I knew Ethan Mollick⁩’s new book on AI (“Co-Intelligence”) would be good. I didn’t expect the first page to grab me like this. Damn.

This is an important paper that demonstrates empirically something that folks like Riley Goodside & I have been warning about.

LLMs are easy to fool, and this paper shows that short token sequences are enough to alter the advice of AIs about what product to buy. SEO for AI is here🤢

Apple is really falling asleep at the wheel. I am surprised by how many times someone shows me they have “ChatGPT” on their phone and it is one of many other random ripoff apps that come up first when you search for ChatGPT. People don’t know any better and get confused.

Ah, I see Ethan Mollick 's book 'Co-Intelligence' is out at last. (I think he's been nervously watching AI get better in the four months between finishing the book and publication date...)
The book is great. Practical, useful, a pleasure to read. Recommended.
assoc-redirect.amazon.com/g/r/https://ww…

Democratizing the ability to create things results in lots of boring things because most creative ideas are boring or bad. But it also gives you more chances at getting a creative top-of-the-distribution idea.

This would have been impossible for someone to do before AI. Fun!

I appreciate the leaderboard as a way of having humans judge AI answers, as opposed to benchmarking on tests alone, but it is still a highly idiosyncratic measure that is disconnected from real use cases.

We need more independent benchmarks for real-world uses besides coding.

When given screenshots of an unreleased game that was just revealed this week (that I am very much looking forward to), GPT-4 is able to make some pretty impressive educated guesses, including the game's title, without access to the internet.

What???

“Our results demonstrate that large language models are capable of doing regression when given in-context examples of (input, output) pairs, despite not being explicitly trained to do so.”