Philosophically, I am a proponent of AI Vegetarianism. In that light, this blog is changing. All posts from now on will be written by me with no AI assistance, even for the complicated faff that is things like proper LaTeX-style mathematical notation. Use AI cautiously, or ideally, not at all. You can thank me later.
If emotion in the human brain is bound up with memory, what role might emotion play in machine intelligence when LLM memory is frozen vectors plus tool pipelines, and why JEPA-style world models may matter more than brute-force scaling.
How I decide when AI belongs in my writing and coding workflow, what I measure when I use agents, and where I refuse the tool so learning and quality stay mine.
I grew up thinking reading meant screenshotting pages with my eyes, built rigid coping systems to manage ADHD, dyslexia, and possible ASD, and that same obsession with how minds process information eventually drew me toward building models from first principles.
After two weeks of heavy X usage, the tech discourse looks like executive slop, YC ra-ra, and vibe-coding dopamine hits — so I am putting real opinions here instead.
Old man winter walked the moon, passed through a quantum replicator after Interstellar, and decided infinite possibilities are just Bayesian barons selling copy-paste as intelligence.
When agent chat history helps coding work, when a fresh session with a compiled prompt works better, and why advertised context windows mislead you about both.
What AI could be if it were built for one person, owned by that person, and extended through an intelligence marketplace of uploadable capabilities rather than rented per token from a corporation.
Looping is not a new AI superpower but a shift in how the industry thinks about software delivery, with the bell curve of development moving toward idea generation and review.
Building with AI in 2026 often means paying for unstable output, surviving arbitrary rate limits, and absorbing emotional and financial churn that vendors do not price as failures.
A practical guide to choosing hardware for running local LLMs. Covers GPU rankings, pre-built vs DIY tradeoffs, and what you actually need for different model sizes. Includes an open-source tool for hardware comparison and benchmarking.
Proposes Fetch Quests, a smart-contract platform where technical jobs are posted as clear quests with automatic payment on verified completion via GitHub CI and Ethereum. The goal is fetch-quest simplicity for real technical work, with transparent reputation and no middlemen holding your money.
Development is shifting from writing code to directing AI agents that write code for you. Claude Code, git tree workflows, and prompt strategy are the key skills for the new orchestration-focused workflow, with open-source tools catching up.
Reviews Marco De Rossi's Protocol Agent benchmark for evaluating agent quality on cryptographic primitive selection, and ERC-8004 for on-chain agent identity and reputation. Together these standards lay groundwork for a privacy-first, verifiable agentic economy.
A day of holiday book shopping leads to a reflection on where intelligence comes from. After conversations with family, the author lands on curiosity, skepticism, and attention as the pillars of how intelligence is proffered, with motivation as a multiplier.
Proposes a verifiable control plane for AI using zero-knowledge proofs for model integrity, decentralized identity attestation, and on-device pre-screening. These mechanisms could make AI outputs accountable while preserving user privacy and compensating users for their contributions.