Questions for Exploration: Human Intelligence and AI

Open questions about thought, learning, and the evolution of intelligence, organized with related articles, papers and projects.

Part 1: From Humans to AI

Where Do New Ideas Come From?

Spontaneous thoughts, memory, and new input

Q: Why do possible solutions sometimes appear during a shower or a walk?

Further reading: Stanford’s article on walking and creativity (research overview) explains experiments comparing walking with sitting during creative-thinking tasks. Baird et al., Inspired by Distraction (free author PDF) examines creative incubation during an undemanding activity. These explore walking and mind wandering, without establishing a single explanation for everyday inspiration.

Q: How might new input and existing memories contribute to an idea?

Further reading: Addis, Wong, and Schacter on remembering and imagining (free author PDF) compares neural processes in remembering past events and imagining future ones. Schacter and Addis on constructive memory (free full-text review) explores how memory supports imagined experiences. Both inform the memory side of this question; neither directly explains how new sensory input becomes a creative idea.

Novelty and idea generation in AI

Q: Can an LLM produce an idea that is genuinely new?

Further reading: Sakana AI’s introduction to The AI Scientist (project explainer) describes a system that generates ideas, checks related literature, and runs experiments. Si et al., Can LLMs Generate Novel Research Ideas? (free arXiv paper) evaluates generated NLP research proposals using expert judgments. Novelty judged by reviewers or literature searches does not establish that an idea is absent from the model’s training data.

Autonomy, goals, and consciousness

Q: Could an agent initiate work without a new user prompt?

Further reading: Anthropic’s Building Effective Agents (engineering article) explains how agents choose actions and use feedback within an ongoing task. Generative Agents (free arXiv paper) implements memory, reflection, and planning in a simulation. It demonstrates activity without prompting each action, while relying on a configured environment and agent descriptions. Project code and setup guide.

Q: Would initiating work or describing its own thoughts imply consciousness?

Further reading: Butlin et al., Consciousness in Artificial Intelligence (free arXiv report) introduces scientific theories of consciousness and derives proposed indicators for AI systems. It provides a framework for examining consciousness separately from initiative or self-description; the indicators are not a validated diagnostic test.

Can Intelligence Build and Revise Knowledge Through Experience?

Once an agent can act on its own, perhaps through a robot, could it use what it already knows to explore, test new ideas, and revise its understanding? The questions below follow that process, from choosing an investigation to retaining what the experience teaches it.

Using existing knowledge to guide exploration

Q: How could an agent use existing knowledge to formulate new, testable ideas about its environment?

Further reading: Cambridge’s introduction to Adam, the robot scientist (research overview) describes a system that generated and experimentally tested hypotheses about yeast biology. CMU’s introduction to Coscientist (research overview) explains how language models, chemical information, and laboratory tools can work together to plan and perform experiments. These offer bounded examples within human-defined research settings.

Q: Without a person specifying every task, how could an agent decide what is worth exploring next?

Further reading: The Plan2Explore project page (illustrated overview and code) explains how a learned world model guides exploration before downstream tasks are specified. Oudeyer, Kaplan, and Hafner on intrinsic motivation (free author PDF) studies exploration driven by progress in predicting sensorimotor outcomes. These implement particular exploration objectives, rather than unrestricted human-like curiosity.

Testing ideas through interaction with the world

Q: How could a robot choose actions that distinguish between competing explanations, rather than merely complete a task?

Further reading: Cornell’s article on a robot that discovers its own body (research overview) explains how a robot chooses actions that distinguish competing models of its body, then refines those models from observations. Cambridge’s article on Adam (research overview) provides a laboratory counterpart: generating hypotheses and testing them through automated experiments. Both connect action selection to resolving uncertainty in constrained domains.

Q: What would count as new knowledge: a better prediction, a reusable understanding, or a discovery previously unknown to humans?

Further reading: Plan2Explore (project overview) learns a predictive model through exploration and evaluates its usefulness for later control tasks. Cambridge’s account of Adam’s discoveries (research overview) describes biological findings that the researchers checked through manual experiments. These illustrate different standards: acquiring useful knowledge for an agent and contributing new scientific knowledge for humans.

Correcting earlier beliefs and retaining what is learned

Q: When observations contradict an agent’s expectations, how could it identify and revise the part of its existing understanding that is wrong?

Further reading: Cornell’s account of a robot adapting to injury (research overview) describes a robot revising its body model after physical damage and developing a new gait. Plan2Explore (project overview) describes learning a predictive world model from exploration. These study model revision through experience; neither provides a general method for distinguishing an incorrect belief from sensor noise, execution failure, or a changed environment.

Q: How could discoveries and corrections become lasting knowledge that remains useful in unfamiliar situations?

