Research & Benchmarks

A new class of model claims frontier-level instinct at 40–1,000× cheaper than LLMs. We ran Jev against keyword rules, a logistic regression, llama3.1 and Claude Haiku on three real OnlyData classification tasks — measured accuracy, milliseconds and dollars. It beat Haiku on the 40-class task at a thirtieth of the cost, lost to a ten-second logistic regression everywhere we have labels, and failed our crawl-score task exactly the way its own limitations page predicts.

Evals used to be a spreadsheet of prompts and right answers. Then agents happened. The unit of evaluation has moved four times in five years — completion, task, process, outcome — and most teams are still grading the first one. Step contracts, chain math, the three clocks ground truth arrives on, the prediction ledger, and why six structural fixes come before you ever touch the prompt.

An Anthropic researcher quit saying the labs are “gambling with our lives” — and Anthropic’s own alignment lead agreed the number is over 10%. So we built the number: a weighted log-odds pool of researcher surveys, superforecasters, prediction markets, frontier-lab insiders and the skeptical case, all converted to one 10-year horizon, backcast to 2014, and re-run every Monday. Every input cited, both sides represented, and the dial moves on the data.

How hot, how long, how cold? Benefit and risk curves you can drag, a temperature × time sweet-spot map, a sweat-loss experiment to run with your crew (weigh-in/weigh-out, the towel audit, the waterproof pan), and the physics of why a löyly steam burst stings at temps dry air shrugs off. Shout out The Rabbit Hole. 🐇

A spoiler-gated companion to Gideon, Harrow, and Nona the Ninth. Tell it how far you've read and it unlocks exactly that much: full-spoiler summaries, a decoder ring for the confusing parts, the Nine Houses and their specialties, a ten-thousand-year timeline, character crossover — an interactive who-kills-whom graph where unread characters render as nameless ghosts, and a chapter-by-chapter map of Harrow with every card individually sealed.

68% of US GDP is us buying stuff — and GLP-1 drugs are the first technology that shrinks appetite at population scale. Grocery baskets, bariatric surgery, insulin, and alcohol are already denting; the price dam ($1,349 → $15 generics abroad) is breaking. Sourced evidence, winners and losers, and an interactive 2026–2035 GDP model with sliders. 🎧 With audio edition.

CaChink users name a song and watch an agent build them a playable tab, live — solos complete, original rig documented, every note validated before deploy. The trick: a hardened app on traditional infra, one queue row of coupling, and all the intelligence on a Mac Studio at ~$0 — swappable for API inference the day it matters. A pattern already powering four apps.

40+ scheduled jobs, 315 tables, two chat listeners, a fleet of local models — and silent rot. One typed graph of everything that runs, with freshness contracts, a model registry, and enforcement that blocks commits. In its first hour it caught a job that had been failing every 15 minutes for days.

The demo is a Telegram message; the product is everything under the waterline. Why almost every personal AI tool dies as a demo, the layered architecture (context repo → dispatch → protocols → cron → propagation rules) that made ours actually change daily life — and a public starter repo to clone.

Subscription buffets trained a generation of token-maxers. The buffet is closing — Claude Code is out of the $20 plan, weekly caps are tightening, and rolling windows are shrinking. Here's the math behind why it had to, and the incentive flip we should actually want.

From a 'Not found' page to 9 of 9 hard data invariants enforced. A six-hour autopilot cleanup that collapsed 411 duplicate company rows, canonicalized 1,042 ugly slugs, and backfilled 21,968 missing profiles — and the daily job that keeps it that way.

Of 90 companies the ai_native_v1 classifier tagged as genuinely AI-native, only ONE scored above AR 50. Character.AI scored 0. Cognition/Devin scored 11. The fix isn't a learned model — it's a one-line floor.

We embedded 24K companies into 384-dimensional vector space and explored what the geometry tells us about industry boundaries, agent readiness, and invisible connections. Interactive D3 scatters, spider charts, and force-directed graphs.

Our auto-classified Vibe Coder Stack had 94 cos in 38 categories — including a gas station and an optometrist. We pointed Claude at it through the OnlyData MCP, got a brutal audit, and rebuilt it as a 90-company curated list across 9 sublists.

We shipped the OnlyData MCP, then pointed it at our own catalog and found a massive blind spot in the agentic AI layer. 199 companies promoted from a private list to a public dataset — 155+ of them net new.

We scanned 8,250+ business websites for AI agent readiness with Algorithm E (Spread, v3). Six categories, subdomain probing, 0-100 score. Average 24. Only 9.3% grade A. Just 3% cross the well-behaved-SaaS ceiling of 50. Digits tops the list at 93.

We tested 5 local LLMs with 3 prompt strategies on 30 real Boise businesses. Chain-of-thought reasoning took accuracy from 17% to 77%. gemma4:e2b hits 73% on name alone.

5 models, 5 rounds, 30 real businesses. We went from 0% NAICS accuracy to 83% by splitting semantics from codes. The taxonomy was the bottleneck, not the model.