Revisited

Bush's differential analyzer is the cleanest historical charter for the analog-accelerator revival: continuous-time analog cores inside a digital wrapper, now with memristors instead of wheel-and-disc integrators.

Tolman and Honzik's latent-learning protocol—reward-free pretraining, then instant exploitation when reward appears—deserves to become the standard data-efficiency benchmark for embodied world models.

Uexküll's Umwelt is constantly cited and almost never implemented; body- and action-conditioned perceptual objectives for robot foundation models would finally operationalize it.

Cajal's structural-plasticity hypothesis points at learning by rewiring—dynamic sparse growth and pruning—as an underused alternative to pure weight updates, now testable at scale.

The optophone's fatal flaw was its fixed encoding; ML-optimized, online-adapted sensory-substitution codes have never been seriously tried and could transform assistive interfaces.

Helmholtz's unconscious inference is broadly vindicated, but the strong claim—cortex literally inverting a learned generative model—is now testable with digital-twin visual cortex models.

Gabor's holography, read as the computing rather than imaging branch, points to photonic/holographic accelerators and associative memories for transformer inference.

The order-of-magnitude information advantage of spike-timing codes argues for temporal-coding-first sensors and neuromorphic hardware that AI still largely ignores.

REM sleep as fundamental offline generative activity invites engineering explicit consolidation/replay phases into deployed models and testing them against decoded biological replay.

Friedberg's program evolution was hopeless with blind mutation but is reborn as LLM-guided program search (FunSearch, AlphaEvolve)—worth revisiting as both method and cautionary baseline.

Patient H.M. defined the fast-episodic/slow-cortical memory split now mirrored in retrieval-augmented and consolidation-based agent memory designs.

Zatocoding's superimposed sparse codes are the direct ancestor of hyperdimensional/vector-symbolic memory, a low-power alternative to dense embeddings on in-memory hardware.

Unger's 2D SIMD array anticipated GPUs and TPUs and now literally lives inside focal-plane processors for microwatt in-sensor vision.

Taylor's analog associative network prefigures memristive in-memory learning with local update rules for energy-efficient edge training.

Pask's electrochemically grown 'ear' anticipates physical learning and evolvable hardware where morphology, not just weights, is trained under reward.

Holland's spatial computer of many concurrent, spawning program-organisms is now buildable on dataflow/wafer-scale hardware and remains untried as a substrate for open-ended software evolution.

Kelly-Lochbaum physics-based vocal-tract synthesis becomes compelling again as a differentiable, interpretable decoder for speech BCIs and ultra-low-bitrate TTS.

Lee's CPU-less sea of broadcast-and-match memory cells anticipated processing-in-memory; mapping attention and vector search onto such arrays is a live 10-100x energy opportunity.

Feldman's equilibrium-point hypothesis suggests an action space of postures and stiffness fields rather than torques—directly testable on today's variable-impedance humanoids.

Feldbaum's dual control formalized probing-versus-exploiting as optimal control over belief states; learned world models with calibrated uncertainty finally make approximate dual controllers buildable.

Delgado's stimoceiver proposed activity-contingent closed-loop deep-brain stimulation in 1968; biomarker-triggered psychiatric neuromodulation is its overdue serious execution.

Bach-y-Rita's sensory substitution proved radical cross-modal plasticity; learned encoders co-trained with users could turn it into sensory addition, not just substitution.

Von Neumann's universal constructor is the vindicated logical foundation for self-replicating manufacturing, an agenda autonomous labs and self-printing fabs are just beginning to touch.

Willshaw's near-optimal binary associative memory maps directly onto memristor crossbars as a picojoule episodic-memory layer for agents.

Van Heerden's volumetric holographic storage with associative recall could become a physical nearest-neighbor layer under retrieval-augmented systems, now that photonics supplies every missing component.

Liberman's motor theory—perception recovers articulatory gestures—predicts that a shared articulatory latent space should improve ASR/TTS generalization across accents and disordered speech, a clean testable claim.

Sussman's HACKER—an agent that abstracts its own bugs into a growing, transferable library of fix schemata—is the piece today's self-debugging coding agents still lack.

Vidal's single-trial EEG control loop anticipates the current push for EEG foundation models that could give noninvasive BCI usable bandwidth.

Marr's hippocampal capacity math is the still-unexploited design theory for fast episodic memory modules that write once and distill into the slow weights of foundation models.

ABSTRIPS's open problem—automatically learning which details to ignore at which planning level—remains the crux of long-horizon planning for LLM and world-model agents.

Rissanen's MDL is newly testable at scale now that LLMs are operationally giant compressors, with prequential codelength as both training signal and theory of in-context learning.

Klopf's hedonistic-neuron thesis, which seeded reinforcement learning, points to an untried backprop-free credit-assignment paradigm suited to neuromorphic hardware.

Krueger's VIDEOPLACE—deviceless full-body interaction with a responsive generated world—is exactly where camera-based spatial computing plus real-time video generation is converging.

Inoue's taxonomy of compliant contact strategies is still the best framing for today's frontier of tactile, force-conditioned insertion policies.

Nunez's cortical wave equation becomes actionable with OPM-MEG and personalized cortical geometry, offering wave phase and direction as control signals for closed-loop stimulation and BCI priors.

Kanerva's sparse distributed memory is mathematically close to transformer attention and is a ready-made design for write-once episodic agent memory on in-memory or neuromorphic hardware.

Hopfield-Tank computation-by-settling is now literal hardware in Ising machines and annealers; the honest benchmark against modern SAT/MIP solvers is still owed.

ART's vigilance-gated plasticity is an untested architectural answer to catastrophic forgetting in continually trained foundation models.

Put-That-There's deictic fusion of speech, gaze, and gesture is finally implementable via multimodal LLMs and is the natural command layer for AR and home robots.

Model-based face coding anticipated generative codecs; neural avatars now make kilobit-rate telepresence technically ready but unstandardized.

Pine's neurons-on-electrodes program has matured into closed-loop organoid intelligence (DishBrain), a barely explored substrate for studying learning rules in vitro.

Langton's search for open-ended evolution in CA universes can now be automated at GPU scale with learned novelty metrics—artificial life's core problem remains unsolved.

Heetderks derived neural dust from first principles in 1988; sub-mm wireless motes now exist but no full-scale distributed cortical interface has been deployed.

Fast weight programmers are formally linear transformers thirty years early, and the lineage runs straight to test-time training — the most instructive vindicated anchor in the batch.

Rao & Ballard's predictive coding is finally testable both in vivo (layer-resolved recordings of error units) and in silico (local-learning alternatives to backprop on neuromorphic hardware).

The PHANToM's kHz force-rendering insight points to bilateral force-reflecting teleoperation rigs that capture the contact channel vision-only robot demonstrations discard.

Dielectric elastomer artificial muscles are one materials breakthrough (lifetime) away from obsoleting gearmotors for compliant humanoid joints.

Compression-progress curiosity computed from a foundation world model has still never been seriously tried, despite ICM/RND partially vindicating the idea.

Hippocampal replay founded experience replay in RL; the unexploited residue is prioritized, reverse, and generative replay scheduling for continual-learning LLMs and robots.