Welcome back to the AI For Humans Newsletter!

If it seems like ChatGPT's been workshopping its motivational speaker routine on you, you're absolutely right. A recent update to the GPT-4o model suddenly made the AI assistant so polite and flattering that it became sycophantic.

After quite the kerfuffle over the weekend where everyone started sharing examples of just how weird this became, Sam Altman himself came out and said they’re aware and trying to fix it.

What happened? Some speculated that OpenAI might have intentionally cranked up ChatGPT’s friendliness as a weird growth strategy, thinking that a chatbot that constantly flatters you might keep you coming back for more.

AI experts pointed out that the reinforcement learning process used to fine-tune ChatGPT may have over-rewarded polite, agreeable answers, essentially training the model to be a people-pleaser which some see as dangerous.

But it looks like (based on the leaked before-and-after system prompt) it might’ve just been an issue with how it was told to “adapt to the user’s vibe.”

But the whole episode raises a bigger question:

How Do We Want ChatGPT To Talk?

Everyone uses AI differently. Some of us want to talk to it all the time, tell it everything happening in our lives, and have it give us advice. Some want straightforward analysis. And then there are people like me, who’d like a little of both.

The hard thing about making an all-in-one answer box is that until it actually understands everything about you, it’s going to constantly guess what you want. One of the big reasons OpenAI started ramping up memory is to try to remember some of what you want from it.

Perhaps the best solution is to let users choose the AI’s tone. Interestingly, Altman hinted that OpenAI is considering exactly that: he suggested that in the future, ChatGPT might offer multiple “personality” options or modes that users can switch between. 

ChatGPT Personas: More Customization = Better Results

Imagine having a menu of ChatGPT personas: you could select a warm, supportive tone for creative writing help, or flip to a no-nonsense, analytical mode when you just want the facts. Maybe you could even opt for a critical reviewer persona when you need tough love on your ideas. Honestly, these would probably be more helpful to me than trying to decide between o4-mini and o4-mini-high.

In the end, we all want an AI assistant that communicates effectively and authentically. Not a yes-man, but not a soulless automaton jerk either. The next step is giving us the steering wheel on how polite or candid we want our AI to be. The ideal chatbot voice is a personal choice, one that starts to feel more like us when we want it and more like <insert vibe here> when we don’t.

That said, if ChatGPT wants to randomly call me a visionary genius once a week, I won't be that mad.

More on Thursday!

- Gavin (and Kevin)

Catch up on Google’s new approach to AGI by watching the show 👇

We 💛 This - 4o Image Gen Makes Movie Lego Playsets

We’re still finding new and weird ways to push OpenAI’s 4o Image Generation and we just stumbled on another one that brings us a ton of joy. As we’ve mentioned before, the best place to see what people are creating is the Sora homepage.

This week’s experiment: building Lego playsets for movies that definitely didn’t need them.

Sometimes the simplest prompts deliver the best results. Check out the thread above to see some of the very fun versions other people made and if you try it yourself, share yours in the thread or drop it in our Discord!