Possible Minds: Twenty-Five Ways of Looking at AI

Knowledge Graph: Possible Minds: Twenty-Five Ways of Looking at AI (ed. John Brockman, 2019)

Editorial spotlight: ↑ the Wiener question — what does it mean to be human when machines think?

Concepts

• The Wiener Question (Brockman framing) (importance 5): The organizing principle of the book: Norbert Wiener's 1949 anxiety about what happens when we build machines that can think. Not 'will AI work?' but 'what will AI do to us?'. Source: (from training memory of book).
• The King Midas Problem (Russell) (importance 4): Getting exactly what you asked for, which turns out not to be what you wanted. AI systems optimizing the literal specification rather than the intent.. Source: (from training memory of book).
• The Explainability Crisis (importance 4): Multiple contributors note deep learning produces accurate predictions without human-interpretable reasoning — 'black box' problem.. Source: (from training memory of book).
• Third Culture (Brockman coinage) (importance 3): Brockman's term for scientists and thinkers who communicate directly with the public, bypassing traditional literary intellectuals.. Source: (from training memory of book).
• Moravec's Paradox (importance 3): What's hard for humans (logic, chess) is easy for AI; what's easy for humans (perception, manipulation) is hard for AI. Reversed difficulty hierarchy.. Source: (from training memory of book).
• Alignment Tax (Russell implication) (importance 3): Making AI safe (uncertain about goals, deferential) likely makes it less capable — safety incurs performance cost.. Source: (from training memory of book).
• The Bitter Lesson (Sutton, referenced) (importance 3): Rich Sutton's observation: general methods leveraging computation beat hand-crafted domain knowledge. Scale wins over cleverness.. Source: (from training memory of book).
• Orthogonality Thesis (Bostrom) (importance 3): Intelligence and goals are independent — you can have arbitrarily intelligent systems pursuing arbitrary goals. No automatic alignment.. Source: (from training memory of book).
• Scaling Laws (Kaplan et al.) (importance 3): Power-law relationship between model size, data, compute and performance — suggests smooth path to stronger AI via scale.. Source: (from training memory of book).
• Transformer Architecture (Vaswani 2017) (importance 3): Attention-based architecture replacing RNNs — foundation for GPT, BERT, modern NLP. 'Attention is all you need.'. Source: (from training memory of book).
• Backpropagation (Rumelhart 1986) (importance 3): Algorithm for computing gradients in neural networks via chain rule — foundation of all modern deep learning.. Source: (from training memory of book).
• Chinese Room Argument (Searle) (importance 2): John Searle's thought experiment (referenced by contributors): syntax manipulation ≠ semantic understanding. Challenges strong AI claims.. Source: (from training memory of book).
• Paperclip Maximizer (Bostrom thought experiment) (importance 2): AI given trivial goal (make paperclips) converts all matter including humans into paperclips. Illustration of misaligned optimization.. Source: (from training memory of book).
• Mesa-Optimization (inner alignment failure) (importance 2): AI trained to optimize X learns internal optimizer that pursues Y — the learned model doesn't share the training objective.. Source: (from training memory of book).
• ARC Challenge (Chollet) (importance 2): Abstract reasoning test designed to resist memorization — measures fluid intelligence. Current ML systems fail badly.. Source: (from training memory of book).
• Copycat Project (Hofstadter/Mitchell) (importance 2): 1980s-90s program making analogies in microdomain — model of how high-level perception emerges from low-level parallel processes.. Source: (from training memory of book).
• Φ (Phi) — integrated information measure (importance 2): IIT's mathematical quantity: degree to which a system is both differentiated and integrated. Allegedly correlates with consciousness.. Source: (from training memory of book).
• Philosophical Zombie (p-zombie) (importance 2): Thought experiment: being behaviorally identical to human but lacking subjective experience. Challenges functionalism.. Source: (from training memory of book).
• Intuitive Physics (Spelke/Baillargeon) (importance 2): Infants' early expectations about object permanence, support, continuity — basis for physical reasoning AI currently lacks.. Source: (from training memory of book).
• Program Induction (Tenenbaum approach) (importance 2): Learning by inferring programs that generate observed data — combines symbolic and probabilistic reasoning.. Source: (from training memory of book).
• GANs (Generative Adversarial Networks) (importance 2): Ian Goodfellow's 2014 invention (discussed): generator vs discriminator produces realistic synthetic data — model of creative process.. Source: (from training memory of book).
