From our discussions marks work that came up directly in our research conversations. Added for context marks work we've added to round out a topic and haven't yet discussed in depth.
This isn't a canon or an exhaustive review. A paper appearing here doesn't mean it supports our broader thesis. For each one, we try to ask what was supplied, what emerged, what was measured, what resources were used, and what actually generalizes. Entries marked Source checked have been checked against the primary source for this update; that does not mean Rubedo has reproduced or independently validated the reported results.
For the current agenda, start with the frontier sources below and the research questions they motivate. For the broader inquiry, start with the OEE workshop paper, novelty search and evolvability, then compare curiosity, predictive maps, POET, cultural evolution and the civilization-scale work.
Frontier learning, search and verified discovery
Primary sources checked on 14 October 2026. This is a selected evidence map, not a ranking of current models or an exhaustive survey. Reported findings remain bounded by their tasks, compute and evaluation procedures.
Reports reasoning improvements from RL on a pretrained base, with additional stages in the final R1 pipeline. Motivates separating training ingredients and inherited capabilities.
Limits: reported benchmark gains do not isolate every contribution or establish open-ended discovery. Starting without preliminary supervised fine-tuning is still starting from pretraining.
Reports better small-k performance but limited or reduced large-k coverage for tested RLVR setups. Makes sampling efficiency versus capability expansion an empirical question.
Limits: finite sampling probes an empirical boundary, not every possible strategy. Results do not establish a universal impossibility for RL or multi-turn interaction.
Reports improved pass@1 and pass@256 after locating difficult examples, adding targeted teacher guidance and consolidating with RL.
Limits: teacher guidance introduces information, so this does not demonstrate expansion by reward-only learning. The sampled boundary is an empirical proxy; Rubedo has not reproduced it.
Studies verifier-guided search and adaptive response revision, finding that effective compute allocation depends on problem difficulty.
Limits: mathematical tasks and particular proposal/verifier setups. Extra inference computation differs from lasting changes to model weights, and verifier cost belongs in the comparison.
Demonstrates interleaved reasoning and actions, including external information access, in selected question-answering and interactive environments.
Limits: prompting a tool loop is not itself RL training. Task demonstrations and tool interfaces supply structure; success does not certify reasoning-trace faithfulness.
Combines LLM proposals, evolutionary program search and automated evaluation; reports mathematical discoveries and infrastructure improvements.
Limits: human problem design, evaluators and pretrained models supply substantial structure. A verified construction or infrastructure gain does not establish unlimited autonomous science or recursively improving general intelligence.
Maintains an archive of coding agents and empirically tests modifications to their implementations; reports improvements on coding benchmarks.
Limits: empirical selection is not a proof that every modification is beneficial. Agent implementation improvements in a bounded evaluation do not establish indefinite recursive capability growth.
Controls tools, prompts and compute across agent configurations. Reports task-dependent gains and losses, coordination overhead and differing error propagation.
Limits: evaluated architectures and benchmarks support conditional conclusions. They do not establish that more agents always help or that shared memory produces cumulative discovery.
MAP-Elites preserves high-performing solutions across user-defined behavioral dimensions, offering a concrete archive mechanism for quality-diversity search.
Limits: descriptors embed design choices. Diverse archive coverage is not evidence of transfer or open-ended innovation; applying it to LLM workflows is a Rubedo hypothesis.
Open-ended search and stepping stones
Why objectives can be deceptive, and how search can keep finding new things without converging.
A starting point for the whole problem. It separates the observable hallmarks of open-ended evolution from candidate mechanisms, and lays out the terminology and disagreements.
Limits: it doesn't prescribe a universal mechanism or guarantee ever-increasing intelligence.
Shows how objective-driven search can be deceptive, and demonstrates behavioural novelty as an alternative selection criterion. The intermediate discovery doesn't have to look closer to the goal.
Limits: novelty needs a behavioural comparison, which carries assumptions. Success in the studied domains isn't proof that novelty alone yields general intelligence.
The accessible, book-length version of the novelty-search argument, extended to science, careers and institutions. It's close to our stepping-stones question in spirit.
Limits: an argument and a perspective, not an empirical result about any particular system.
Generates environments and optimizes agents for them together, transferring solutions between problem–solution pairs. It's a clear example of indirect stepping stones and a self-generated curriculum.
Limits: the environment family and generator are engineered. It doesn't show arbitrary autonomous invention of environments.
The canonical quality-diversity algorithm. It keeps the best solution found in each cell of a behavioural space, preserving diverse alternatives rather than a single optimum.
