Midwit writing used to be woke.Now midwit writing is all AI.“It’s not this, it’s that.”A superintelligence, yet midwit. ...

I think we’ll need new ideas. Deterministic logical reasoning on top of the probabilistic AI models we have today. Otherwise error rates are just too high when you do tasks of any complexity or duration.Vik ⚡️ (@its_V1k):@balajis 100% AI will be optimal at some point though surely? ...

0% AI is slow.But 100% AI is slop.So the optimal amount of AI is actually between 0-100%. The exact figure varies by situation, but just the idea that 0% and 100% are both suboptimal is useful.It's the Laffer Curve, but for AI. ...

There are many different proof-of-human mechanisms that you can layer on top of each other. They have different and hence complementary failure modes.Use as many as you can cheaply introduce. From 2FA to MFA, multi-factor authentication.Steven Sinofsky (@stevesi):@balajis How do you solve the realID and association with an account problem (in case authentication is breached)? Or does this depend on algorithmic determination of bot'ness? ...

Of course, but you can get very far by just detecting generic ChatGPT output, and perhaps also the top few models. Most people don’t change defaults.It’s like Snapchat setting norms on disappearing messages. Not perfect, but works well enough to set the culture of the app.Dhru (@0Dhru):@balajis If it’s trivial for an algorithm to flag generated text, it is trivial to program a reward function for a LLM to produce undetectable text. ...

Yes. But then comes the third level of defense, which is trusted human moderators doing occasional bot-or-not flagging to train the algorithms. I think in practice you could get fairly good at this if the system was built for it, and if most humans in the network cooperated.Yishan (@yishan):@balajis I think this will run into the “motivated bears are smarter than the laziest humans” problem and any system that detects all bots will have a high false positive rate.This is probably ok in practice because huma ...

Yes, but then the second layer of defense comes into play: the anti-AIrcraft guns, namely the AI detectors.If it’s easy for a human to flag ChatGPT-generated text, it’s trivial for an algorithm.Conor (@jconorgrogan):@balajis problem is, humans will just type in what they should post into their llm and repeat what slop it saysalready seeing this with even things like tributes to people who died- friends/celebrities etc posting on socials clearly copying the slop mannerisms of 4o output for ozzy ...

The closest thing to this right now (besides IRL, of course) is probably a corporate Slack. But that has limited scale.There’s room for a new kind of app with no AI at all within its walls. No slop, no spam.Like Twitter’s original 140 character limit. But limited to humans. ...

An important kind of social network will be one where no bots whatsoever are allowed. ...

Most people live here. https://t.co/NlrkVp6tkX ...