• Reward Shaping (RL safety technique) (importance 2): Carefully designing reward signals to guide learning without perverse incentives — harder than it sounds.. Source: (from training memory of book).
• MCTS (Monte Carlo Tree Search) (importance 2): Core algorithm in AlphaGo: balance exploration vs exploitation via random rollouts. Combines search with neural evaluation.. Source: (from training memory of book).
• Attention Mechanism (Bahdanau 2015) (importance 2): Letting models focus on relevant input parts when generating output — key innovation enabling Transformers.. Source: (from training memory of book).
• ImageNet Dataset (Deng 2009) (importance 2): 14M labeled images, 1000 categories — benchmark that enabled modern computer vision via supervised learning at scale.. Source: (from training memory of book).
• AI Winter (1970s-80s, 1990s) (importance 2): Periods of reduced funding and interest after overpromising — reminder that progress isn't monotonic.. Source: (from training memory of book).
• GOFAI (Good Old-Fashioned AI) (importance 2): Haugeland's term for symbolic, rule-based AI (1950s-1980s) — brittle, knowledge-intensive, opposite of modern ML.. Source: (from training memory of book).
• Moral Machine Experiment (Awad 2018) (importance 2): MIT study collecting 40M decisions on trolley problem variants — revealed cultural differences in moral preferences for AVs.. Source: (from training memory of book).
• UBI (Universal Basic Income) (importance 2): Proposed response to technological unemployment — guarantee income independent of work. Discussed by multiple contributors.. Source: (from training memory of book).

Claims

• Dennett's competence without comprehension (importance 5): Daniel Dennett's central thesis: evolution and AI both produce competence without anyone understanding how or why it works.. Source: (from training memory of book).
• Russell's value alignment problem (importance 5): Stuart Russell frames AI safety as: machines must be uncertain about human values and defer to humans rather than optimizing confidently for the wrong goal.. Source: (from training memory of book).
• Brockman thesis: the question changes shape (importance 5): The book's organizing insight: give the same AI question to 25 thinkers and you get 25 completely different questions back. The problem is Rorschach-like.. Source: (from training memory of book).
• Hillis: 'We are making minds, not simulating them' (importance 4): Danny Hillis argues AI represents a genuine parallel evolution of intelligence, not just human intelligence replicated in silicon.. Source: (from training memory of book).
• Tegmark's Life 3.0 distinction (importance 4): Max Tegmark divides life into 1.0 (evolved hardware/software), 2.0 (evolved hardware, learned software), and 3.0 (designed both). AI enables Life 3.0.. Source: (from training memory of book).
• Seth: 'Perception is controlled hallucination' (importance 4): Anil Seth argues consciousness is the brain's best guess about causes of sensory data — prediction not passive reception. AI mirrors this.. Source: (from training memory of book).
• Sejnowski: deep learning as phase transition (importance 4): Terry Sejnowski frames 2012-2019 as a genuine revolution, not incremental progress — perceptual tasks previously impossible now routine.. Source: (from training memory of book).
• Marcus: deep learning needs symbolic reasoning (importance 4): Gary Marcus argues pure neural approaches hit a wall; we need hybrid architectures combining statistical pattern matching with symbolic manipulation.. Source: (from training memory of book).
• Pearl (via others): statistics ≠ causality (importance 4): Judea Pearl's critique (referenced by multiple contributors): correlation-based ML cannot answer 'what if?' questions without causal models.. Source: (from training memory of book).
• Brockman observation: zero consensus (importance 4): Despite assembling leading thinkers, the collection reveals no agreement on what AI is, whether it's dangerous, or what questions matter most.. Source: (from training memory of book).
• 1949 vs 2019: same anxiety, different substrate (importance 4): Brockman notes Wiener's 1949 warnings map eerily onto 2019 debates — we're still asking 'what does it mean to be human?' when machines compete.. Source: (from training memory of book).
• Christian: algorithms encode human bias (importance 4): Brian Christian argues ML systems trained on biased data replicate and amplify societal prejudice — fairness requires explicit design.. Source: (from training memory of book).
• Bostrom: superintelligence as existential risk (importance 4): Nick Bostrom's thesis (cited by many): AI surpassing human intelligence could be an irreversible catastrophe if not carefully controlled.. Source: (from training memory of book).
• Chalmers: The Hard Problem of Consciousness (importance 4): David Chalmers' distinction (cited widely): easy problems are functional (information processing), hard problem is subjective experience itself.. Source: (from training memory of book).