Limits: the behavioural descriptors are usually hand-chosen, which is where the designer's assumptions enter.
Archive interesting states, return to them, then explore further, rather than losing progress to random wandering. It produced large gains on notoriously sparse-reward games.
Limits: something still has to define which states are "interesting." In a rich world, handing that to humans brings back the milestone problem.
Argues for generating general AI by meta-learning architectures and learning algorithms and by automatically generating environments. These are close to Rubedo's evolvability and self-generated-curriculum questions.
Limits: a research agenda and position paper, not a demonstration.
Proposes a working definition of open-endedness in terms of novelty and learnability from an observer's perspective, and argues that open-ended systems are necessary for superhuman capability.
Limits: a position paper. The observer-relative definition is useful, but it's a choice, not a settled consensus.
A lead into recent work on evolving complexity, evolvability and organizational transitions.
Limits: we haven't yet reviewed the full issue, so treat any summary as provisional.
Evolvability and learning across generations
How inherited organization shapes future variation, and how evolution and lifetime learning interact.
Connects complex adaptation to the organization of variation. It's how we distinguish "well adapted now" from "able to produce useful variation later." Developmental organization may matter as much as policy parameters.
Limits: it doesn't show that every evolving system becomes more evolvable.
The classic, tiny simulation of the Baldwin effect: lifetime learning smooths a needle-in-a-haystack fitness landscape so evolution can find the needle.
Limits: a deliberately minimal model.
Selection shapes initial parameters and learning hyperparameters so that supervised and reinforcement learning are fast within a lifetime. It's an operational bridge between evolution and learning.
Limits: uses designed inner learning algorithms and task distributions. It doesn't evolve a general learner from nothing.
Draws formal parallels between evolution and learning theory. Evolving systems can, under some conditions, generalize from past selective environments. It's central to our view of evolvability as a kind of learning.
Limits: the correspondence holds for particular models and conditions, not evolution in general.
Examines transitions in which independent replicators become parts of a larger whole and information transmission changes. Adaptation sometimes changes the unit of organization itself.
Limits: applying it to AI is an analogy that needs its own definitions and evidence.
An update to the major-transitions framework, including the role of new inheritance systems, which is relevant to cultural and artificial inheritance.
Limits: a theoretical synthesis, not a computational recipe.
Origin of the error threshold: replication fidelity limits how much information a lineage can maintain under selection. This was an early Rubedo interest, as a constraint on developmental genomes.
Limits: derived for molecular quasispecies. Carrying it over to neural developmental systems is our own analogy.
NEAT evolves network structure as well as weights, using speciation to protect innovation. HyperNEAT adds an indirect, developmental encoding. Both are reference points for genome → phenotype mappings.
Limits: speciation uses a hand-designed compatibility distance, exactly the kind of mechanism we hoped would emerge on its own.
In the Avida digital-evolution platform, complex functions evolved by building on simpler ones that were rewarded, and some intermediate steps were even deleterious. It's a concrete stepping-stone result.
Limits: the rewarded simpler functions were designer-chosen.
CPPNs are small networks of composed functions that paint patterns over coordinates, an abstraction of development without simulating chemistry. They're the encoding behind HyperNEAT and our MNIST experiment.
Limits: strongly biased toward smooth, regular, geometric structure, which is a poor fit for layers with no natural geometry.
One network generates the weights of another, trained end to end with gradients. It's the gradient-based cousin of an evolved developmental generator.
Limits: naive hypernetworks scale badly with the size of the target network, so useful versions use low-rank, shared or factorized outputs.
Learns not just weights but how much each connection should change during a lifetime. It's a concrete way to make plasticity rules themselves trainable or heritable.
Limits: demonstrated on modest tasks. Plasticity is learned by gradients, not evolved.
Networks can compute the same function with neurons in different orders. Aligning those permutations first can let two independently trained models be merged. It's relevant to recombining two agents' learned weights without destroying both.
Limits: works best for models trained on the same task and data. Alignment between very different agents is open.
Self-replicating machine-code programs competing for CPU time and memory. Parasites, hyper-parasites and size optimization emerged without being designed. It's the ancestor of our heredity toy.
Limits: evolution tended to plateau rather than continue open-endedly.
Generalized optimizers and learning to learn
Can a system improve its own way of generating solutions? Theory on the limits of general optimization, and results that move parts of the learning machinery into the learned system.
