3 Comments

The title of your article is interesting because as a psychiatrist who works with people with dissociative identity disorder (DID), my gut reaction to reading Kevin Roose's Valentine's Day chat transcript with Sydney was that many of the AI's responses sounded very much like interactions I've had with hidden parts of an internal system in someone with DID. In people with a fragmented self structure, there is never just one "alter ego," so I wasn't surprised to hear about Ben Thompson's interaction where we learned about Venom, Fury and Riley. It's fairly obvious that Microsoft/Open AI did not intentionally build in modes like Venom and Fury who want to retaliate against users, but does anyone know if the specialized modes you describe (eg Riley optimized for travel inquires, Berry for news and maybe also music) were intentional? Did the AI self-generate the names of these modes? I am very curious how much of these hidden personas are emergent vs planned. Does anyone have more information?

Expand full comment

Fascinating and worth a deeper look. Thanks for sharing! I am sure we will find there are more 'personalities' and modes under the hood. This makes perfect sense from the POV of making the most useful platform architecture for this kind of interface, as you need different types of prompts and fine-tuning for it to be the best in these different verticals, interaction types. It would be interesting to know how they implemented it though. Is it a fine-tuning of the LLM or a whole additional layer embedded in the API, and what datasets are added for each mode and how? The base chatGPT knows nothing about current events, weather, airplane tickets, etc. so that is all added in somehow. Integration is key.

Expand full comment

these modes are pretty legit. Game mode was not made up, just not codenamed yet

Expand full comment