Further reading: Voyager (illustrated project explainer) explains its growing library of executable skills and their reuse in new Minecraft worlds and unseen tasks. Plan2Explore (project overview) examines whether a world model learned during exploration supports later tasks. These test skill and model reuse, while leaving open how an embodied agent could retain and generalize corrected factual knowledge over long periods. Voyager code.

How Can Agents Build Knowledge Together?

If agents can share observations, memories, code, and model weights, how could their separate experiences contribute to collective learning? The questions here concern direct sharing, joint learning, verification, and specialization, without assuming that knowledge must pass through biological-style generations.

Sharing experience and capabilities

Q: What should agents share: observations, conclusions, reusable skills, or model updates?

Further reading: Google Research’s introduction to federated learning (technical blog) explains how devices contribute model updates while keeping training data local. Large Language Models as Tool Makers (free arXiv paper) illustrates a different kind of sharing: reusable programs created by one model and used by another. These offer concrete examples of what can be transferred, without supplying a universal rule for choosing what to share.

Shared models, specialization, and independent judgment

Q: When is one shared model sufficient, and when should agents retain specialized models or local knowledge?

Further reading: IMPALA (open-access paper; free PDF on the page) trains a shared policy across multiple tasks. Smith et al., Federated Multi-Task Learning (free arXiv paper) learns related but distinct models for clients with different data distributions. These provide contrasting approaches to sharing and specialization; their different experimental settings do not establish a general winner.

Part 2: From AI to Humans

How Does the Brain Think and Remember?

Processing information and storing memories

Q: How do neurons work together to process information, and what changes when we learn or remember something?

Further reading: Ebitz and Hayden on neural populations (free arXiv manuscript) examines how groups of neurons support attention, working memory, and decision-making. Bailey, Kandel, and Harris on learning and memory (free full-text review) explores how learning changes connections between neurons.

Comparing the brain with AI

Q: How does the complexity of the brain compare with that of an LLM?

Further reading: Beniaguev, Segev, and London, Single Cortical Neurons as Deep Artificial Neural Networks (free preprint) uses artificial networks to approximate the responses of simulated biological neurons, offering a perspective on complexity beyond neuron and parameter counts.

Q: Could a computer model fully reproduce how a human brain works?

Further reading: Izhikevich and Edelman, Large-Scale Model of Mammalian Thalamocortical Systems (free author PDF) examines how neural structure and interactions give rise to patterns of activity in a large-scale simulation. The model combines evidence from several species and studies selected brain systems; it does not reproduce an entire human brain.

What Can We Learn from Recording Brain Activity?

Measuring brain activity

Q: What can brain recordings tell us about the activity of neurons, and what information do they leave out?

Further reading: Buzsáki, Anastassiou, and Koch on recorded brain signals (free full-text review) explains where electrical recordings come from. Wendel et al. on EEG and MEG (free full-text review) examines the challenges of locating the activity behind those measurements.

Reading perception and thought

Q: Could brain recordings reveal what a person is seeing, thinking, or remembering?

Further reading: Tang et al., Semantic Reconstruction of Continuous Language from Non-invasive Brain Recordings studies language reconstruction from fMRI. Kriegeskorte and Douglas on interpreting decoding models discusses what successful decoding does and does not show. These help distinguish information recovered from brain activity from a model’s predictions.

Could We Provide Input to the Brain and Influence Its Activity?

Artificial sensory input

Q: Could artificial neural input create an interactive world in which someone experiences people, sounds, and touch that are absent from their physical surroundings?

Further reading: Beauchamp et al. on visual cortex stimulation (free PDF) examines electrically evoked visual forms. NIDCD’s guide to cochlear implants (public overview) describes artificial input to the auditory nerve. UPMC’s account of artificial touch feedback (research overview) describes artificial sensations linked to a person’s control of a robotic arm. These address individual sensory functions, without demonstrating a complete, lifelike simulated environment.

Influencing ongoing thought

Q: Could stimulating the brain trigger a particular thought or change how someone interprets information and reaches a decision?

Further reading: Histed, Ni, and Maunsell on cortical stimulation (free full-text review) examines effects on perception and behavior. Fox on the origins of self-generated thought (free arXiv chapter manuscript) reviews recording and stimulation studies of internally generated experiences. These investigate influences on cognition without demonstrating reliable control over arbitrary thoughts.

Reading and changing memories

Q: Could a specific memory be read, modified, or transferred to another brain?

Further reading: Roy et al. on memory traces across brain regions (free full text) studies neurons involved in a particular memory in mice. Liu, Ramirez, and Tonegawa on experimental memory manipulation (free full text) examines the formation of a fear association in mice. These explore parts of memory storage and modification, rather than a general method for copying or rewriting human memories.

Questions for Exploration: Human Intelligence and AI

https://kongchenglc.github.io/blog/2026/10/11/AI-Human/

Author

Chili

Posted on

2026-10-11

Updated on

2026-11-16

Licensed under

Comments