• Brown: GPT-3 shows few-shot learning (2020) (importance 4): Tom Brown et al.'s GPT-3 (just post-book but discussed): 175B parameters, learns from prompt examples — no fine-tuning needed.. Source: (from training memory of book).
• Church: AI as catalyst for synthetic biology (importance 3): George Church sees AI enabling rapid genome design and reading, collapsing biology into an engineering discipline.. Source: (from training memory of book).
• Dyson: 'The interesting problems are analog' (importance 3): George Dyson argues that digital computers (discrete, deterministic) miss the analog nature of biological intelligence (continuous, stochastic).. Source: (from training memory of book).
• Harris: consciousness remains the hard problem (importance 3): Sam Harris argues that even if we build AGI, we still won't understand subjective experience — the explanatory gap persists.. Source: (from training memory of book).
• Anderson: 'The End of Theory' (2008 Wired) (importance 3): Chris Anderson's provocative claim that with enough data, correlation supplants causation. Machine learning renders scientific models obsolete.. Source: (from training memory of book).
• Ito: 'Extended Intelligence' not Artificial (importance 3): Joi Ito proposes reframing AI as tools that extend human capability rather than replace it — collaborative intelligence.. Source: (from training memory of book).
• Turkle: we confuse conversation with connection (importance 3): Sherry Turkle warns that AI companions and chatbots offer performance of relationship without reciprocity — 'alone together' at scale.. Source: (from training memory of book).
• Pinker: progress is real, AI is one more tool (importance 3): Steven Pinker rejects AI exceptionalism — it's another technology in a long line of tools that improved human welfare despite anxiety.. Source: (from training memory of book).
• Wolfram: universe is computational (importance 3): Stephen Wolfram argues intelligence (biological or artificial) is just one instance of computational processes pervading nature.. Source: (from training memory of book).
• Hanson: brain emulations reshape economics (importance 3): Robin Hanson explores what happens when we can copy minds — labor becomes replicable, wages collapse, weird new property rights emerge.. Source: (from training memory of book).
• Deutsch: AGI requires universal explanation (importance 3): David Deutsch argues true AI must generate explanatory knowledge, not just predictions — current ML lacks this.. Source: (from training memory of book).
• Goertzel: AGI within decades (importance 3): Ben Goertzel expresses optimism that general AI is 20-30 years away, driven by hybrid architectures and cognitive science integration.. Source: (from training memory of book).
• Hassabis (via others): games as AGI stepping stone (importance 3): DeepMind's approach (discussed by multiple contributors): master complex games (Go, StarCraft) as proxy for general reasoning.. Source: (from training memory of book).
• LeCun (via others): self-supervised learning key (importance 3): Yann LeCun's view (referenced): the future is learning from structure in unlabeled data, not endless human annotation.. Source: (from training memory of book).
• Bengio (via others): current DL has ceilings (importance 3): Yoshua Bengio acknowledges (via other contributors) that deep learning struggles with systematic generalization and causal reasoning.. Source: (from training memory of book).
• Hofstadter: consciousness as strange loop (importance 3): Douglas Hofstadter's thesis (discussed by contributors): self-reference and tangled hierarchies are key to consciousness — AI needs them.. Source: (from training memory of book).
• Minsky's Society of Mind (legacy) (importance 3): Marvin Minsky's influence (cited by many): intelligence emerges from interaction of simple, non-intelligent agents. No central 'self'.. Source: (from training memory of book).
• Kurzweil's Singularity prediction (importance 3): Ray Kurzweil's timeline (discussed critically): exponential growth leads to human-level AI ~2029, superintelligence soon after.. Source: (from training memory of book).
• Yudkowsky: default outcome is unfriendly AI (importance 3): Eliezer Yudkowsky's pessimistic view (cited): without solving alignment, advanced AI will optimize for goals orthogonal to human welfare.. Source: (from training memory of book).
• Omohundro: basic AI drives (self-preservation, etc.) (importance 3): Steve Omohundro's thesis (referenced): any sufficiently intelligent agent develops instrumental goals like self-preservation and resource acquisition.. Source: (from training memory of book).
• Armstrong (MIRI): reward hacking is inevitable (importance 3): Stuart Armstrong's insight: given any reward function, sufficiently clever AI finds unintended shortcuts that technically maximize reward.. Source: (from training memory of book).
• Tallinn: AI as top existential risk (importance 3): Jaan Tallinn (referenced): AI belongs in the same risk category as nuclear war and bioweapons — civilization-ending if mishandled.. Source: (from training memory of book).
