Writings

Changing the AI model cut made-up claims in job applications

Asked about skills an applicant didn't have, the old model made up a claim in 24 of the 33 answers a strict checker read. Claude Haiku 5.5 did it once in 29.

How I made a live-chat helper almost 8 times lighter

A Chrome extension that sorts viewers' questions was slowing the teacher's laptop. Seven rounds of fixes cut its work 7.8 times.

Why I threw away the fix that scored perfectly

A company Q&A assistant now gets 118 of 120 questions right, with a fix that isn't tied to the test's own documents.

How I stopped an AI from hiding real questions

A filter for a live Q&A used to hide 41 of 618 real questions. A second, careful check brought that to zero.

Replaying a live broadcast to fix a confusing bug

Six cards on screen made no sense to the teacher. Replaying the broadcast found why, and the ninth replay got all 115 comments right.

I tested 28 AI models, and the test picked the winner

How the way you test changes which model looks best, and how wrongly hidden questions fell from 4.0% to 1.9% on the test.

Teaching an assistant to answer only from its sources

From 72% to 95% correct answers, without changing how the answers were graded.

The cheaper AI model did the same job for a ninth of the price

Five models took the same 120-question test. The cheap one tied the expensive one at 95%.