The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence

Definition and motivation

• Robust AI should synthesize knowledge from varied sources, reason flexibly about the world, and transfer learning across contexts in a systematic, reliable way.This reliability is necessary for AI that people can trust in homes, transportation, healthcare, businesses, and communities, though reliability alone does not guarantee trustworthiness.Source: §1 · p. 3

Limits of current AI

• Current deep-learning systems remain brittle, data-hungry, and weak at generalizing beyond training distributions, often exploiting spurious patterns instead of learning flexible meaning.They can perform extremely well on specific tasks yet fail when situations deviate slightly from training data or when novel tasks require repurposing.Source: §1 · p. 5
• GPT-2 illustrates these limits: it can produce fluent continuations, yet often predicts the correct response category while failing to track concrete details or simple quantities.Examples include placing dry-cleaned clothes at the wrong location and reporting seventeen frogs after six frogs lose two and gain three.Source: §1 · pp. 5–6

Proposed direction

• The paper proposes refocusing AI on systems that acquire, represent, and manipulate abstract knowledge while building, updating, and reasoning over complex internal models of the world.This approach revisits classical AI’s concerns with knowledge, internal models, and reasoning using modern techniques.Source: §1 · p. 6

Open-ended reasoning

• Knowledge-free general methods may work in closed-ended domains, but open-ended commonsense requires internal world models that update over time, such as tracking where objects have been left.The paper contrasts game success with unexpected real-world planning and conversational understanding, where computational scaling alone remains pointillistic and unreliable.Source: §1 · p. 8

• Robust AI should perceive the world, construct and update structured internal cognitive models, and use them to reason and make decisions.Cognitive models may be incomplete or inaccurate, yet they guide an organism’s understanding and action; analogous cycles appear in language processing and much robotics.Source: §2 · pp. 8–9
• The paper’s central claim is that AI lacking detailed world models, substantial knowledge, and reasoning will remain unreliable when real-world circumstances deviate from training data.Such systems may get some things right through correlations but will not robustly understand or handle distribution shifts.Source: §2 · p. 9
• Adding hidden layers or more nodes does not resolve this out-of-distribution generalization problem.The limitation extends beyond the toy identity example to abstract-pattern failures in natural-language systems and remains a central obstacle to progress (Bengio, 2019).Source: §2 · pp. 11–12
• Multilayer perceptrons can learn training examples and interpolate within the training distribution, but they fail to recover broad rules such as identity outside it.A network trained on f(x)=x for even binary numbers may predict f(1111)=1110 rather than extending the identity relationship to odd cases.Source: §2 · pp. 11–12
• These generalization failures make systems difficult to trust in open-ended domains, where performance can collapse just beyond familiar examples.One math-learning system handled sums through 1+1+1+1+1+1=6 but failed at 1+1+1+1+1+1+1=7 and larger cases, while symbolic inductive program synthesis can be more effective in some such problems.Source: §2 · p. 12

• Neural networks’ sensitivity to item order, catastrophic interference, and task-environment particulars makes their generalization depend critically on training details.Catastrophic interference can overwrite earlier associations with later ones, and proposed solutions have not resolved the issue.Source: §2 · p. 13
• Their inability to extrapolate beyond training distributions conflicts with commonsense knowledge, which supports abstract causal generalizations across open-ended classes of entities.For example, people generalize that breaking a bottle containing liquid will probably cause some liquid to escape, across variations in bottle properties and contents.Source: §2 · p. 13
• Backpropagation-trained multilayer perceptrons alone are therefore not the right tool for representing and manipulating such abstract knowledge, requiring alternative learning mechanisms.The challenge is to learn, represent, and extend knowledge that applies beyond specific entities to whole classes of things.Source: §2 · p. 13

2.1.1. Symbolic operations over variables offer the only known solution, but the solution is partial

• Robust, knowledge-driven AI requires symbol-manipulation because it is the only known machinery that reliably represents and manipulates abstract knowledge, while it must be combined with learning mechanisms.Symbolic operations provide abstraction and reliability independently of experience, whereas deep learning is stronger at learning from large datasets.Source: §2.1.1 · pp. 14, 17

