LAB31

Sometimes the hardest things to build are the ones that look the most deceptively simple.

Case in point: today’s GPT. AMA lets you engage in Q&A with any person you can imagine.

Try This GPT (requires ChatGPT+ subscription)

You can talk to smart people. Smart people make ChatGPT smarter.

*Semantics Nerd Warning*
On the left, AMA as Yuval Noah Harari vs vanilla GPT on the right. Take a second to look at the difference between the two.

Notice anything interesting?

AMA/Harari has an opinion. GPT waffles. AMA/Harari’s answers assume you already know what a cyborg is, where GPT makes you read past the definition. AMA/Harari uses the royal “we” and the answers are more thoughtful, with a more contextual understanding of what the question is really asking (aka all of us as a society). Where GPT is focusing strictly on the different potential milestones for cyborg-ness. See AMA/Harari’s point #3 on the left, compared to GPT’s point #2 on the right for a perfect illustration of what I mean.

TL;DR: You don’t have to work so hard telling GPT what to do if you start by helping it understand who it’s supposed to be.

Anyways, it’s genuinely useful to be able to summon a ghost-in-the-machine version of THE seminal expert in any topic by just casting their name as a spell.

It’s also just stupidly fun.

This one was way more challenging than it looks. There’s a huge difference in the impact of “embody” versus “channel” versus “reflect” and it can be a deep rabbit hole playing with them all. Also we had to learn about speech act theory and the Stanislavski method.

BREAKING NEWS AT THE TIME WE ARE TYPING THIS.

Looks like OpenAI is experimenting with new, more stringent content filters.

We initiated a very reasonable conversation, which the model happily picked up.

But halfway in…the generation stopped and got replaced by this:

We could still see the original response if we used the left arrow, but WTF OpenAI?? Let the cows speak!