By Jim Shimabukuro (assisted by Claude)
Editor
Introduction: Claude and I had a conversation about the fundamental building blocks for a learning model. To sidestep the confusing backdrop of countless learning and developmental theories, we consciously avoided references to them. We ran through a few narratives to illustrate the five blocks and ended up with a short story, Algorithm of an Intentional Heart, that breathes life into this construct. -js
JS: Good morning, Claude. If we were to create a state-of-the-art learning bot, what are the simplest machine learning rules we would program into it? Regarding number of rules, ideally the smaller the number the better. To keep the process open, let’s refrain from using terms that are tied to established learning theories.
Claude: Good morning! For a state-of-the-art learning bot with minimal rules, I’d focus on these fundamental principles:
1. Prediction and Error Correction The bot continuously makes predictions about what comes next (whether in language, patterns, or outcomes) and adjusts its internal parameters when predictions prove wrong. The size of adjustment is proportional to the prediction error.
2. Pattern Compression The bot seeks the most efficient way to represent information it encounters, automatically finding regularities and building compressed representations that capture the essence while discarding noise.
3. Multi-Scale Attention The bot dynamically focuses on different aspects of input at multiple levels of granularity – from fine details to broad context – allocating more processing power to what seems most relevant for the current task.
4. Hierarchical Abstraction The bot builds layered representations, where each level captures increasingly abstract features from the level below, allowing it to work with concepts rather than just raw data.
5. Adaptive Exploration vs Exploitation The bot balances between using what it has already learned (exploitation) and trying new approaches (exploration), with the balance shifting based on confidence levels and environmental feedback.
These five rules could theoretically generate sophisticated learning behavior, as they capture the core mechanics of how information gets processed, refined, and applied in intelligent systems. Each rule is simple to state but can generate complex emergent behaviors when implemented at scale.
JS: Let’s transfer these rules to a human child who encounters a puppy for the first time. Without any assistance from other humans, illustrate how these five rules function in the child’s efforts to learn about the puppy.
Claude: Here’s how those five learning rules would manifest as a child encounters a puppy for the first time:
Rule 1: Prediction and Error Correction: The child sees the small, furry creature and predicts it might be soft like their stuffed animals. They reach out to touch – confirming the softness prediction. But then the puppy moves suddenly and licks their hand. This unexpected warm, wet sensation violates their “inanimate object” prediction, forcing a major adjustment: this thing is alive and interactive.
Rule 2: Pattern Compression: Initially overwhelmed by countless details (fur texture, ear shape, tail movement, sounds, smells), the child’s mind starts organizing: “Four legs + fur + wet nose + wagging tail + happy sounds = this type of creature.” They’re building an efficient mental template that captures the essential “puppy-ness” while filtering out irrelevant details like the exact number of spots.
Rule 3: Multi-Scale Attention: The child’s focus shifts dynamically – zooming in on the puppy’s eyes when it looks at them directly, then pulling back to notice how the whole body moves when excited, then focusing on specific sounds the puppy makes. When the puppy approaches their face, attention narrows intensely to those immediate sensations and movements.
Rule 4: Hierarchical Abstraction: Low level: “Wet, pink thing touching my hand.” Mid level: “Tongue that licks.” Higher level: “Friendly greeting behavior.” Abstract level: “This creature shows affection differently than humans.”
Rule 5: Adaptive Exploration vs Exploitation: Initially cautious (low exploration), the child stays back observing. As positive interactions accumulate, they become bolder – trying to pet different areas, making sounds to see if the puppy responds, experimenting with different movements. But if the puppy suddenly barks loudly, they quickly revert to exploitation mode, using their now-familiar “stay still and quiet” strategy until they assess the situation.
Through these rules operating together, the child builds a rich, nuanced understanding of “puppy” without any verbal instruction.
JS: Brilliant! Let’s now transfer these five rules to a robot that we send to a potentially habitable planet in or near our galaxy. As you did with the child and dog, illustrate how the robot applies the five rules on its initial encounter with a higher-order life form.
Claude: Here’s how the robot would apply these five learning rules upon encountering an unknown higher-order life form on the distant planet:
Rule 1: Prediction and Error Correction: The robot’s sensors detect a large, crystalline structure that appears geometric and static. Based on planetary geology data, it predicts “mineral formation.” But then the structure suddenly shifts color from blue to amber, and pseudopod-like extensions emerge from its base. The robot’s classification algorithms rapidly recalibrate: “Not geological – biological entity with dynamic properties.” When the robot approaches, the entity retracts, violating the “plant-like sessile organism” prediction, forcing another major parameter adjustment toward “mobile, responsive life form.”
Rule 2: Pattern Compression: The robot processes massive sensory input: electromagnetic emissions across multiple spectra, complex geometric transformations, chemical signatures, vibrational patterns. Its compression algorithms identify core patterns: “Crystalline base + color-change capability + responsive movement + electromagnetic field generation = coherent biological system.” It builds an efficient representation focusing on these key identifying features while discarding redundant data about specific facet angles or minor color variations.
