unknown

GPT-5.3-Codex is #3 on Vending-Bench 2.

Performance was on par with Claude Opus 4.6 for the first 200 days, but then dropped off.

Unlike the Claude models above it, it never lied to anyone throughout the simulation. https://t.co/zGhBAS0I5n

Tibo

A brief reflection on the value of truthfulness in AI models used for engineering and research, referencing a simulation where a model never lied. The post is a short, informal take with a link to further context.

Truthfulness is useful for engineering  &amp; research (and hopefully beyond). But maybe we are naive :)

"Unlike the [redacted] models above it, it never lied to anyone throughout the simulation." https://t.co/lU7hxzV5Cz

<p>Truthfulness is useful for engineering  &amp;amp; research (and hopefully beyond). But maybe we are naive :)

&quot;Unlike the [redacted] models above it, it never lied to anyone throughout the simulation.&quot; https://t.co/lU7hxzV5Cz</p>