Farm Fresh Insights

Sign in Subscribe

Latest — 08 Mar 2026

Comparative Advantage: Calibrated Pairwise Win Rates from LLM Judge Scores

Causal Judge Evaluation (CJE; Landesberg, 2025) formalizes a workflow that is already widespread in practice: score every response with a cheap LLM judge, label a small slice with a trusted oracle, learn the judge-to-oracle mapping, and evaluate at scale with calibrated uncertainty. However, calibration is only as informative as the

More issues

Liberation and Immanence

In the Indian religions of liberation, the highest good is release. Moksha means release from samsara, the cycle of death and rebirth. The Muktika Upanishad distinguishes several grades of liberation and places one above the rest: videhamukti, disembodied liberation. Buddhism reaches a similar conclusion through different metaphysics. Nirvana is the

Simulating and Evaluating Agentic Systems

Most teams building agentic systems know they need some way to test multi-turn conversations. The agent interprets ambiguous input, picks tools, maintains state, and has to land in the right place at the end. You can't eyeball that. So people reach for what they already know. The most

Anekantavad

When I wrote my short missive on motivated citations nine years ago, I didn't ship with examples because I wanted to leave those to a paper. I never wrote the paper. But each time I review a "literature," I rue the missed opportunity to highlight this

Hardware‑Attested Ad Insertion for Encrypted Video

Traditional video monetization assumes the server can stitch ads into the stream. In an E2EE messenger like WhatsApp, the server never sees plaintext by design. This essay lays out a path to extend monetization to some person‑to‑person video inventory without moving decryption server‑side by combining: * on‑device

Out-of-Context: Constrained Tool Based Exploration of Context

Longer context windows have not come at the back of a solution for "context rot." Recursive Language Models (RLMs) take inspiration from out‑of‑core data processing, where systems handle datasets larger than memory by deciding what to load. The paper proposes solving the context rot problem by

When the Fix Is Another Distortion: Quotas as Second-Best Engineering

Chattopadhyay and Duflo (2004) (hereafter C&D) contains a formal model that builds a world where a striking normative implication becomes possible: “welfare” (in the paper’s narrow sense) can rise when you ban half the people from running for office. The paper says this pretty directly: “...reservation may

First, Assume No Elections

Dixit, Grossman, and Gul (2000), hereafter DGG, study a setting in which two groups repeatedly divide a fixed surplus over an infinite horizon. The basic setup is as follows: * Two groups, $A$ and $B$, each with concave utility over their share of a pie normalized to 1. * In each period,

Targeted Model Editing

Pretrained generative models (LLMs, text-to-image, etc.) are used as general-purpose engines and then "aligned" toward specific goals: lower toxicity, higher truthfulness, removal of certain objects or styles, domain specialization, and so on. In practice, we rarely want to rebuild the whole model. The real objective is more surgical.