2.1.2 Hybrids are often effective

• Hybrid systems have shown practical effectiveness across search, information retrieval, visual question answering, planning, mathematics, and reasoning by combining symbolic structures with neural learning.Google Search combines Knowledge Graph tools with BERT and RankBrain; AlphaGo combines Monte Carlo Tree Search with deep-learning position evaluation, while deep learning alone produces weaker play.Source: §2.1.2 · pp. 18–19

2.1.3. Common objections to hybrid models and symbol-manipulation

• Objections that symbols are biologically implausible, ineffective, or unscalable lack compelling support because psychological evidence, active research, large systems, and differentiable hybrids provide counterexamples.The text notes evidence for abstract generalization in infants, adults, and bees, while Google Knowledge Graph and recent neurosymbolic systems challenge scalability concerns.Source: §2.1.3 · pp. 20–21

2.1.4. Determining whether a given system is a hybrid system is not always trivial

• Determining whether a system is hybrid is not always straightforward, since symbolic components may be embedded without being acknowledged, as in OpenAI’s Rubik’s cube solver.The section also allows that deep-learning variants might achieve symbolic-like computation through a different substrate, with human-like limitations while supporting learning and uncertainty.Source: §2.1.4 · pp. 21–22
• Determining whether a neural system is hybrid requires discovering mappings between its architecture and symbol-manipulating operations, rather than relying on its outward description.Synthetic systems can be engineered through search, but understanding what makes them work remains a reverse-engineering task involving discovery and rejection of possible mappings.Source: §2.1.4 · pp. 22–23
• A successful neural system could still require a separate scientific explanation to establish whether its components implement symbol-manipulating machinery.Complex neural systems, like the brain, may work effectively while remaining opaque, so performance alone does not reveal the underlying computations.Source: §2.1.4 · p. 23

2.1.5. Summary

• Hybrid models combine large-scale data learning with abstract syntactic and semantic representations, but may be necessary rather than sufficient for robust intelligence.Symbol-manipulation also supports extrapolation beyond training regimes and structured representations, while lacking a satisfactory learning framework.Source: §2.1.5 · pp. 23–24
• Greater study of hybrid architectures is warranted because relatively few resources have examined them compared with pure deep learning, despite indications of their value.A principled account of how symbolic and neural components work together could also reveal computational efficiencies and expose hidden symbolic dependencies.Source: §2.1.5 · pp. 23–24

2.2. Large-scale knowledge, some of which is abstract and causal

• Transformers such as GPT-2 appear to absorb and integrate commonsense knowledge without explicit knowledge engineering, but primarily model word usage rather than underlying concepts.Their co-occurrence-based approximations can produce fluent behavior while lacking reliable conceptual understanding, relevance, plausibility, and causal structure.Source: §2.2 · pp. 24–26
• This shallow knowledge undermines downstream inference, planning, language understanding, interpretability, and explainability because unreliable concepts yield unreliable decisions and inconsistent explanations.Systems may treat plausible and implausible fire-starting methods alike and lack the cognitive context needed to disambiguate language.Source: §2.2 · pp. 26–27
• Systematic ways to induce, represent, and manipulate large databases of structured, often causal knowledge are presented as prerequisites for robust intelligence.Statistical approximation alone produces shallow inferences and cannot reliably support robustness in complex real-world environments.Source: §2.2 · p. 27

2.2.1 What kind of knowledge will robust artificial intelligence require?

• Robust AI knowledge should be largely learned but include external cultural information, abstract relationships, rules with exceptions, causal and counterfactual structure, broad domain coverage, and a small innate component.Web scraping readily captures explicit facts but misses unstated commonsense, while integrating relevant abstract knowledge into real-world situations remains essentially unsolved.Source: §2.2.1 · pp. 27–28

2.2.2 Case Study: Containers

• The container case study proposes deriving unfamiliar-object behavior from general logical axioms about space, time, objects, liquids, and containers rather than memorizing every specific fact (Davis, Marcus, & Frazier-Logue, 2017).Such axioms could support inference about unusual kettles, yarn feeders, and other objects, although current neural networks do not naturally represent the required object-centered knowledge.Source: §2.2.2 · pp. 30–31
• The container framework illustrates the need for architectures that can acquire, represent, and manipulate symbolic knowledge, potentially combining innate and learned axioms through neurosymbolic systems.The relevant frameworks cluster around domains such as space, time, and causality, but no existing system is identified as capable of learning axioms of this kind.Source: §2.2.2 · p. 31