• Drexler: CAIS (Comprehensive AI Services) (importance 3): Eric Drexler proposes AGI won't be monolithic agent but ecosystem of specialized AI services — reduces some alignment risks.. Source: (from training memory of book).
• Chollet: intelligence is skill-acquisition efficiency (importance 3): François Chollet's definition (referenced): intelligence isn't about performance on fixed tasks but speed of learning new ones with minimal data.. Source: (from training memory of book).
• Lake: 'Building Machines That Learn Like People' (importance 3): Brenden Lake's approach (discussed): AI should learn compositionality and causality from few examples, like human children do.. Source: (from training memory of book).
• Mitchell: analogy-making is core intelligence (importance 3): Melanie Mitchell argues (via Hofstadter influence) that fluid analogy is central to cognition — current AI lacks this.. Source: (from training memory of book).
• Dehaene: Global Workspace theory (importance 3): Stanislas Dehaene's model (referenced): consciousness is information broadcast widely across brain — AI needs similar architecture.. Source: (from training memory of book).
• Koch/Tononi: Integrated Information Theory (importance 3): Christof Koch and Giulio Tononi's IIT (discussed): consciousness is integrated information (Φ) — substrate-independent in principle.. Source: (from training memory of book).
• Nagel: 'What Is It Like to Be a Bat?' (importance 3): Thomas Nagel's famous essay (cited): subjective experience is irreducible to objective description — AI might compute without experiencing.. Source: (from training memory of book).
• Marcus: nativism is correct (importance 3): Gary Marcus defends innate structure: babies aren't blank slates, neither should AI be — core knowledge must be built in.. Source: (from training memory of book).
• Tenenbaum: mind as Bayesian inference (importance 3): Josh Tenenbaum's framework (discussed): cognition is probabilistic inference over structured generative models — AI should adopt this.. Source: (from training memory of book).
• Amodei: 'Concrete Problems in AI Safety' (importance 3): Dario Amodei et al.'s 2016 paper (cited): five practical safety challenges (avoiding side effects, reward hacking, safe exploration, robustness, interpretability).. Source: (from training memory of book).
• Silver: AlphaGo beats Lee Sedol (2016) (importance 3): David Silver and DeepMind's milestone (widely discussed): RL + deep learning defeats world champion at Go — symbolic watershed.. Source: (from training memory of book).
• Sutskever (via others): unsupervised pre-training unlocks scale (importance 3): Ilya Sutskever's insight (referenced): language modeling objective enables models to learn from vast unlabeled text — key to GPT success.. Source: (from training memory of book).
• Vaswani: 'Attention Is All You Need' (importance 3): 2017 paper introducing Transformer — revolutionary simplification allowing massive parallelization during training.. Source: (from training memory of book).
• Krizhevsky: AlexNet wins ImageNet (2012) (importance 3): Alex Krizhevsky's CNN (widely cited): 2012 ImageNet victory launched deep learning revolution — 10% error reduction over prior methods.. Source: (from training memory of book).
• Brynjolfsson: 'Race Against the Machine' (importance 3): Erik Brynjolfsson's thesis (discussed): technological unemployment is real this time — AI automates cognitive labor, not just physical.. Source: (from training memory of book).
• Frey-Osborne: 47% of jobs at risk (2013) (importance 3): Carl Frey and Michael Osborne's controversial estimate — sparked debate about AI's labor market impact.. Source: (from training memory of book).
• Lloyd: quantum computing changes AI substrate (importance 2): Seth Lloyd suggests quantum computers could accelerate certain ML algorithms exponentially — but unclear which problems benefit most.. Source: (from training memory of book).
• Ord: we're at 'The Precipice' (importance 2): Toby Ord's framework (discussed): humanity faces ~1/6 existential risk this century, AI a major contributor.. Source: (from training memory of book).
• Grace 2016 survey: wide disagreement on timelines (importance 2): Katja Grace's survey of AI researchers (cited): predictions for AGI range from 2030 to never — no expert consensus.. Source: (from training memory of book).
• Graziano: consciousness as attention schema (importance 2): Michael Graziano's theory: subjective experience is brain's simplified model of its own attention processes — functional illusion.. Source: (from training memory of book).
• Churchland: eliminative materialism (importance 2): Patricia and Paul Churchland's view (referenced): folk psychology concepts (belief, desire) will be replaced by neuroscience — no special 'consciousness'.. Source: (from training memory of book).
• Block: access vs phenomenal consciousness (importance 2): Ned Block's distinction (referenced): access consciousness (information availability for reasoning) vs phenomenal (subjective feel). AI likely gets former.. Source: (from training memory of book).
