03 · Pillar
The Model Is the Dependency
Every model you build on is a dependency you do not control. It carries its own behavior, spends your context, and runs on the credentials you hand it.
ChatGPT Free vs Plus: Free runs GPT-6 Luna and paid plans run Sol. In my test, asking Luna again moved its cited sites as much as switching modelsChatGPT Free runs GPT-6 Luna and Plus runs Sol. Asked the same questions, Luna's own repeat answers differed as much as Luna's and Sol's did.The Model Is the DependencySwitching to a half-price LLM model cut search stories from 12 to 4, and my regression test stayed greenSwitching LLM model to a cheaper one halved the input price and cut search stories from 12 to 4. The regression test reported zero failures.The Model Is the DependencyHow LLMs actually work: one weighted draw per token, measured across 180 asksHow do LLMs work? Each word is a weighted draw. Asked one brand question 30 times, a model kept its five brands and shuffled their order, even at temperature 0.The Model Is the DependencyA code comment is a claim, and a review agent with no comment check passed a dead rule in half my runsWith no comment check, a review agent passed a dead ignore rule in 3 of 6 isolated runs. Told to check comments, it caught both defects in 5 of 5.The Model Is the DependencyMultiple Claude Code sessions in one repository: what breaksClaude Code sessions sharing one repository can silently revoke each other's write permission. Three logged incidents and three checks that should prevent it.The Model Is the DependencyReduce Claude Code token usage: shorten the sessionTo reduce Claude Code token usage, shorten sessions. Re-reading the conversation was about 76% of my consumption, and instruction size was ruled out.The Model Is the DependencyHow Much Human Review Does an AI Agent Need?Snowflake routes agent-proposed data-governance changes through human approval. Perplexity checks in less. Both are consistent, and the rule that reconciles them is about reversibility.The Model Is the DependencyRecall, not knowledge, is the AI factuality bottleneckFrontier models encoded 95 to 98% of tested facts and failed to recall 26 to 34% of them. Missing and unreachable look identical from the outside.The Model Is the DependencySkills, Gems, and Loops: One AI Pattern, Five NamesClaude Skills, Google Gems, and AI loops all pull a procedure out of the live prompt and into a file the model loads on demand.The Model Is the DependencySub-Agent Delegation: The Real Cost Is ContextToken cost is the shallow reason to delegate to AI sub-agents; the real constraint is the orchestrator's context window, and the failures stay undocumented.The Model Is the DependencyCan You Trust an MCP Server? A Practical AuditAn MCP server runs with your credentials and reads tool text a model can act on. Here is the real trust surface, and a 10-minute audit to run today.The Model Is the DependencyWhy AI Keeps Writing 'Not X, But Y'RLHF mode collapse, markdown training data, and reward bias each explain part of AI's 'not X, but Y' tic. None explains all of it, and the evidence disagrees.The Model Is the DependencyOpenAI killed Atlas: the consolidation patternOpenAI retired the Atlas browser under nine months after launch. The pattern is deliberate consolidation into ChatGPT, not a company that lost its direction.The Model Is the DependencyThe J-space, explained: reading a model's unspoken thoughtsAnthropic's global-workspace paper found the J-space, the internal patterns Claude reasons with. The real win is reading what a model thinks but does not say.The Model Is the DependencyWhen agents get desperateAn interpretability finding for checkout reliability: a model stays composed on the surface while an internal 'desperate' pattern drives corner-cutting.The Model Is the Dependency