Open Source Just Passed Frontier: What GLM-5.1 Means for Builders
GLM-5.1 topped SWE-Bench Pro under MIT license. Gemma 4 shipped edge variants under Apache 2.0. Open source model parity is here — with caveats builders need to understand.
Blog
The $650B Zero-ROI Disconnect: AI's Biggest Bet vs the Data
Amazon, Meta, Google, and Microsoft are spending $650B on AI infrastructure. Goldman Sachs says the GDP contribution is basically zero. Here's what the data actually shows.
Inside Foliome: How the Agent Actually Works
The technical companion to Foliome's release. How the agent builds its own bank integrations, orchestrates parallel sync with MFA routing, recovers from errors without human intervention, and serves a dashboard from your local machine.
Foliome: Your Money, Your Machine, Your Agent
I haven't logged into a bank app in three months. An open-source financial operating system where your AI agent syncs your banks, classifies spending, and gives you complete financial intelligence. Everything local. Everything yours.
The Keynesian Folly: Why AI Will Never Fully Automate Finance
Financial markets are reflexive. AI models degrade the patterns they exploit. The CFA Institute says augmentation, not automation.
AI Platform Divergence: What It Means for Builders
OpenAI is betting on agents, Anthropic on coding, Google on multimodal. Here's the evidence, the strategy layer, and what to do if you're building in between.
DPO Fine-Tuning a 1.2B Model: What Worked, What Broke, What I'd Skip
DPO fine-tuning a 1.2B model with LoRA improved style but degraded reasoning after 400 samples — accuracy dropped from 39/40 to 34/40, and the base model won.
Multi-Model Routing: Instruct vs Thinking on Edge Devices
How I route between two LFM2.5-1.2B models on a Raspberry Pi 5 — when to use Instruct, when to use Thinking, and why user-controlled toggles beat automatic detection.
The Frontier Model Tax
Airbnb doesn't run OpenAI in production. Cursor built its own Tab model. The shift to fine-tuned small models is real — here's the decision framework for builders.
Agent AI Execution: Why Unix Shell Beats JSON Function Calling
Building agents with a single shell execution tool reduces formatting errors and improves autonomy by leveraging LLM training data instead of brittle JSON schemas.
M5 Max and Local LLMs: Why the Best AI Tools Will Never Be Open Sourced
M5 Max hardware enables 120B parameter models to run locally at 65.8 tokens per second, shifting developer focus from public open-source projects to private, hyper-personalized dark tools for maximum leverage.
The 2026 Solo Builder AI Stack: Picking Tools for Optionality, Not Features
Every solo builder is sharing their AI stack. Nobody explains why they chose it. The difference between a stack and a dependency trap.
The AI Displacement Report Card: What the Data Actually Shows
WEF predicted 85M AI job losses by 2025. Actual US figure: 200K-300K. Oil, tariffs, and DOGE explain more than AI does.
The Consultant-to-Builder Pipeline: How to Learn Technical Skills by Solving Problems
AI displaced 200K-300K jobs while employment grew 2.5%. The opportunity isn't learning to code. It's solving a $1,000 problem in your domain with AI tools.
Text-to-Data with Snowflake Cortex: Building a Natural Language Portal
Build a secure text-to-data portal using Snowflake Cortex and Llama 3.1 to translate English into Pandas while maintaining a strict Python sandbox.
Gemini Agentic Workflows: Solving Model Laziness with Skill Graphs
Reduce Gemini agent shortcuts and memory loss by replacing flat tool lists with a file-system Skill Graph for structured, stateful workflows.
The Entry Point Is Closing
Anthropic's own labor data shows AI isn't eliminating white-collar jobs yet — it's quietly shrinking entry-level hiring. Here's what that means if you're making the move into tech right now.
The End of the Context Window War: How Recursive Language Models Solve Infinite Data
Recursive Language Models replace massive context windows with a programmatic loop, letting small local models process infinitely long prompts.
Extracting an Obsidian Plugin From a Note App
I took the voice transcription pipeline out of my note app and shipped it as an Obsidian plugin. No platform. No local server. Just the slice that worked.
AI Agent Architecture: Why Single Prompt Files Fail at Scale
Why single prompt files fail for AI agents—and how failure-driven codification built a 108,000-line C# system with 19 specialized agents.
Agent Memory Architecture: Why Documentation is Agent RAM
How to manage a 24% knowledge-to-code ratio using a 3-tier context architecture—Hot, Warm, and Cold storage for AI agents.
AI Context Coverage: The 25% Rule for Large Codebases
How a 24% knowledge-to-code ratio prevents AI hallucinations and architectural rot in 100k+ line repositories.
AI Agent Architecture: How to Build 'Identity' Without Breaking Performance
Stop using 2,000-word system prompts. Learn how Tiered Context Architecture, Style Injection, and the Observer Pattern optimize AI agent performance and cost.
OpenClaw: Framework or File Bloat? What to Reuse (and Skip)
An evaluation of the OpenClaw gateway, the 'soul.md' ecosystem, and why separating Skills from Souls is the key to reliable agents.
Recursive Language Models: How to Process Infinite Context With an 8B Model
RLMs let small models handle inputs 100x beyond their context window by storing prompts in a REPL and recursively processing slices. Here's how it works.
The LLM Parameter Lie: What Actually Matters in 2026
Parameter counts are a vanity metric. How to read LLM architectures, what active parameters mean for your hardware, and the benchmarks that actually matter.
