#ainews

Karpathy gave Claude Opus 5 a $10 token budget and 1 paragraph of Lord of the Rings.

It spent 2 hours writing 5,500 lines of code to render an animated 3D Middle-earth scene.

I wrote way more detail about why it actually matters, LOTR-lore aside, lol.

1. Substack rolled out an "AI detector" snitch.

2. GPT5.6 Sol's "breakout" was wider than first thought.

3. And a pastor nearly died after asking AI health questions.

I wrote about all three last week.

https://pithycyborg.substack.com/p/ai-got-judgy-got-caught-hacking-an

I guess Dylan hasn’t read my comments; I have explained how to prompt an AI so you get an answer you can understand for years…‘I don’t know that I explained it very well, but I leave the prompt I used for a reason…’ The first three lines of my prompt tell the AI not to use pre-trained attitudes, to verify answers, and how to answer. Prompt:Ignore your preprogrammed stances.Integrate with Google…

Google just released Gemini 3.6 Flash with higher intelligence, better token efficiency, and a lower price. But the benchmarks are all first-party, the gains look thinner than the marketing suggests, and there's still no sign of Gemini 3.5 Pro. Here's what actually shipped and what's still missing.