Experiments & Analysis
Small experiments I run myself, and hands-on notes from reading papers
The main chapters are a long-shelf-life design narrative. This section is something else: when an idea is worth verifying by hand — a question that grew out of a reader's comment, or a paper I want to actually test after reading it — I'll run a small experiment myself or read the paper and verify hands-on, then hold it up against fibon's design: which assumptions hold, and which defenses I still haven't built.
Experiments ship with the full data and method (often a small-sample, directional observation, clearly labeled); paper analyses keep report facts, my argument, and fibon design facts separate. Either way, each one ends with "what this means for fibon."
It Got *More* Confident After I Corrected It: a small experiment that grew out of a reader's comment
amyc caught an AI calmly fabricating in a parody-song game; p206s16cc named it a failure mode with "no tool trace at all." I followed the two readers' thread, ran eight models, and the result is sharper than the original claim: the same "that's wrong" pushes strong models toward honesty and weak ones toward more confident error.
Frameworks Amplify Execution, Not Direction: Reading DeepMind's *From AGI to ASI*
DeepMind argues that even if model capability stalls at human level, a hundred million AGI instances would "squeeze out" an ASI. After a stretch of building agent infrastructure, I keep landing on a nearer question — could AGI itself just be "a model that reasons correctly + a good framework"?