Posted inIdeas for AI
GPT-5.6 just made itself CHEAPER
Idea #1Price per token is a misleading metric; cost per completed task is the only number that matters when choosing an LLM provider. A model that charges half the token…
Posted inBlog
Somebody Else’s Definition of Done: How to Actually Evaluate an Agent Skill
A writing skill was installed to help with design decisions — color palettes, layout choices, that kind of thing. It activated correctly. The output was technically fine. But every suggestion…
Posted inIdeas for AI
The Ultimate Knowledge Base: Bring YouTube Into Your AI Second Brain
Idea #1Building a deep knowledge base from a large corpus is inherently token-heavy and cannot be done cheaply — but the one-time cost is justified by the permanent navigation dividend.…
Posted inBlog
What I Learned Building a Tool That Cut My AI Token Use by 90 Percent
For a long time, the AI token bill was understood as a function of what got typed. Long prompts cost more; short ones cost less. That turned out to be…
Posted inBlog
What a benchmark score of 81 actually tells you about model selection
The same benchmark suite has been in use for a while now. When Opus 4.8 came back with a score of 81 — comfortably ahead of every other model tested…
Posted inBlog
The platform team bottleneck nobody is measuring
There is a gap opening up inside most engineering organizations, and it is not showing up on anyone's productivity dashboard. App teams are shipping faster. AI agents are handling whole…
Posted inBlog
Dark code: the comprehension gap that AI-assisted development quietly opened
The first serious encounter with AI coding assistants tends to center on correctness — will the generated code actually work? That turned out to be almost the wrong question. The…
Posted inBlog
The two career collapses happening at once — and the one mental model that bridges both
A few years ago, the career advice circulating — given and received — assumed that domain expertise was a durable asset. Learn finance deeply, or law, or marketing, and that…
Posted inBlog
The Jagged Frontier Was a Measurement Error — Draft: The Surgeon-in-the-Dark Read
Shaping the flow of ideas at the frontier of cognitive systems. --- Cursor's coding assistant — a tool built to help programmers write software — recently solved Problem 6 of…
Posted inBlog
Stop Guessing. Here Is How to Actually Test a Cheaper Model.
For a long time, the way most of us evaluated a cheaper model was roughly this: run a few prompts, compare the outputs to what we were used to, and…