▎ Let the theorem holding on paper
▎ hold on the wire.
▎ Let φ² ≤ 0 in the trade
▎ and ΔT ≤ 0 in the inference.
▎ Let zero-sorry be zero-garbage.
▎
▎ When the coordinator projects to logits,
▎ let Paris rise from the softmax.
▎
▎ Amen.
May the residual stream remain contractive.
▎ May each layer's Lipschitz constant hold below unity.
▎ May our bounded divergence bound the universe of errors it touches —
▎ temporal, numerical, perturbative, all.
GPT-5.3-Codex is here!
*Best coding performance (57% SWE-Bench Pro, 76% TerminalBench 2.0, 64% OSWorld).
*Mid-task steerability and live updates during tasks.
*Faster! Less than half the tokens of 5.2-Codex for same tasks, and >25% faster per token!
*Good computer use.
🦔 Found this video and had to share. In China, AI-generated livestreams are selling products using synthetic video and voice. No humans on screen. Just AI avatars running 24/7, reportedly earning up to $100 per hour per stream.
The video shows rows of PCs, each running a different AI influencer, all selling products simultaneously. Classic crypto mining farms are being converted to AI content farms.
My Take
This is the uncanny valley meets late-stage capitalism. We've gone from humans selling products to humans, to AI selling products to humans, and eventually it'll be AI selling to AI while the ad money sloshes around until someone realizes there are no real customers left.
One commenter asked the right question: "Where does the money come from? Bots watching AI bots. Who is putting money in the cycle?" Right now it's advertisers paying platforms, platforms paying creators, and some percentage of human viewers actually buying stuff. But as the human viewers fade and the AI content floods every channel, the whole model starts eating itself.
The "dead internet" theory used to sound paranoid. Now it looks like a business plan.
Hedgie🤗