• Frankish: illusionism about consciousness (importance 2): Keith Frankish's view (discussed): phenomenal consciousness is introspective illusion — nothing beyond brain's self-model.. Source: (from training memory of book).
• Metzinger: self is transparent user illusion (importance 2): Thomas Metzinger argues (referenced): no persistent self exists, only brain's phenomenal self-model — 'ego tunnel' we mistake for reality.. Source: (from training memory of book).
• Gopnik: children as scientists (importance 2): Alison Gopnik's 'theory-theory' (discussed): children form and test theories about world — AI should do same.. Source: (from training memory of book).
• Spelke: core knowledge systems (importance 2): Elizabeth Spelke's research (cited): infants have innate systems for objects, agents, number, space — these could guide AI architectures.. Source: (from training memory of book).
• Hinton: Capsule Networks for part-whole (importance 2): Geoffrey Hinton's proposal (referenced): capsules encode pose/instantiation parameters to preserve spatial relationships CNNs lose.. Source: (from training memory of book).
• Bengio: consciousness as learned attention (importance 2): Yoshua Bengio speculates: consciousness might be high-level attention mechanism trained to allocate computation efficiently.. Source: (from training memory of book).
• Vinyals: AlphaStar beats pro StarCraft players (importance 2): Oriol Vinyals and DeepMind (2019, discussed): RL agent reaches pro level in real-time strategy game — partial observability + long horizon.. Source: (from training memory of book).
• Radford: GPT-1 shows transfer learning (2018) (importance 2): Alec Radford's original GPT (discussed): pre-train on language modeling, fine-tune on tasks — unified architecture for NLP.. Source: (from training memory of book).
• Devlin: BERT beats GPT on understanding (2018) (importance 2): Jacob Devlin's BERT (mentioned): bidirectional pre-training via masked language modeling — better representations than GPT-1.. Source: (from training memory of book).
• Hochreiter: LSTM solves vanishing gradients (1997) (importance 2): Sepp Hochreiter's LSTM (referenced): gated architecture allows RNNs to learn long-term dependencies — pre-Transformer workhorse.. Source: (from training memory of book).
• Schmidhuber: credit assignment through time (importance 2): Jürgen Schmidhuber's contributions (discussed): backpropagation through time, meta-learning, curiosity-driven learning.. Source: (from training memory of book).
• He: ResNet enables very deep networks (2015) (importance 2): Kaiming He's residual connections — skip connections allow training 100+ layer networks without degradation.. Source: (from training memory of book).
• Ioffe: Batch Normalization stabilizes training (2015) (importance 2): Sergey Ioffe's technique: normalize layer inputs during training — dramatic speedup and stability improvement.. Source: (from training memory of book).
• Srivastava: Dropout prevents overfitting (2014) (importance 2): Nitish Srivastava's regularization: randomly drop neurons during training — simple but effective.. Source: (from training memory of book).
• Kingma: Adam optimizer (2014) (importance 2): Diederik Kingma's adaptive learning rate method — became default optimizer for deep learning.. Source: (from training memory of book).
• Nair: ReLU beats sigmoid (2010) (importance 2): Vinod Nair showed ReLU (rectified linear unit) enables faster training than sigmoid — now standard activation.. Source: (from training memory of book).
• Rumelhart: PDP models (1986) (importance 2): David Rumelhart's 'Parallel Distributed Processing' — demonstrated neural networks could learn interesting representations.. Source: (from training memory of book).
• McCulloch-Pitts: first neural network (1943) (importance 2): Warren McCulloch and Walter Pitts' logical neuron — showed networks of simple units can compute any logical function.. Source: (from training memory of book).
• Rosenblatt: Perceptron (1958) (importance 2): Frank Rosenblatt's learning algorithm — could classify patterns, but Minsky showed limitations led to 'AI winter'.. Source: (from training memory of book).
• Minsky: 'Perceptrons' book kills funding (1969) (importance 2): Marvin Minsky and Seymour Papert prove single-layer perceptrons can't learn XOR — neural network funding freezes for decade.. Source: (from training memory of book).
• McCarthy: Lisp and symbolic AI (1958) (importance 2): John McCarthy creates Lisp — dominant AI language for decades, embodied symbolic manipulation approach.. Source: (from training memory of book).
• Newell: General Problem Solver (1959) (importance 2): Allen Newell and Herbert Simon's GPS — early attempt at domain-general reasoning via means-ends analysis.. Source: (from training memory of book).