Averaged over all possible problems, no optimizer beats any other. "General" optimization must therefore mean matching the structure of the problems that actually occur, which is a strong argument for studying inductive biases.
Limits: the uniform average over all problems is not the distribution of problems in our world.
Defines intelligence as an agent's ability to achieve goals across a wide range of environments, weighted by simplicity. It builds on Hutter's AIXI, the idealized and uncomputable general agent.
Limits: uncomputable in its pure form. Useful as a conceptual reference point, not a recipe.
A theoretical design for a system that rewrites any part of itself, including its search procedure, once it can prove the rewrite is beneficial. It's the formal extreme of recursive self-improvement.
Limits: proof search makes it impractical. It marks a boundary rather than a path.
Replaces a hand-designed optimizer with a learned one, so the update rule itself becomes a trained network.
Limits: generalization of learned optimizers beyond their training distribution has been a persistent difficulty.
Learns an initialization from which a few gradient steps adapt quickly to new tasks. It's a machine-learning cousin of the Baldwin effect.
Limits: assumes a task distribution and a fixed inner learning rule.
Discovers changing hyperparameter schedules through population-level selection and perturbation. It's the nearest common practice to search that adjusts its own settings.
Limits: its exploration and replacement procedures are themselves engineered.
Evolves entire learning algorithms from basic mathematical operations and rediscovers techniques such as gradient descent along the way. It's a striking demonstration of search producing learners.
Limits: small problems, an engineered search space and large compute.
Meta-learns an RL update rule, including what the agent should predict, from interaction with a population of environments. It's a precursor to DiscoRL.
Limits: discovered in simple environments. Transfer was promising but bounded.
Learned an RL update rule from populations of agents across diverse environments. The rule performed strongly on benchmarks, including environments held out during discovery. It's evidence that useful learning machinery needn't all be specified by hand.
Limits: a designed meta-learning framework, bounded by the environments and evaluation it used.
Meta-learns evolution-strategy update rules that transfer to unfamiliar problems, and even uses a learned strategy to drive its own outer optimization. It searches for better ways of generating solutions, not just solutions.
Limits: bounded by the tested problem classes. Not evidence of unrestricted self-improvement.
Self-modifying networks that improve themselves on bandit and classic-control tasks without a separate meta-optimizer. It shows self-modification can be studied experimentally.
Limits: simple tasks and a designed resource-allocation procedure. It doesn't show that self-modification leads to ever-greater intelligence.
An evolutionary coding system that found improvements to data-centre scheduling and to components used in training the models it runs on. It's a bounded, real example of search improving its own supporting infrastructure.
Limits: heavily engineered, with human-defined evaluators. Not an autonomous, accelerating resource loop.
Two reference points for black-box evolutionary optimization: ES scaling across thousands of workers, and CMA-ES adapting the shape of its own search distribution. That second point is a modest, well-understood form of a search that structures its own variation.
Limits: these optimize fixed objectives. They're tools, not open-ended systems.
General methods that scale with computation have repeatedly beaten methods built on human knowledge. It's a strong argument against prescribed milestones, and a reminder that scaling assumes the compute is there.
Limits: an essay drawn from history, not a theorem. "General methods" still embed substantial engineering.
Curiosity, development and autonomous goals
What drives exploration when nobody specifies what's worth learning.
Maps the many computational meanings of intrinsic motivation (surprise, novelty, competence, learning progress) and exposes ambiguities in the term.
Limits: a taxonomy, not one universal drive.
Separates sampling information for a known purpose from exploring to discover what's worth learning, and bridges developmental learning and neuroscience.
Limits: informs the question without giving a complete artificial mechanism.
Agents that represent, generate, prioritize and pursue their own goals to build skill repertoires. It's the closest literature to our "self-generated developmental goal."
Limits: learned goals still depend on supplied interfaces and representations.
Curiosity as reward for compression progress, meaning improvement in the agent's ability to predict or compress its history. It's an early, formal version of learning progress.
Limits: elegant in theory. Measuring compression progress in practice is hard.
Prediction-error curiosity in deep RL, and a large study showing both how far it goes with no extrinsic reward and how it fails with stochastic distractions, the noisy TV.
Limits: prediction error isn't learning progress. These papers illustrate the problem as much as the solution.
Explores without a task by seeking states where an ensemble of world models disagrees, then adapts to tasks later. It's discovery of what in a world is worth learning.
Limits: doesn't by itself connect discoveries to an eventual objective.
Learns a repertoire of diverse behaviours without extrinsic reward, which a higher-level policy could later choose among.
