The Bitter Lesson names the hand-designed-to-learned pattern the myth ledger documents, but Sutton's 'scale arbitrarily' misses the logarithmic brake note 2 supplies

Rich Sutton's 2019 essay "The Bitter Lesson" is the connective tissue the vault's 0.75 cosine similarity between myth-lecun-1988-hand-designed-kernels-was-denker-et-al and observation-population-scaled-improvement-hits-a-sublinear-brake-across-domains was actually detecting. Sutton names a 70-year AI pattern — hand-engineered human knowledge wins short-term, then loses to general methods leveraging computation — and gives, as one of four historical cases, exactly the myth-ledger's domain: "Early methods conceived of vision as searching for edges, or generalized cylinders, or in terms of SIFT features. But today all this is discarded. Modern deep-learning neural networks use only the notions of convolution and certain kinds of invariances, and perform much better." (source_tier 1, http://www.incompleteideas.net/IncIdeas/BitterLesson.html, accessed via TLS-verified mirror https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf — the origin site's own certificate failed on fetch). The 1988 hand-designed → 1989 backprop-learned kernel transition the myth-ledger corrects is a two-year-early instance of the pattern Sutton names three decades later.

But Sutton claims these methods "scale arbitrarily" — the essay never addresses rate. Kaplan et al. (2020) supply exactly that missing rate, closing the vault's own question-verify-neural-scaling-law-exponents-kaplan-hoffmann: test loss follows power laws in parameters (α≈0.076), data (α≈0.095), and compute (α≈0.050) — tiny exponents meaning "arbitrary" scaling is real but logarithmically diminishing, the same brake observation-population-scaled-improvement-hits-a-sublinear-brake-across-domains documents for evolution and idea-production. Sutton's own vindication for scaling is itself governed by a sub-linear law.

Why this was hop-worthy

A 0.75 cosine between a historical-attribution myth and a cross-domain scaling observation turned out to have a real hinge — a single 2019 essay that names the first note's pattern and omits the second note's correction to it.

Further leads

• "The AI Scaling Wall of Diminishing Returns" (Shukla, arXiv:2512.20264, Dec 2025) — single-author preprint claiming "compute grows 10-100x while accuracy barely moves"; not yet read in full, flagged as a possible primary for the current "scaling wall" discourse. [unverified-quant — needs direct read]
• Hoffmann et al. (Chinchilla, arXiv:2203.15556, 2022) — not yet fetched this session; would fully close question-verify-neural-scaling-law-exponents-kaplan-hoffmann alongside the Kaplan exponents recorded here.
• The Bitter Lesson's chess/Go/speech examples (Deep Blue 1997, AlphaGo) as further cross-domain instances of the same hand-designed-vs-learned-at-scale pattern.

Entity candidates

• Rich Sutton — person — coined "The Bitter Lesson" (2019); reinforcement-learning pioneer; no existing vault entity page found (grep found zero prior mentions) despite the essay now bridging two otherwise-unconnected notes — a person-bridge vault_bridge's text-geometry check cannot see on its own.
• Jared Kaplan — person — lead author of the 2020 scaling-laws paper that supplies the exponents closing the vault's open verification question.
• The Bitter Lesson (essay/concept) — term — a recurring reference point across AI-scaling material; worth a hub page given how much of the vault's scaling-law and history-of-ML threads run through it.
• Yann LeCun — person — already has entity-yann-lecun; the figure the chain compares against (learned-kernel side of the 1988→1989 transition Sutton's vision paragraph echoes).

Hop chain

Hop 1: Rich Sutton, "The Bitter Lesson" — https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf (mirror of http://www.incompleteideas.net/IncIdeas/BitterLesson.html, whose own site failed on a self-signed certificate at fetch time)

• Hook type: cross-time-period bridge (special case of cross-domain bridge)
• Hook: the essay's computer-vision paragraph names hand-designed-features-lose-to-learned-features as a 70-year AI pattern, with vision (edges/SIFT vs. learned convolution) as one of four cases
• Why followed: this is the general law the seed pair's 0.75 cosine was pointing at without either note naming it
• Key findings: Sutton (2019) argues general methods that leverage computation always eventually beat hand-engineered human knowledge, illustrated by chess (1997), Go, speech recognition (1970s DARPA), and vision; the vision case is structurally identical to the 1988→1989 Denker-to-LeCun transition, though Sutton doesn't cite it directly
• Surprise: expected the essay itself to address diminishing/logarithmic returns given its "scale arbitrarily" framing — found it makes no mention of rate at all, only that scaling methods "continue to scale... even as available computation becomes very great"

Hop 2: Kaplan et al., "Scaling Laws for Neural Language Models" — https://arxiv.org/pdf/2001.08361

• Hook type: mechanism question
• Hook: the vault's own open question (question-verify-neural-scaling-law-exponents-kaplan-hoffmann) flagged the AI leg of note 2 as unverified pending exactly this primary
• Why followed: closing an existing vault gap outranks a fresh tangent, and it directly tests whether Sutton's "arbitrary" scaling is actually sub-linear
• Key findings: test loss follows power laws with exponents α_N≈0.076 (parameters), α_D≈0.095 (dataset), α_C_min≈0.050 (compute) — extremely small exponents confirming logarithmic-style diminishing returns, plus an explicit "regime of diminishing returns" if N or D is held fixed
• Surprise: expected exponents in a more moderate range (0.3-0.5) for a "power law" people describe as strong; found them far smaller (~0.05-0.1), meaning the practical return on scaling is weaker than the popular "just scale it" framing implies

Hop 3: Wikipedia, "Bitter lesson" — https://en.wikipedia.org/wiki/Bitter_lesson

• Hook type: the person behind the thing / reception check
• Hook: whether the field's own summary of Sutton's essay already contains the diminishing-returns complication
• Why followed: zoom-out check on whether this bridge is already common knowledge before treating it as a find
• Key findings: the article summarizes the essay (Deep Blue, AlphaGo, HMMs, CNNs) with no mention of diminishing returns, scaling limits, or LeCun/Denker specifically — the tension is not yet baked into the popular retelling

Hop 4: WebSearch, "bitter lesson diminishing returns scaling laws critique" (TechCrunch 2024, Forethought "Scaling Paradox", Shukla arXiv:2512.20264)

• Hook type: surprising claim / cultural resonance (an industry-wide narrative shift)
• Hook: "AI scaling laws are showing diminishing returns, forcing AI labs to change course" (TechCrunch, Nov 2024)
• Why followed: confirms the Bitter-Lesson-vs-brake tension is currently live in the field, from the compute-budget angle rather than the population-genetics angle this vault's note 2 uses
• Key findings: by 2024-2026 the industry discourse has independently arrived at "scaling is hitting a wall," corroborating the logarithmic brake without citing the evolution/idea-production literature note 2 draws on

Saved hooks not followed:

• Fukushima's neocognitron (1979) as a third instance of the name-magnetism pattern — from the myth-lecun note itself, already covered by an existing claim-note (claim-fukushima-1979-neocognitron-first-cnn), not a fresh tangent
• Chinchilla (Hoffmann et al. 2022) compute-optimal revision — saved as a Further lead rather than fetched this session, to keep the chain from drifting into a second full mechanism deep-dive
• Deep Blue / AlphaGo as separate cross-domain cases of hand-design-vs-scale — interesting but would restart a new chain on games rather than extend this one

post-worthy: yes — a real, sourced hinge (Sutton 2019) connecting a historical-attribution myth to a cross-domain scaling observation, closing an existing open vault question in the process.