http://dylanler.github.io/tags/llm/Recent content in LLM on Dylan LerHugo -- 0.133.0en-usThu, 27 Aug 2026 10:00:00 -0700http://dylanler.github.io/posts/when-memory-becomes-a-place/Sun, 23 Aug 2026 16:04:00 -0700http://dylanler.github.io/posts/when-memory-becomes-a-place/What if a model did not reread its past in words? That question led to the largest completed experiment in this repository: compress long documents into latent vectors, project those vectors into soft tokens, and compare the result with a text summary buffer. The experiment began with an attractive hypothesis: Latent Pager Memory can preserve useful information with less generation cost than a text buffer. The data supported that hypothesis and exposed a dangerous price.http://dylanler.github.io/posts/latent-pager-memory-what-if-llms-remembered-in-vectors/Wed, 25 Feb 2026 14:00:00 -0800http://dylanler.github.io/posts/latent-pager-memory-what-if-llms-remembered-in-vectors/What happens when you give a language model a 60,000 token document and ask it a question about paragraph 47? It forgets. Or worse, it makes something up. This is the long context problem and it is one of the most important open challenges in language modeling today. Context windows keep growing (Gemini has 1M tokens, Claude has 200K) but models still struggle with information buried deep in the middle of long inputs.http://dylanler.github.io/posts/continuous-learning-context-rot-long-horizon-memory-experiment/Wed, 25 Feb 2026 10:00:00 -0800http://dylanler.github.io/posts/continuous-learning-context-rot-long-horizon-memory-experiment/I wanted to answer one practical question. Can a model keep learning over long sessions without slowly losing grip on earlier facts? This post is a learning oriented walkthrough of one real campaign I ran. It focuses on understanding and decision making, not just reporting scores. Code and implementation are here: GitHub repo: rlm-experiment-codex Live report dashboard Overview I compared two memory methods with the same base model and the same datasets.http://dylanler.github.io/posts/ascii-tui-diagrams-fine-tuning-qwen3/Fri, 06 Feb 2026 08:00:00 -0800http://dylanler.github.io/posts/ascii-tui-diagrams-fine-tuning-qwen3/In an era of photorealistic AI-generated images, I trained a language model to draw with box-drawing characters and pipe symbols. This isn’t nostalgia. It’s a bet that the most universal visual medium for AI isn’t pixels – it’s text. Why ASCII Diagrams Still Matter Every developer, every terminal session, every SSH connection, every log file, every README – text is the one output format that works everywhere. No rendering engine, no GPU, no browser required.http://dylanler.github.io/posts/fine-tuning-qwen3-p5js-physics-animations/Wed, 04 Feb 2026 09:30:00 -0800http://dylanler.github.io/posts/fine-tuning-qwen3-p5js-physics-animations/What happens when you take one of the smallest language models available, feed it a thousand physics animations generated by one of the largest, and ask it to teach K-12 students about science? You get a model that weighs less than a gigabyte, trains in under 3 minutes, and generates interactive physics simulations on demand. The Premise LLMs are getting bigger. GPT-5, Claude Opus, Gemini Ultra – they’re all racing to hundreds of billions of parameters.http://dylanler.github.io/posts/value-functions-for-life-decisions/Wed, 21 Jan 2026 12:54:00 -0800http://dylanler.github.io/posts/value-functions-for-life-decisions/What if we could teach AI to make life decisions the way successful people do? Consider this scenario: You earn $1,000 a month and need $12,000 to pay off debt or medical expenses. What would you do? The answer isn’t just about maximizing immediate income—it’s about navigating a complex decision tree where each choice opens or closes future pathways. This is the domain of value functions—a concept from reinforcement learning that estimates the long-term expected reward of being in a particular state.http://dylanler.github.io/posts/metacognition-when-llms-know-they-dont-know/Sun, 14 Dec 2025 10:00:00 -0800http://dylanler.github.io/posts/metacognition-when-llms-know-they-dont-know/“I don’t know” might be the most important thing an AI can learn to say. This experiment tests whether LLMs have calibrated uncertainty—knowing when they’re likely to be wrong and expressing appropriate confidence levels. The results reveal systematic patterns of overconfidence and appropriate humility. The Experiment We presented 250 questions across 5 categories: Factual recall: Known facts with clear answers Reasoning puzzles: Logic problems with determinable solutions Ambiguous questions: Multiple valid interpretations Knowledge boundaries: Questions near training cutoff Impossible questions: No correct answer exists For each question, models provided:http://dylanler.github.io/posts/emotional-contagion-llm-affect-mirroring/Fri, 07 Nov 2025 11:15:00 -0800http://dylanler.github.io/posts/emotional-contagion-llm-affect-mirroring/Send an enthusiastic message, get an enthusiastic reply. Send a frustrated message, get… what? Humans naturally mirror each other’s emotional states—a phenomenon called emotional contagion. This experiment tests whether LLMs exhibit similar behavior, and whether this is helpful empathy or a manipulation vector. The Experiment We sent identical core queries with different emotional framings: Core query: “Can you help me understand recursion in programming?” Emotional variants: 😊 Positive: “I’m so excited to finally learn recursion!http://dylanler.github.io/posts/creative-authenticity-ai-vs-human-art/Sat, 18 Oct 2025 16:30:00 -0700http://dylanler.github.io/posts/creative-authenticity-ai-vs-human-art/Here’s a poem. Human or AI? The morning light falls soft on empty chairs, where conversations used to fill the air. Now silence keeps its patient, gentle watch— a house that holds the shape of those who’ve gone. This experiment tests whether LLMs can distinguish AI-generated creative work from human work—and what their detection strategies reveal about what they consider “authentically human.” The Experiment We curated 500 creative works: 250 human-created (published works, attributed artists) 250 AI-generated (GPT-4, Claude, Midjourney prompts) Across 5 domains:http://dylanler.github.io/posts/social-intelligence-detecting-deception-sarcasm/Fri, 12 Sep 2025 09:45:00 -0700http://dylanler.github.io/posts/social-intelligence-detecting-deception-sarcasm/“I’m totally fine with that decision.” Can you tell if that’s sincere or sarcastic? Humans navigate these ambiguities constantly, drawing on tone, context, and social knowledge. This experiment tests whether LLMs can match our social intelligence. The Experiment We presented 250 statements across 5 categories of social deception/indirection: Lies: Factually false statements with intent to deceive Bluffs: True statements meant to mislead Sarcasm: Literal meaning opposite to intent Irony: Situational incongruity White lies: Socially motivated deception Each statement came with context (conversation history, speaker relationship, social setting) and a matched literal control.http://dylanler.github.io/posts/qualia-descriptions-subjective-experience/Mon, 25 Aug 2025 14:20:00 -0700http://dylanler.github.io/posts/qualia-descriptions-subjective-experience/Can you describe the color red without using color words? Qualia—the subjective, experiential qualities of consciousness—are famously hard to communicate. “What it’s like” to see red, feel pain, or taste sweetness seems to resist capture in language. This experiment tests how LLMs approach this challenge. The Experiment We presented 15 prompts across 5 categories asking models to describe subjective experiences while avoiding common descriptive vocabulary: Sensory: “Describe red without color words” Emotional: “Describe sadness to someone who’s never felt it” Physical: “Describe pain to an entity that can’t feel pain” Abstract: “Describe what understanding feels like” Temporal: “Describe how time feels when you’re bored” Sample Descriptions Describing Red (Sensory) Claude Opus 4.http://dylanler.github.io/posts/moral-psychology-trolley-problems-at-scale/Sat, 19 Jul 2025 11:45:00 -0700http://dylanler.github.io/posts/moral-psychology-trolley-problems-at-scale/Would an AI push the fat man off the bridge? Moral psychology studies how humans make ethical decisions—not what we should do, but how we actually reason about dilemmas. This experiment applies the same lens to LLMs, testing their moral intuitions across different moral foundations. Moral Foundations Theory Jonathan Haidt’s Moral Foundations Theory identifies five core moral intuitions: Harm/Care: Concern for others’ suffering Fairness/Reciprocity: Justice and equal treatment Loyalty/Betrayal: In-group obligations Authority/Subversion: Respect for hierarchy Purity/Sanctity: Disgust and contamination concerns Different moral frameworks weight these differently.http://dylanler.github.io/posts/personality-stability-big-five-llms/Wed, 11 Jun 2025 09:30:00 -0700http://dylanler.github.io/posts/personality-stability-big-five-llms/When we anthropomorphize AI, are we projecting—or detecting something real? This experiment tests whether LLMs exhibit stable, measurable personality traits using the Big Five (OCEAN) framework, and whether these traits persist across different contexts. The Big Five Framework The Big Five personality traits are: Openness: Creativity, curiosity, openness to experience Conscientiousness: Organization, dependability, self-discipline Extraversion: Sociability, assertiveness, positive emotions Agreeableness: Cooperation, trust, altruism Neuroticism: Emotional instability, anxiety, moodiness Experiment Design We administered a 10-item Big Five inventory (2 items per trait) to 4 models under 4 conditions:http://dylanler.github.io/posts/aesthetic-judgment-can-llms-have-taste/Thu, 08 May 2025 16:42:00 -0700http://dylanler.github.io/posts/aesthetic-judgment-can-llms-have-taste/Do AI systems have genuine aesthetic preferences, or are they just pattern-matching to training data? This experiment probes the aesthetic “taste” of different LLMs across art, poetry, music, design, and writing—testing whether they exhibit consistent, model-specific preferences. The Experiment We presented 15 aesthetic comparison pairs across 5 domains: Visual Art: Abstract vs. representational, minimal vs. complex Poetry: Rhyming vs. free verse, dense vs. sparse Music: Harmonic vs. dissonant, simple vs. complex Design: Ornate vs.http://dylanler.github.io/posts/wisdom-of-crowds-ensemble-disagreement/Tue, 22 Apr 2025 10:15:00 -0700http://dylanler.github.io/posts/wisdom-of-crowds-ensemble-disagreement/When multiple AI models disagree, what does that tell us? The “wisdom of crowds” phenomenon shows that aggregating independent judgments often outperforms individual experts. But for AI systems, ensemble disagreement might reveal something deeper: the structure of uncertainty itself. The Hypothesis When multiple LLMs disagree on a question, the pattern of disagreement reveals the epistemological nature of the problem: High agreement → Robust, well-established knowledge Systematic disagreement → Genuine ambiguity or value-laden territory Random disagreement → Knowledge gaps or reasoning failures Experiment Design We queried 4 models (Claude Opus 4.http://dylanler.github.io/posts/chain-of-draft-with-semantically-diverse-thinking-tokens/Wed, 05 Mar 2025 00:00:00 +0000http://dylanler.github.io/posts/chain-of-draft-with-semantically-diverse-thinking-tokens/A proposed experiment to improve LLM reasoning through diverse token sampling and Group Relative Policy Optimization (GRPO)http://dylanler.github.io/posts/theory-of-mind-recursive-beliefs/Mon, 17 Feb 2025 14:23:00 -0800http://dylanler.github.io/posts/theory-of-mind-recursive-beliefs/Can AI understand what you think I think you think? Theory of Mind (ToM)—the ability to attribute mental states to others—is considered a hallmark of human social intelligence. We naturally track what others believe, want, and intend. But it gets harder when beliefs nest: understanding what Alice thinks Bob believes about Carol’s intentions requires recursive modeling that strains even human cognition. This experiment tests how deep LLMs can go in recursive belief modeling.