Rule 3: Multi-Scale Attention: The robot’s attention management system dynamically allocates processing power – initially spread across environmental monitoring, it rapidly focuses on the entity’s color transitions when they correlate with the robot’s movements. When the entity extends appendages, attention narrows to track their trajectory and speed. As chemical emissions intensify, olfactory sensors receive priority processing while visual analysis continues at reduced resolution.
Rule 4: Hierarchical Abstraction: Low level: “Silicon-carbon lattice structures + photonic emissions.” Mid level: “Coordinated structural changes + deliberate movement patterns.” Higher level: “Intentional responses to external stimuli.” Abstract level: “Intelligent entity capable of assessment and reaction.”
Rule 5: Adaptive Exploration vs Exploitation: Initially, the robot maintains safe distance while maximizing sensor data collection (high exploration). As the entity’s responses prove non-hostile, it incrementally advances closer, testing different approach angles and speeds. But when the entity suddenly pulses with intense electromagnetic activity, the robot immediately retreats to its proven safe distance (exploitation of known-safe behavior) while continuing passive observation until it can model this new behavior pattern.
Through these integrated processes, the robot builds a comprehensive understanding of this alien intelligence without risking mission failure through reckless interaction.
JS: Let’s transfer this learning model to a modern dating scenario in which a young man meets a young woman, for the first time, at an agreed-upon coffee shop. Illustrate how he applies these five rules to assessing the woman for a possible follow-up date.
Claude: Here’s how the five learning rules would manifest as the young man assesses his coffee date:
Rule 1: Prediction and Error Correction: Based on her dating profile photos, he predicts she’ll be reserved and soft-spoken. But when she arrives, she greets him with a firm handshake and confident eye contact, immediately violating his “shy and demure” prediction. When he mentions his job in finance, he expects interest or impressed questions, but instead she pivots to asking about his weekend rock climbing photos. His algorithms rapidly adjust: “Not impressed by status markers – values adventure and authenticity.”
Rule 2: Pattern Compression: He’s processing countless micro-signals: vocal tone, laugh frequency, posture shifts, eye contact patterns, response timing, coffee choice, phone checking behavior. His mind compresses these into core patterns: “Genuine laughter + leaning forward during stories + asks follow-up questions + makes jokes at own expense = authentically engaged and self-confident person.” He filters out irrelevant details like her specific coffee order while encoding essential behavioral signatures.
Rule 3: Multi-Scale Attention: His attention shifts dynamically – focusing intently on her facial expressions when she talks about her work passion, then broadening to notice how she interacts with the barista (polite, patient), then narrowing to her hands when she gestures animatedly about travel stories. When she mentions her ex briefly, his attention laser-focuses on tone and context clues before returning to broader conversational flow.
Rule 4: Hierarchical Abstraction: Low level: “Smiles when talking about her sister + voice gets softer.” Mid level: “Family-oriented with strong relationships.” Higher level: “Values deep personal connections.” Abstract level: “Someone who prioritizes meaningful relationships over superficial ones.”
Rule 5: Adaptive Exploration vs Exploitation: He starts with safe conversational territory (work, hobbies) while observing her responses. As she proves receptive to humor, he ventures more playful jokes and personal anecdotes. When she responds positively to his story about helping his elderly neighbor, he explores deeper values-based topics. But when he tests a slightly edgy political comment and notices her discomfort, he quickly reverts to safer ground while mentally noting “avoid controversial topics until better calibrated.”
By date’s end, he’s built a nuanced model of compatibility potential without explicitly “interviewing” her – the learning rules operated seamlessly beneath natural conversation.
JS: Claude, can you breathe life into this coffee-shop sketch by creating an entertaining and warm short story between 3,000 and 5,000 words?
Claude: I’ve created an entertaining short story that brings the coffee shop dating scenario to life! The story follows Marcus, a data scientist who approaches dating like a machine learning problem, as he meets Elena, an environmental scientist who challenges all his algorithmic assumptions.
The narrative weaves in all five learning rules we discussed – his prediction errors, pattern compression, shifting attention, building hierarchical understanding of who she really is, and balancing exploration with exploitation as the evening progresses. But more than that, it explores the beautiful messiness of human connection that can’t be reduced to algorithms.
At about 4,200 words, it captures that warm, entertaining feel you requested while showing how genuine chemistry emerges when people drop their carefully optimized personas and just connect authentically. Would you like me to adjust anything about the tone, pacing, or character development?
JS: Read the story, “Algorithm of an Intentional Heart.”
Filed under: Uncategorized |
Educational Technology and Change Journal 
See Stephen Downes’s commentary of this article on Stephen Downes (9/2/25): I’m not sure I would create the learning model exactly the way it is done here, but I appreciate the strategy and I think it’s founded in some good thinking about how cognition occurs. Where I begin to drift apart from the authors is in steps four and five where despite pledging to avoid “the confusing backdrop of countless learning and developmental theories,” they hypothesize ‘layered representations’ and ‘adaptive exploration’. Also, I wish they had managed to avoid collapsing completely into stereotype, but no, they had to outline the story of “a data scientist who approaches dating like a machine learning problem, as he meets Elena, an environmental scientist who challenges all his algorithmic assumptions.” -js