2.2.3 Innate Frameworks for Knowledge

• Innate knowledge can make learning easier by narrowing the hypothesis space, as demonstrated by convolutional structure improving digit-recognition accuracy.The section argues that restricting priors to architectural details such as convolution while excluding broader innate constraints is an arbitrary machine-learning convention.Source: §2.2.3 · p. 32
• AI should optimize for priors that best facilitate learning rather than minimizing their storage, since modern systems have ample memory and need richer conceptual commitments.Relevant priors include object persistence and temporal calculi, which current deep-learning approaches have largely neglected.Source: §2.2.3 · pp. 33–35
• Innate frameworks for time, space, and causality could provide the conceptual basis for learning, with objects, sets, places, and temporal structure supporting the acquisition of further knowledge.The paper identifies time, space, and causality as recurring proposals for innateness and notes related developmental-psychology arguments.Source: §2.2.3 · pp. 33–34
• Near-blank-slate systems trained on massive datasets have not reliably reasoned about time, space, or causality, limiting their robustness in variable real-world settings.This motivates a more nativist approach to AI.Source: §2.2.3 · p. 34
• Although humans learn millions or tens of millions of facts, a significant fraction of that knowledge is abstract and causal, likely requiring hybrid architectures rather than purely neural representations.Learned knowledge can guide action and decision-making, while causal and abstract content may require neurosymbolic or other hybrid mechanisms.Source: §2.2.3 · pp. 35–36
• Pure prewiring cannot suffice because the world changes and systems must learn new causal principles, as children do when encountering novel technologies.A system should therefore acquire new causal knowledge rather than depend exclusively on fixed innate representations.Source: §2.2.3 · pp. 35–36
• Learning abstract and causal knowledge entirely from experience is inefficient and unreliable; core physical and psychological knowledge is needed to direct subsequent common-sense learning.The paper uses object persistence under occlusion as an example of knowledge that should not be relearned from scratch.Source: §2.2.3 · pp. 35–36
• The third major claim is to initialize AI systems with domain frameworks for time, space, and causality, thereby speeding learning and massively constraining the hypothesis space.These frameworks might be represented in formal logic or other mechanisms, but no amount of innateness can replace learning.Source: §2.2.3 · p. 36
• Knowledge alone is insufficient: robust systems must put it into practice through reasoning tools and cognitive models.These topics motivate the paper’s subsequent discussion.Source: §2.2.3 · p. 36

Reasoning versus memorization

• Reasoning applies general truths to specific cases instead of memorizing the world’s exponentially complex probability distribution.Current approaches require increasingly large amounts of data, whereas inference can generalize from principles such as human mortality.Source: §2.3 · pp. 36–37

Rich cognitive models

• Rich cognitive models encode complex scenarios with time-specific information about what protagonists do, know, intend, and expect.Broad commonsense knowledge can be represented explicitly or derived online, enabling sophisticated and interpretable inferences.Source: §2.3 · p. 38

Limitations and requirements

• CYC’s proof of concept depends on extensive manual knowledge engineering, so automatically generating comparable representations would be a major breakthrough.The approach also requires structured representations, variable operations, and records for individuals.Source: §2.3 · pp. 38–39

CYC and rich cognitive models

• CYC provides an existence proof that combining large-scale abstract knowledge with higher-order formal reasoning can support subtle, interpretable inferences.Its Romeo and Juliet example combines a rich cognitive model of events, protagonists, intentions, and expectations with commonsense knowledge about actions and consequences (Lenat, 2019).Source: §2.3 · pp. 38–39

Limitations and requirements

• CYC remains brittle and difficult to apply because most problems lack logical-form inputs, while its handling of uncertainty, incomplete knowledge, and statistical distributions is unclear.Its limited natural-language and visual interfaces restrict immediate applications, and performance depends on database content and scenario mappings.Source: §2.3 · p. 39

