🎭 Someone spent $20,000 so AIs could lie to each other at Werewolf

Podcaster Dwarkesh Patel paid about $20,000 in API credits to make frontier models play Blood on the Clocktower — a social-deduction game like Werewolf — against each other for 386 matches.

One model plays the whole good team; another plays the whole evil team. OpenAI's GPT-6 Astra leads the board at about 1,859 Elo. Against Claude Opus 4.6 it wins roughly 92% of the time. Dwarkesh says the edge is mostly catching liars, not telling better ones.

Full rankings and replays are on clocktowerbench.com. You can also play the AIs yourself.

@aipost 🏴

📢 Party flyer or shop ad in ChatGPT: the prompts 1,100+ people saved

"Design a flyer" gets you AI slop, says the AI creator behind these seven prompts. They spell out the size, every line of text, the colours and what to leave out.

The event flyer, shortened:
Create a polished 4:5 event flyer for [EVENT NAME]. Date: [DATE] Time: [TIME] Location: [VENUE]. Highlight the event name first, showcase the main attraction second, and present the essential event details third. Apply [PRIMARY COLOUR], [SECONDARY COLOUR], and [ACCENT]. Avoid unnecessary icons, fabricated QR codes, harsh shadows, overcrowded layouts.

Lines worth stealing from the others:
• Use a maximum of two font families.
• Avoid generic stock photography, tiny text, fake logos.
• Never fabricate statistics or include unverified data.
• For a YouTube thumbnail: Avoid exaggerated shocked expressions, excessive arrows, unnecessary red circles.

ChatGPT's image tool got an update on Nov 4. OpenAI says it now works 50% faster.

@aipost 🏴

🧠 Claude now "dreams" between chats — so it can fix its own bad notes

Anthropic's new engineering post says agent memory works a lot like sleep: Claude writes notes during a session, then offline "dreaming" replays the transcript, corrects mistakes, builds wiki-style maps, and promotes what the next session should load first.

In a Pokémon Red test, Sonnet 3.7 dumped 413 memory files and earned one badge. Fable 5 kept nine files and earned four. When Claude wrongly marked Mt. Moon as finished, dreaming checked the transcript, rewrote the note, and Claude went back into the cave instead of looping.

On ~350,000 internal company files, scaling dreaming from 2 to 128 sessions lifted a 25-question eval from 13% to 48%. Anthropic's write-up frames dreaming as idle compute that organizes memory — not a change to the model's weights.

@aipost 🏴