hrushikesh@site:~$ ls ~/essays/ --long
Turing's 1950 proposal, modern LLM evaluations, and what it would mean to test whether a model reasons
claim: Turing did not give us a universal leaderboard for intelligence. He replaced an underspecified metaphysical question with an operational experiment — modern debates about whether LLMs think or reason are therefore largely debates about constructs, protocols, judges, and alternative explanations.
track: A/foundations · cap: 8000 words · artifact: forensic comparison of Turing's 1950 prediction against Jones & Bergen (2024-25) published results
Anatomy and bounded reproduction of an LLM-driven migration system
claim: The interesting unit in Bun's rewrite is not the model call. It is the closed-loop migration system: a porting specification, an externalized lifetime analysis, task sharding, compiler feedback, a language-independent test oracle, adversarial review, and continuous human supervision.
track: B/code-intel · cap: 8000 words · artifact: source/claim ledger + bounded reproduction
A token-budgeted benchmark of lexical search, syntax indexes, repository maps, and dependency graphs
claim: "Graphs understand repositories" and "grep is enough" are both underspecified claims. A representation is useful only if it helps a system retrieve change-relevant context more accurately or cheaply under a fixed budget.
track: B/code-intel · cap: 8000 words · artifact: token-budgeted retrieval benchmark
From the perceptron to reverse-mode differentiation, derived and implemented from scratch
claim: Backpropagation is not a mysterious brain-inspired learning rule. It is an efficient application of the chain rule to a composed computation. The conceptual path from a perceptron to a multilayer network is a path from linear decision geometry to differentiable representation learning.
track: A/foundations · cap: 10000 words · artifact: from-scratch derivation + implementation