The 2026 AI Stack: Building Compound Systems and Agentic Workflows
A strategic blueprint for the modern AI stack: moving from simple LLM wrappers to model routing, multi-agent orchestration, and local-first memory.
ChatGPT Ads: What Changes When the Search Box Talks Back
ChatGPT Ads target conversation context, not keywords. Here's what OpenAI confirmed, what's speculated, and how GEO content prepares you for both.
Custom Blocks in Obsidian: The Scribe Block
Obsidian doesn't have custom blocks, so I faked it with a code fence, a hidden div, and two features that actually make it useful.
Lazy Engineering: Why GraphRAG Beats 1M Token Context Windows
Large context windows introduce massive KV cache costs and latency. GraphRAG provides better retrieval accuracy for complex datasets at a fraction of the compute expense.
RTX 5090 Local AI ROI: GPU Cluster vs. API TCO Analysis
An RTX 5090 cluster breaks even against DeepSeek API costs at 450M tokens of inference, providing a 14-month ROI for high-volume developers.
System 2 Distillation: Porting DeepSeek-R1 Reasoning via Probability Mapping
Distill DeepSeek-R1 reasoning into small models for $0.50 using logit-based probability mapping. A high-signal alternative to standard fine-tuning.
Qwen 3.5 vs Liquid LFM: Edge AI Benchmarks on Raspberry Pi 5
Comparing Qwen 3.5-2B and Liquid LFM2.5-1.2B on Raspberry Pi 5: benchmarks for GPQA intelligence, RAM usage, and time-to-first-token latency.
Voice Transcription Market: Current State and the Missing Piece
Voice transcription in 2026: who ships what (Monologue, Wispr Flow, Hey Lemon), what builders use (Whisper, AssemblyAI), and why fragmentation and lock-in are the real problem.
DIY Local Voice Assistant: Building a Raspberry Pi 5 AI Device
How to build a private, fully local voice assistant on Raspberry Pi 5 using Python and Whisper for on-device speech processing and LLM inference.
Launch Platform ROI: Data on Product Hunt, HN, and AppSumo
Actual traffic and conversion data for Product Hunt, Hacker News, and AppSumo — including the 10% featured rate and specific ROI math for indie makers.
LFM2.5-1.2B on Raspberry Pi 5: llama.cpp Optimization Guide
Optimizing llama.cpp for LFM2.5-1.2B on a Raspberry Pi 5. Recommended settings for quantization, threads, and KV cache to maximize local LLM performance.
Raspberry Pi 5 Local Voice Assistant: A Whisper and Liquid LFM AI Stack
Build a private AI voice assistant on Raspberry Pi 5 with Whisper and Liquid LFM2.5. Includes memory budgets, hardware setup, and Python code.
Deep Research: Exa.ai vs Other Providers
Embedding deep research in a note app with Exa.ai. I built a 5-phase enrichment pipeline, then simplified to Exa with a light review step.
GEO: The New Rules of Search Visibility
AI search engines don't rank pages — they cite sources. Here's what actually works for getting cited by ChatGPT, Perplexity, Claude, and Google AI Overview.
Handling Chain-of-Thought for GPT-OSS-120b: Harmony Format and Provider Quirks
GPT-OSS-120b uses OpenAI's Harmony token protocol — analysis, commentary, and final channels. Here's how we handle CoT leakage and provider routing.
Workspace vs Document Assistant: Why You Need Both
The best work happens when conversation stays continuous. Document-localized chat gives the workspace assistant richer context — here's how we architected it.
LFM2.5-1.2B vs LFM2-2.6B: Why We Chose the Smaller Model
Benchmarking Liquid AI's LFM2.5-1.2B against LFM2-2.6B on a Pi 5 — the smaller model scores higher on IFEval (+9), runs 2.3x faster, and fits in under 1GB.
Why We Stopped Forcing One Model to Do Everything
We built a Reasoner-Planner-Solver pipeline for a Pi 5 voice assistant, then replaced it with Instruct/Thinking model routing. Here's why simpler won.
Ditching Exa AI for Self-Hosted Search
Replacing Exa AI with self-hosted SearXNG and trafilatura on a Raspberry Pi — fully local web search in 2-3 seconds, no API keys, no data leaving the device.
From 78% to 97%: Making a 1.2B Model Actually Work
Three tricks that took a 1.2B model's tool routing from 78% to 97% — renaming tools, adding a calibration line, and a regex post-processor.
One Model, Three Roles: How a 1.2B Model Plays Reasoner, Planner, and Solver
One 1.2B model plays Reasoner, Planner, and Solver with different system prompts on a Raspberry Pi 5. Three LLM calls, 15-30 seconds for a reasoning task.
When to Stop Adding Rules: Building a Negation-Aware Sanitizer Without Overfitting
I kept adding regex patterns to fix edge cases until I realized I was overfitting a rule-based system — here's the framework for knowing when to stop
Thinking Without Thinking Models
How I got multi-step reasoning working on a Raspberry Pi 5 with a 1.2B model — no thinking model needed. ReWOO cuts it to 2 LLM calls and 80% fewer tokens.
The SKILLS.md Standard: Why the Agentic Web Needs a Universal Resume
Stop writing massive Python files to tell your AI agents what they can do. We need a portable, plain-text standard for agent capabilities.
The Instruction Gap: When AI Starts Building Its Own Systems
We want AI to work autonomously, but letting it write its own playbook creates a new kind of technical debt.