier Radar #5 is cited as tracking 45 neolabs betting on gaps Sutton and Tworek describe.

None of that replaces peer review. It does tell builders what senior researchers are arguing about in public while METR acceleration thresholds and frontier model launches dominate product news.

Sutton’s longer-standing LLM critiques (context)

The Decoder reminds readers Sutton did not spring from nowhere:

• He has criticized scaling LLMs alone as insufficient for continual intelligence.
• He has called reliance on synthetic data a “big mistake” in summarized accounts — the real world is vastly larger than any simulation.
• He wants agents with world models that keep learning without catastrophic forgetting — aligned with Oak Lab’s public story in our July coverage.

If you build agents that loop for days, that is the tension: static weights versus runtime learning. explainx.ai’s harness reliability work shows how brittle long-horizon agent stacks are even without solving Sutton’s continual-learning problem.

Tworek and the neolab wave

Rising scaling curve with green points along it, framing the compute-versus-algorithm argument

The Decoder places Tworek as an ex-OpenAI researcher now running Core Automation, hunting algorithms beyond pretraining plus reinforcement learning. That sits in the same bucket as Sutton’s Oak Lab, John Carmack’s Keen lineage, and other “neolabs” explainx.ai has tracked when founders leave frontier labs to chase non-transformer-default paths.

For practitioners, neolabs matter as option value: most production code still calls frontier APIs or open-weight transformers. Neolabs are where failure modes get explored early — continual learning, world models, planning at low watts — even when shipping calendars stay on GPT-class and Claude-class models.

The scaling optimists’ side in The Decoder’s account

The Decoder quotes Sam Altman arguing LLM scaling is not exhausted and pointing to a model that reportedly disproved a math conjecture humans had not cracked. It quotes Dario Amodei describing scaling as a “rainbow with no end.” The piece notes both men lead companies whose economics assume frontier scaling continues — disclosure that matters when you weigh quotes.

Builders should read those lines as strategic commitments, not as falsified science. Altman’s example, if accurately reported, is evidence that current stacks still surprise on some closed domains. Sutton’s counter is that surprise on math benches does not equal continual learning in the open world.

Jayden Teoh’s nuance — ideas do not have to “scale forever”

The Decoder summarizes Jayden Teoh arguing that ideas once seen as un-Bitter-Lesson were later replaced by cleaner general methods — without making the original ideas “wrong.” Research culture, in that telling, over-punishes heterodox work because scaling theology became social law.

That matters if you fine-tune, distill, or invent architectures that will not train at trillion-parameter scale on day one. The community may still adopt your trick as a stepping stone — a pattern visible in open-weight license and merge culture even when headline labs stay centralized.

What this means for what you build this month

Sutton’s November post will not change your Tuesday API bill. It should change how you read roadmap claims:

| If you mostly consume APIs | Keep evaluating models on your evals, not on scaling slogans alone. | | If you train or fine-tune | Treat static datasets as a ceiling; plan for refresh and regression. | | If you build agents | Separate harness engineering (agent skills, MCP) from learning at runtime — today’s default is still prompt plus tools, not Sutton-style continual RL. | | If you pick small models | Cost wins in local MLX stacks remain real even if frontier scaling continues. |

Chess, Go, and speech — why Sutton’s essay still gets misread

The 2019 essay is not abstract philosophy. Sutton walks chess, Go, speech, and vision in sequence. Each time, teams invested in human structure until search plus learning at scale won — often bitterly, for the researchers who bet on craftsmanship.

That history is why Tworek’s “reward hacked” line lands: the community may be repeating the premature human-structure phase, but this time with transformer scaling as the structure. Sutton’s November post says the next chapter still needs new methods, not only more FLOPs on the same architecture.

Readers who only know Sutton from memes should read the original once. It is short, dated March 13, 2019, and free on incompleteideas.net.

Altman’s math example and the limits of anecdotal scaling proof

The Decoder reports Altman citing a model that disproved a math conjecture. Treat that as an existence proof for surprises, not as proof that scaling laws have no knee:

• A single benchmark win does not show continual learning in production agents.
• Math domains with generatable training data are the easy regime Sutton himself separates from messy real-world interaction.
• Corporate leaders have structural incentives to emphasize runway; Sutton has structural incentives to emphasize unexplored algorithm space.

Good engineering keeps both ideas in head at once: ship with frontier APIs while watching neolab outputs for something that changes your fine-tuning or RL stack.

Classroom and hiring effects

Jayden Teoh’s point — ideas dismissed for not “looking Bitter Lesson enough” — has HR-shaped consequences. Labs may filter research proposals that do not mention scale in the first paragraph. Startups may over-rotate to MoE width because investors recognize the vocabulary.

If you hire researchers or staff applied scientists, ask for problem-first pitches, not scale-first pitches. If you are junior, document small ideas that unlocked big runs — Teoh’s argument is that those stepping stones deserve credit even when a later general method supersedes them.

Synthetic data, world models, and robotics

Green chip beneath a cloud outline, contrasting cloud-scale LLM training with embodied world models

The Decoder ties Sutton’s synthetic-data skepticism to robotics world models as a place scaling optimists still invest — Altman is quoted calling world models useful while favoring more LLM scaling. That split is the practical debate: language priors versus sensorimotor experience.

Readers following Gemini robotics or NVIDIA embodied stacks should hear Sutton as warning that internet text is a thin slice of state space — not as telling you to uninstall your simulator.

Related reading

• Richard Sutton Oak Lab — algorithms beyond static LLM training
• Small models, Luna economics, Bitter Lesson HN thread
• METR and Anthropic acceleration thresholds
• Open-weight licenses and merge culture
• Agent skills complete guide
• Harness leaderboard reliability
• MLX on Apple Silicon local LLM guide

Sutton quotes and third-party paraphrases are accurate as of November 8, 2026, per The Decoder and the 2019 Bitter Lesson essay on incompleteideas.net — not as a verified full-text archive of Sutton’s X post.