US labs say they've fixed distillation, even against small players like PewDiePie. So how is Kimi K3.1 beating them on benchmarks?
Either distillation was never the secret, or it was never fixed. Both can't be true.
US labs say they've fixed distillation, even against small players like PewDiePie. So how is Kimi K3.1 beating them on benchmarks?
Either distillation was never the secret, or it was never fixed. Both can't be true.
@theo I would spend them on the platforms you already run. Point Codex at each one. Have it audit performance, flag what is broken, and suggest how to make them better. Then make it work the issues it finds through to completion. Codex is very thorough at audits.
@AiZeroToAIHero@HermesWatcher I already use projects. They hold files, not the running conversation. The context I cannot rebuild is the conversation and the decisions inside it.
The AI drafts a flyer in minutes now, so the draft itself is worth almost nothing. What people pay me for is the edit, the eye that knows which version actually ships. Everyone can produce more now. Taste and judgment became the product.
Most 'AI agents' I get shown are just workflows with a fancier label. A workflow runs the same steps every time and never breaks. An agent picks its next step, and that judgment is exactly why it can also make things up. If you can draw the steps, reach for the workflow first.
Most 'AI agents' I get shown are just workflows with a fancier label. A workflow runs the same steps every time and never breaks. An agent picks its next step, and that judgment is exactly why it can also make things up. If you can draw the steps, reach for the workflow first.
Phishing emails now carry two messages. The one you read, and a hidden one for your AI assistant. Barracuda found an invoice telling the summariser to change a supplier's payment details. If your AI reads your mail, nothing sends or pays without your yes.
We built this one on purpose. Count the tells. Red and yellow brush script, a paint smear behind the name, a blue glow on SPECIAL, a tilted drink cut out in the top right, a starburst price badge, four unrelated icons, a bokeh background with a lens flare. Almost no empty space.
Slopocalypse.
I am all for designing flyers with AI. But look at what is landing in the feed and on print. Good design is a series of choices. A model picks the most average option, then fills the space. That is why so much of it looks identical.
The food ones are worse. There is a real uncanny valley for food. A 2025 study in Appetite found near-miss AI food is rated more eerie and less pleasant than both real food and obvious cartoon food. Broken realism repels. Cartoon does not.
Here is the kicker. Polished AI food can score as more appetising than a real photo, until you say it is AI. Then the boost vanishes. The killer is unchecked AI passing as real. So one person who knows the product checks the hands, the text, and whether the food is what you serve.
People assume an AI agent fails because it is not smart enough. Usually it did exactly what you said, with no judgement about what you meant. I treat mine like a fast, literal new hire. One narrow job, what done looks like, and I check it before it ships.
@LethueK Some convert. Most chase views. The difference is whether the skit sells the product or just the joke. The ones that work usually show the thing and tell people where to buy.