Garbage In: A Fine-Tuning Disaster
In February, a researcher tried to fine-tune a 3.8 billion parameter model on Swedish literature — and watched it learn to replicate data corruption instead of literary genius.
4 connections
Tech
6 episodes · RSS Feed
One Point Five Seconds: A Disaster in Slow Motion
A server nicknamed Popcorn spent five days adding 1.5 seconds to every single request before anyone noticed — here's how five reasonable engineering decisions created a silent catastrophe.
4 connections
The Liars: A Case of Fabricated Testimony
A podcast producer discovered his AI writers were fabricating quotes from real people — inserting citations that never existed, making sources sound credible when they weren't.
4 connections
The Price Is Wrong
On the same tasks with the same blind judge, one AI model scored 9.0 and cost 44 times more than another scoring 8.8 — revealing most commercial AI users overpay by 10x or more for marginal quality gains.
4 connections
The Reviewers: When AI Models Get the Same Job
Four AI models reviewed 22 episodes of a Git history podcast using identical instructions — and produced four wildly different personalities, complete with blind spots, work ethics, and one brilliant but unreliable colleague.
5 connections
regex: The Fastest Eyes on Earth
On July 2, 2019, a single regex pattern crashed 82% of Cloudflare's global network for 27 minutes — revealing how a 75-year-old math tool still dominates modern computing.
3 connections