re: Fast followers in AI
In distillation, you have a student model sampling traces from a teacher model. But in certain cases, the student may receive such high quality traces that it surpasses the teacher model, despite the teacher being much larger.
Sonnet 3.5 is a real example of this: despite being a smaller and cheaper model than Opus, it performed noticeably better!
So when people bring up the water skier-boat analogy to say that the water skier (the student) always lags the boat (the teacher), they are wrong. The student can always surpass the teacher.
This is also why the rhetoric of a certain senator-elect is confusing: he argues that we can and should pause AI because China is nothing but a fast follower. So while this infantilization is wrong on its face given how many Chinese researchers are at American labs, it is also, on a technical level, incorrect.
Hudson River Trading's head of AI Iain Dunning on what happens in the second after a jobs report:
"We have like a Bloomberg headlines feed that is like pretty low latency."
"You can do everything from having kind of a handcrafted logic to look for keywords through to putting it through like an AI model."
"There are options trading firms that have thousands of people that are essentially cyborg trading options."
"For a hackathon, we got a PlayStation controller and kind of gave people the chance to try and practice reacting to events very fast. It's really tough, but it's a learnable skill."
"It's fast for any normal day-to-day thing, but for markets, it's kind of slow."
Companies cannot tell that agents are using their apps.
For reference, agents personally booked 6 Uber rides for me this quarter, so I'm pretty sure the below numbers are wrong (lol)
i was wrong: i am constantly sending my friends messages intended for an LLM and look like a complete idiot
@Apple please extend the length of time wherein we can unsend messages lol
BTW personal assistants sending emails on your behalf very easily fool Pangram
Reason: there often aren't enough words in emails + assistants are doing their best to replicate your speech
As AI diffuses to the public via person assistants, detection will basically become impossible unless watermarking tools become available to the detectors