Limits: simpler domains. Not shown to scale to technology-tree-level skills.
Empowerment is the channel capacity between an agent's actions and its future states, a task-independent measure of future controllability. It's a candidate formal reason to prefer infrastructure over immediate payoff.
Limits: expensive to compute, and over-weighting it produces agents that never cash in.
Prediction, world models and abstraction
How recurring structure becomes reusable knowledge, and how learners use imagination without being fooled by it.
Explains hippocampal representations with the successor representation: encoding expected future occupancy, not just location.
Limits: policy-dependent. Not a full causal model or a theory of autonomous abstraction.
Links schema learning, RL, dimensionality reduction and memory reactivation. It's a bridge from recurring structure to reusable knowledge.
Limits: a perspective, not proof that compositional concepts follow from prediction.
A model that learns when to spend computation on internal rollouts, relating them to human behaviour and replay.
Limits: replay-like rollouts are designed into the model. It doesn't show autonomous invention.
One algorithm learns many control tasks by training largely on imagined trajectories from a learned world model. It also collected diamonds in Minecraft from pixels.
Limits: the Minecraft result used 12 milestone rewards and abstracted crafting. "Same algorithm across domains" isn't one agent learning cumulatively across them.
Plans inside a learned model without being given the environment's rules.
Limits: tree search becomes intractable in open worlds without learned temporal abstraction.
A manager picks latent goals, not human-named ones, for a worker, all trained in imagination. It's the closest precedent for a system choosing its own milestones.
Limits: tasks far simpler than an open technology tree. The mechanism exists; the scale doesn't.
Short model rollouts branched from real experience avoid compounding model error. Useful imagination doesn't mean simulating everything.
Limits: tested settings. Doesn't solve real-world planning generally.
Two sides of continual learning: protecting important weights so old skills survive, and the discovery that continually trained networks can gradually lose the ability to learn at all.
Limits: partial remedies, each with costs. Stable and plastic remain in tension.
Agent architecture: memory, actions and communication
What a single learning agent needs before a population can do anything interesting. See Agents.
Embeds actions in a continuous space and looks up nearest neighbours, so an agent can choose among very large numbers of discrete actions without a giant output layer.
Limits: depends on a good action embedding. Nearest neighbours can miss the best action.
Choose a discrete action type, then its continuous parameters. That's the clean form of an action that "calls a function."
Limits: large discrete arguments, such as which node or which asset, still need another mechanism.
Output a choice by pointing at one of a variable set of inputs, which is the basis for choosing a counterparty or asset from a candidate set that keeps changing.
Limits: the policy can only choose among candidates that retrieval surfaces.
Learns temporally extended skills (options), the policies inside them, and when to stop them, all end to end. It's the hierarchical version of a high-level choice handing off to a specialized model.
Limits: options often collapse to trivial or one-step behaviour without extra regularization.
A router sends each input to a few expert sub-networks, with load-balancing so no expert dominates. It's a reference point for modular policies.
Limits: designed for supervised scale, not for RL credit assignment.
The Differentiable Neural Computer: a recurrent controller that learns to read and write an external memory. It's an early instance of the "recurrent core plus external memory" tier.
Limits: hard to train and scale. Simpler key–value and attention memories are more common in practice.
Two foundations for strategic communication. Costly signals can keep honesty stable (the handicap principle), and people coordinate on focal points without negotiating (Schelling points), which is the inspiration for our "Schelling-point alignment" idea.
Limits: classic theory. How either applies to learning agents is part of what we want to test.
Multi-agent systems, culture and collective intelligence
Populations, autocurricula, communication, institutions, and knowledge that outlives the individual.
Argues that innovation arises when agents' adaptations keep creating new challenges for one another, a self-generating curriculum. It's very close to Rubedo's original multi-agent thesis.
Limits: a manifesto. Sustained autocurricula remain hard to produce.
Hide-and-seek agents discovered successive strategies, including building shelters, using ramps and locking tools, each in response to the other side, with no reward for any of them.
Limits: the number of distinct innovations was bounded by a small, designed environment.
A procedurally generated universe of multi-agent games and a dynamically adapting task distribution produced agents with broad zero-shot competence.
Limits: very large compute and an engineered task space.
Capture the Flag agents trained as a population, with evolved internal rewards. It's an example of drives shaped by selection rather than hand-tuned.
Limits: one game family. The evolved rewards are weightings over designed signals.
Workhorse methods for centralized training with decentralized execution. These are the practical baselines for any toy market or multi-agent environment.