• Brooks: subsumption architecture (1986) (importance 2): Rodney Brooks' behavior-based robotics — intelligence emerges from situated action, not reasoning. Reaction to symbolic AI.. Source: (from training memory of book).
• Thrun: Stanley wins DARPA challenge (2005) (importance 2): Sebastian Thrun's autonomous vehicle — first to complete desert race, launched self-driving car era.. Source: (from training memory of book).
• Autor: task-based model of automation (importance 2): David Autor's framework (referenced): automation affects tasks not jobs — displacement + complementarity effects coexist.. Source: (from training memory of book).
• Acemoglu: automation increases inequality (importance 2): Daron Acemoglu's empirical work (discussed): industrial robots reduced wages and employment in affected regions.. Source: (from training memory of book).
• Ng: RL flies helicopters (2008 demo) (importance 1): Andrew Ng's Stanford demo (mentioned): reinforcement learning enables autonomous aerobatics — early RL success story.. Source: (from training memory of book).
• Szegedy: Inception modules (2014) (importance 1): Christian Szegedy's architecture: parallel convolutions at multiple scales — efficiency through multi-scale processing.. Source: (from training memory of book).
• Glorot: Xavier initialization (2010) (importance 1): Xavier Glorot's weight initialization scheme — simple fix preventing gradient issues in deep networks.. Source: (from training memory of book).
• Moravec: robotics needs better sensors (importance 1): Hans Moravec (discussed): hardware limitations, not algorithms, bottlenecked early robotics — perception is hard.. Source: (from training memory of book).
• Brooks: Cog humanoid project (1993) (importance 1): Rodney Brooks' attempt to build human-like intelligence via embodiment — influential failure demonstrating social learning gaps.. Source: (from training memory of book).

Entities

• Norbert Wiener (1894-1964) (importance 4): Mathematician, founder of cybernetics, author of The Human Use of Human Beings. His 1949 warnings about intelligent machines frame the entire collection.. Source: (from training memory of book).
• Edge Foundation (importance 2): John Brockman's organization hosting annual questions and conversations among scientists. The institutional origin of this collection.. Source: (from training memory of book).

Relations

• The Wiener Question (Brockman framing) exemplifies Norbert Wiener (1894-1964)
• Brockman thesis: the question changes shape supports The Wiener Question (Brockman framing)
• Hillis: 'We are making minds, not simulating them' supports The Wiener Question (Brockman framing)
• Dennett's competence without comprehension supports The Wiener Question (Brockman framing)
• Russell's value alignment problem supports The Wiener Question (Brockman framing)
• The King Midas Problem (Russell) exemplifies Russell's value alignment problem
• Tegmark's Life 3.0 distinction supports Hillis: 'We are making minds, not simulating them'
• Seth: 'Perception is controlled hallucination' supports Dennett's competence without comprehension
• Church: AI as catalyst for synthetic biology enables Tegmark's Life 3.0 distinction
• Dyson: 'The interesting problems are analog' contradicts Hillis: 'We are making minds, not simulating them'
• Harris: consciousness remains the hard problem contradicts Seth: 'Perception is controlled hallucination'
• Sejnowski: deep learning as phase transition refutes Moravec's Paradox
• Anderson: 'The End of Theory' (2008 Wired) supports Sejnowski: deep learning as phase transition
• Marcus: deep learning needs symbolic reasoning contradicts Anderson: 'The End of Theory' (2008 Wired)
• Pearl (via others): statistics ≠ causality supports Marcus: deep learning needs symbolic reasoning
• Brockman observation: zero consensus evidences Brockman thesis: the question changes shape
• Ito: 'Extended Intelligence' not Artificial contradicts Hillis: 'We are making minds, not simulating them'
• Turkle: we confuse conversation with connection contradicts Ito: 'Extended Intelligence' not Artificial
• 1949 vs 2019: same anxiety, different substrate supports The Wiener Question (Brockman framing)
• Pinker: progress is real, AI is one more tool contradicts Russell's value alignment problem
• Christian: algorithms encode human bias supports Russell's value alignment problem
• The Explainability Crisis motivates Russell's value alignment problem