Path toward robust reasoning

• Robust AI requires systems that efficiently use large-scale background knowledge despite incomplete information, alongside new benchmarks testing sophisticated reasoning in complex scenarios.Neurosymbolic work such as Minervini et al. (2019) and Besold et al. (2017) offers promising starting points, but scalability may still require improvement.Source: §2.3 · pp. 39–40

Cognitive models

• Cognitive models encode entities, their properties, and temporal events, then support updates and inferences about how situations operate and what may happen next.Understanding is characterized as constructing an internal model and reasoning over it.Source: §2.4 · p. 40
• Inferring the right cognitive model for a situation is complex, can have multiple plausible answers, and remains largely unautomated; systems such as CYC therefore rely on hand-coded models.Robust AI would require inducing models automatically from streams such as text or video.Source: §2.4 · pp. 40–41

Limits of model-free learning

• Model-free reinforcement learning can excel in closed environments with enough data but generalizes poorly: minor Breakout changes caused drastic performance reductions and often required considerable retraining.The proposed explanation is that these systems lack rich cognitive models of their environments.Source: §2.4 · pp. 41–42

Limits of current language models

• Current Transformers often predict locally plausible language without representing, updating, or manipulating enduring entity properties, leaving their broader discourse understanding and reasoning unreliable.Failures to track evolving entities, such as dry-cleaning locations, are presented as evidence of missing cognitive models.Source: §2.4 · p. 42

Implications for robust AI

• Rich cognitive models require a hybrid architecture combining prior knowledge and sophisticated reasoning, and inducing, updating, and reasoning over them should be a highest priority for robust AI.The paper argues that this capability is needed across language and perception, including coherent interpretation of scenes, stories, and human interactions.Source: §2.4 · pp. 44, 45, 47

Limits of current language models

• Meena’s fluent, on-topic responses still lacked an enduring internal model: it repeated the same band as both favorite and least favorite and contradicted itself about whether it spoke Cantonese.The paper contrasts outstanding sentence-level prediction with incoherence across a conversation.Source: §2.4 · pp. 45–46

3.1. Towards an intelligence framed around enduring, abstract knowledge

• Robust AI requires hybrid architectures, large-scale knowledge bases, tractable reasoning, and rich cognitive models, with operations over variables, structured representations, and records for individuals.These prerequisites should support language and higher-level cognition, while integrating learning with symbol manipulation rather than treating learning as the sole first-class capability.Source: §3.1 · p. 48, §3.4 · p. 53
• Hybrid architectures are necessary but not sufficient: robust intelligence also needs rich domain knowledge and mechanisms for spatial, physical, psychological, temporal, and causal reasoning.The space of possible systems within the relevant intersection is vast, and only some systems combining learning, structured representations, individual records, and variable operations will instantiate robust intelligence.Source: §3.1 · pp. 49–50, Figure 5 · p. 50

3.2.1 Engineering practice

• Engineering practice must complement cognitive prerequisites through redundancy, specified tolerances, failure detection, causal models, and replicability.The discussion presents these practices as established engineering techniques relevant to robust AI, drawing on Dietterich (2017) and Henderson et al. (2017).Source: §3.2.1 · pp. 51–52

3.2.2. Culture

• AI progress also requires a culture that welcomes critical outside perspectives and interdisciplinary work rather than penalizing researchers who defend symbol-manipulation.The discussion describes aggression toward criticism and students’ fear that supporting symbolic AI could damage their careers as barriers to progress.Source: §3.2.2 · p. 52

3.3. Seeing the whole elephant, a little bit at a time

• Progress may come from combining complementary strengths: deep learning supplies extensive data-derived patterns, CYC illustrates reasoning over rich knowledge, and NS-CL demonstrates possible neuro-symbolic integration.The discussion advocates breaking out of disciplinary silos and building AI on multiple complementary foundations rather than relying on one pole.Source: §3.3 · p. 53

3.4 Conclusions, prospects, and implications

• The proposed four-step program—hybrid architectures, cognitive frameworks and knowledge databases, abstract reasoning tools, and cognitive-model representation and induction—could provide a substrate for richer, more trustworthy AI.It is intended to shift learning toward abstract, language-like generalizations relative to knowledge and cognitive models, but these prerequisites remain a starting point and cannot guarantee a complete cognitive being.Source: §3.4 · pp. 53–54