Limits: engineering baselines, not theories of emergence.
How agents can invent communication protocols, and a survey of what emergent languages do and don't look like.
Limits: emergent protocols are often brittle, uninterpretable or degenerate.
Learning agents sharing a renewable resource reproduce tragedy-of-the-commons dynamics, and some ways out. It's a direct test bed for scarcity-driven social structure.
Limits: small gridworlds.
A simulated economy with learning agents and a learning policy-maker optimizing over them. It's a concrete example of a synthetic economic "wind tunnel."
Limits: highly stylized economy. External validity is the whole question.
Maps the capabilities needed for cooperation (understanding, communication, commitment, institutions) across agents with differing interests.
Limits: an agenda, not results.
Culture as an additional source of open-endedness, and a case where open-endedness itself evolved.
Limits: a theoretical foundation, not a demonstration of artificial cumulative technology.
Cumulative culture, the "ratchet effect", as the driver of human capability. No individual could rediscover what their culture knows.
Limits: about humans. Whether artificial populations can ratchet is open.
Organisms modify their environments and thereby change their own selective pressures. Environment and agent co-evolve.
Limits: a biological framework. Computational analogues are still young.
Real communities sustain shared resources without either markets or central control, through institutions they design themselves. It's evidence that stable cooperative structure can emerge from below.
Limits: empirical case studies of human communities.
Decentralized and distributed AI
Training and running AI without a single centre, and ecosystems of specialized intelligence rather than one monolith.
The foundational federated-learning algorithm: many devices train locally and share only model updates.
Limits: decentralizes data, not goals. All participants optimize one shared objective.
Training and serving large models over volunteer hardware on the open internet, with unreliable, heterogeneous participants.
Limits: these solve coordination of compute, not incentives or open-ended goals.
Islands of compute train mostly independently and synchronize rarely, cutting communication by orders of magnitude. It suggests large training needn't sit in one data centre.
Limits: still one model and one objective.
An attempt to reward independently hosted models through a token-based market for useful outputs. It's directly relevant to our question of rewarding real usefulness in a decentralized network.
Limits: we read it as a case study in incentive design. Evaluator gaming is exactly the risk such systems face.
Argues that general capability may arrive as an ecosystem of specialized services rather than a unified agent, which is close to our "distributed superoptimizer."
Limits: a conceptual argument. Whether service ecosystems stay non-agentic is disputed.
Intelligence as the interaction of many simple, mindless agents. It's an early statement that minds may be collectives.
Limits: a philosophical framework, not a computational model.
Instrumental convergence, power and systemic risk
Why many goals reward capacity-building, and what happens when optimizers, singly or together, pursue it.
Argues that sufficiently capable goal-seeking systems develop drives for self-improvement, self-protection and resource acquisition regardless of their goal.
Limits: informal argument.
Names instrumental convergence: very different final goals converge on similar intermediate incentives. Our paperclip and diamond intuition is a known phenomenon.
Limits: philosophical. It doesn't show any particular learning system develops these incentives.
Formalizes power as keeping access to many future options, and proves conditions under which optimal (and later, more general "retargetable") decision-makers favour it. It's the nearest mathematical grounding for future controllability.
Limits: holds in the formal settings studied. Extending it to large learned agents is analogy.
Introduces mesa-optimization: an optimizer (training) can produce a learned system that is itself an optimizer, with its own objective. That's directly relevant to "optimizers made of optimizers."
Limits: a conceptual framework. Empirical evidence is still developing.
Splits Goodhart's law into regressional, extremal, causal and adversarial forms. It's a precise vocabulary for evaluator failure.
Limits: a taxonomy, not a remedy.
A careful, premise-by-premise examination of the argument from power-seeking to catastrophe, with explicit probabilities.
Limits: the probabilities are subjective by design.
Both foreground multi-stakeholder, multi-system settings, where risk comes from interactions among many AI systems and many humans rather than one misaligned agent.
Limits: taxonomies and research agendas.
Argues that incremental replacement of human participation in economies, culture and states could erode human influence with no single dramatic failure. It's the risk side of "civilization as an optimizer."
Limits: a scenario analysis, necessarily speculative.
A structured account of failure modes specific to many interacting AI agents: miscoordination, conflict and collusion.
Limits: a survey and taxonomy.
Two essays on competitive selection producing outcomes no participant values. They're the clearest statements of an emergent objective nobody chose.
Limits: essays, not models.