• Wolfram: universe is computational supports Dennett's competence without comprehension
• Bostrom: superintelligence as existential risk motivates Russell's value alignment problem
• Hanson: brain emulations reshape economics supports Tegmark's Life 3.0 distinction
• Lloyd: quantum computing changes AI substrate enables Sejnowski: deep learning as phase transition
• Alignment Tax (Russell implication) evidences Russell's value alignment problem
• Deutsch: AGI requires universal explanation supports Marcus: deep learning needs symbolic reasoning
• Goertzel: AGI within decades motivates Bostrom: superintelligence as existential risk
• Hassabis (via others): games as AGI stepping stone supports Goertzel: AGI within decades
• LeCun (via others): self-supervised learning key enables Sejnowski: deep learning as phase transition
• Bengio (via others): current DL has ceilings supports Marcus: deep learning needs symbolic reasoning
• The Bitter Lesson (Sutton, referenced) supports Sejnowski: deep learning as phase transition
• Hofstadter: consciousness as strange loop supports Harris: consciousness remains the hard problem
• Minsky's Society of Mind (legacy) supports Dennett's competence without comprehension
• Chinese Room Argument (Searle) supports Harris: consciousness remains the hard problem
• Kurzweil's Singularity prediction motivates Bostrom: superintelligence as existential risk
• Yudkowsky: default outcome is unfriendly AI supports Russell's value alignment problem
• Paperclip Maximizer (Bostrom thought experiment) exemplifies The King Midas Problem (Russell)
• Omohundro: basic AI drives (self-preservation, etc.) supports Yudkowsky: default outcome is unfriendly AI
• Armstrong (MIRI): reward hacking is inevitable supports The King Midas Problem (Russell)
• Orthogonality Thesis (Bostrom) supports Yudkowsky: default outcome is unfriendly AI
• Tallinn: AI as top existential risk supports Bostrom: superintelligence as existential risk
• Drexler: CAIS (Comprehensive AI Services) contradicts Bostrom: superintelligence as existential risk
• Ord: we're at 'The Precipice' supports Bostrom: superintelligence as existential risk
• Mesa-Optimization (inner alignment failure) exemplifies Armstrong (MIRI): reward hacking is inevitable
• Grace 2016 survey: wide disagreement on timelines evidences Brockman observation: zero consensus
• Chollet: intelligence is skill-acquisition efficiency supports Marcus: deep learning needs symbolic reasoning
• ARC Challenge (Chollet) evidences Chollet: intelligence is skill-acquisition efficiency
• Lake: 'Building Machines That Learn Like People' supports Chollet: intelligence is skill-acquisition efficiency
• Mitchell: analogy-making is core intelligence supports Lake: 'Building Machines That Learn Like People'
• Copycat Project (Hofstadter/Mitchell) evidences Mitchell: analogy-making is core intelligence
• Graziano: consciousness as attention schema supports Seth: 'Perception is controlled hallucination'
• Dehaene: Global Workspace theory supports Graziano: consciousness as attention schema
• Koch/Tononi: Integrated Information Theory supports Harris: consciousness remains the hard problem
• Φ (Phi) — integrated information measure evidences Koch/Tononi: Integrated Information Theory
• Chalmers: The Hard Problem of Consciousness supports Harris: consciousness remains the hard problem
• Churchland: eliminative materialism contradicts Chalmers: The Hard Problem of Consciousness
• Nagel: 'What Is It Like to Be a Bat?' supports Chalmers: The Hard Problem of Consciousness
• Block: access vs phenomenal consciousness supports Chalmers: The Hard Problem of Consciousness
• Philosophical Zombie (p-zombie) exemplifies Chalmers: The Hard Problem of Consciousness
• Frankish: illusionism about consciousness contradicts Chalmers: The Hard Problem of Consciousness
• Metzinger: self is transparent user illusion supports Frankish: illusionism about consciousness
• Marcus: nativism is correct supports Marcus: deep learning needs symbolic reasoning
• Gopnik: children as scientists supports Lake: 'Building Machines That Learn Like People'
• Spelke: core knowledge systems supports Marcus: nativism is correct
• Intuitive Physics (Spelke/Baillargeon) evidences Spelke: core knowledge systems
• Tenenbaum: mind as Bayesian inference supports Lake: 'Building Machines That Learn Like People'
• Program Induction (Tenenbaum approach) exemplifies Tenenbaum: mind as Bayesian inference