Economics, technology and industry
Capital, ideas and coordination: the economic machinery any civilization-scale optimizer runs on.
Prices aggregate dispersed knowledge no one holds in full. That's why markets can coordinate without a centre, and why prices could coordinate machine agents too.
Limits: about human markets. Machine markets may behave differently, for example in speed and in collusion.
Bounded rationality and satisficing: real optimizers work within limits on computation and information, and the structure of the environment determines what simple strategies can achieve.
Limits: about organisms and people, but the lesson generalizes.
Ideas as the engine of growth, and a long-run model in which more people produce more ideas, which support more people. Both are formal versions of a capability feedback loop.
Limits: aggregate models. They describe the loop rather than its mechanisms.
Technology evolves by combining existing technologies into new ones, combinatorial evolution. It's a framework for our question about agents recombining discovered mechanisms.
Limits: a descriptive theory of human technology.
Philosophy and physics of optimization
Ideas about search, selection and optimization as features of nature. Several are contested, and we read them as provocations.
Natural selection as a substrate-neutral algorithm, a mindless procedure that produces design. It's the philosophical basis for "generate, evaluate, select, amplify."
Limits: a philosophical interpretation, not an empirical result.
Rugged fitness landscapes (the NK model), self-organization, and the "adjacent possible," the set of things reachable one step from what exists. It's a useful vocabulary for open-ended expansion.
Limits: influential but partly speculative, especially the later work.
The maximum power principle: selection tends to favour systems that capture and use more of the available energy flow.
Limits: a principle with a long, contested history, not a law.
A thermodynamic bound relating replication to heat dissipation. It's the basis for the "dissipative adaptation" view that physics may tilt toward self-replicating structure.
Limits: the broader interpretation goes well beyond what the bound shows.
A system pushed toward states with the greatest diversity of accessible futures shows tool-use-like and cooperation-like behaviour in simple simulations. It's a physical cousin of empowerment and option value.
Limits: toy demonstrations. The claims about intelligence are speculative.
Persisting self-organizing systems can be described as minimizing a bound on surprise, which unifies perception, action and learning under one quantity.
Limits: whether it's falsifiable theory or reframing is actively debated.
Estimates the total number of elementary operations the observable universe could have performed. It puts "physics performs the computation" on a number line.
Limits: an upper bound on physical operations, not a measure of useful search.
Proposes energy rate density as a common measure of complexity, from stars to societies. It's one attempt at a substrate-neutral account of rising organization.
Limits: a broad synthesis whose single metric is disputed.
The origin of the many-worlds interpretation. We cite it because it often comes up in branching-search analogies.
Limits: nothing in it involves evaluating or selecting branches. The optimization reading is purely our speculation.
The simulation argument, a trilemma about civilizations and simulations.
Limits: a probabilistic argument, not evidence of any simulator or its goals.
Minecraft and embodied-agent references
Engineering references from the diamond testbed discussion.
An inverse-dynamics model labels about 70,000 hours of unlabelled gameplay video, producing a behavioural prior with native mouse and keyboard control. RL fine-tuning reached diamond tools.
Limits: inherits human strategy, the opposite of autonomous discovery.
Goal-conditioned behaviour built on top of a VPT-style policy.
Limits: depends on a human-demonstration prior and a designer-defined goal interface.
A language model writes and stores reusable code skills with an automatic curriculum. It's a useful contrast: its symbolic skill library sidesteps catastrophic forgetting.
Limits: imports human knowledge through the language model. It's excluded from our sensory-to-action benchmark by design.
A large open-world Minecraft benchmark that also explores learned reward functions.
Limits: a benchmark, not a solution to open-ended discovery.
Leads we haven't verified
These came up in our discussions but we haven't yet checked them properly. Treat them as pointers, not foundations.
- AutoQD, reportedly deriving quality-diversity descriptors automatically from occupancy measures (supplied ICLR 2026 link). The title, status and method are unverified.
- A 2024 review of the neuroscience of curiosity and information-seeking (lead).
- Experimental evidence that people choose activities where they're making learning progress (lead). This doesn't mean human curiosity is optimal.
- Dopamine beyond scalar reward-prediction error (lead). It's a reason for caution about simple "the brain is RL" analogies.
- Open-world continual learning (lead) and embodied evolution with learning (lead).
- Recent critiques arguing that existing arguments don't support a universally strong form of instrumental convergence, and recent reviews of self-driving laboratories and closed-loop manufacturing. Several were raised in our industrial-process discussion and still need checking.