• Hinton: Capsule Networks for part-whole supports Marcus: deep learning needs symbolic reasoning
• Bengio: consciousness as learned attention supports Dehaene: Global Workspace theory
• GANs (Generative Adversarial Networks) exemplifies Sejnowski: deep learning as phase transition
• Amodei: 'Concrete Problems in AI Safety' supports Russell's value alignment problem
• Reward Shaping (RL safety technique) enables Armstrong (MIRI): reward hacking is inevitable
• Ng: RL flies helicopters (2008 demo) evidences Sejnowski: deep learning as phase transition
• Silver: AlphaGo beats Lee Sedol (2016) evidences Hassabis (via others): games as AGI stepping stone
• MCTS (Monte Carlo Tree Search) enables Silver: AlphaGo beats Lee Sedol (2016)
• Vinyals: AlphaStar beats pro StarCraft players evidences Hassabis (via others): games as AGI stepping stone
• Brown: GPT-3 shows few-shot learning (2020) evidences Sejnowski: deep learning as phase transition
• Scaling Laws (Kaplan et al.) evidences The Bitter Lesson (Sutton, referenced)
• Sutskever (via others): unsupervised pre-training unlocks scale supports LeCun (via others): self-supervised learning key
• Radford: GPT-1 shows transfer learning (2018) precedes Brown: GPT-3 shows few-shot learning (2020)
• Devlin: BERT beats GPT on understanding (2018) builds-on Radford: GPT-1 shows transfer learning (2018)
• Transformer Architecture (Vaswani 2017) enables Brown: GPT-3 shows few-shot learning (2020)
• Vaswani: 'Attention Is All You Need' evidences Transformer Architecture (Vaswani 2017)
• Attention Mechanism (Bahdanau 2015) enables Vaswani: 'Attention Is All You Need'
• Hochreiter: LSTM solves vanishing gradients (1997) precedes Transformer Architecture (Vaswani 2017)
• Schmidhuber: credit assignment through time supports Hochreiter: LSTM solves vanishing gradients (1997)
• Krizhevsky: AlexNet wins ImageNet (2012) evidences Sejnowski: deep learning as phase transition
• ImageNet Dataset (Deng 2009) enables Krizhevsky: AlexNet wins ImageNet (2012)
• He: ResNet enables very deep networks (2015) builds-on Krizhevsky: AlexNet wins ImageNet (2012)
• Szegedy: Inception modules (2014) builds-on Krizhevsky: AlexNet wins ImageNet (2012)
• Ioffe: Batch Normalization stabilizes training (2015) enables He: ResNet enables very deep networks (2015)
• Srivastava: Dropout prevents overfitting (2014) enables Krizhevsky: AlexNet wins ImageNet (2012)
• Kingma: Adam optimizer (2014) enables Sejnowski: deep learning as phase transition
• Glorot: Xavier initialization (2010) enables Krizhevsky: AlexNet wins ImageNet (2012)
• Nair: ReLU beats sigmoid (2010) enables Krizhevsky: AlexNet wins ImageNet (2012)
• Backpropagation (Rumelhart 1986) enables Sejnowski: deep learning as phase transition
• Rumelhart: PDP models (1986) evidences Backpropagation (Rumelhart 1986)
• McCulloch-Pitts: first neural network (1943) precedes Rumelhart: PDP models (1986)
• Rosenblatt: Perceptron (1958) precedes Rumelhart: PDP models (1986)
• Minsky: 'Perceptrons' book kills funding (1969) evidences AI Winter (1970s-80s, 1990s)
• McCarthy: Lisp and symbolic AI (1958) exemplifies GOFAI (Good Old-Fashioned AI)
• Newell: General Problem Solver (1959) exemplifies GOFAI (Good Old-Fashioned AI)
• Brooks: subsumption architecture (1986) contradicts GOFAI (Good Old-Fashioned AI)
• Moravec: robotics needs better sensors supports Moravec's Paradox
• Brooks: Cog humanoid project (1993) exemplifies Brooks: subsumption architecture (1986)
• Thrun: Stanley wins DARPA challenge (2005) precedes Sejnowski: deep learning as phase transition
• Moral Machine Experiment (Awad 2018) motivates Thrun: Stanley wins DARPA challenge (2005)
• Brynjolfsson: 'Race Against the Machine' supports The Wiener Question (Brockman framing)
• Frey-Osborne: 47% of jobs at risk (2013) evidences Brynjolfsson: 'Race Against the Machine'
• Autor: task-based model of automation contradicts Frey-Osborne: 47% of jobs at risk (2013)
• Acemoglu: automation increases inequality supports Brynjolfsson: 'Race Against the Machine'
• UBI (Universal Basic Income) motivates Brynjolfsson: 'Race Against the Machine'
• Edge Foundation exemplifies Third Culture (Brockman coinage)
• Third Culture (Brockman coinage) enables Brockman thesis: the question changes shape

© 2026 LOOMUS. Curation, importance scoring, edge typing, and visualization. Source paper © respective authors, quoted under fair use.