Ian Mulvany

May 10, 2025

How do you build a product on top of a statistical system?

Hacker news threw up a story today: 

https://www.theregister.com/2025/05/08/google_gemini_update_prevents_disabling/ 
Gemini's latest model update has made the model refuse to discuss topics like sexual assault and rape. A developer is working on a platform to help victims, and their current app is not not working. It seems like there was a training update to an in-flight model. 

The same thing happened a few weeks back with GPT4.5 which had suddenly become obsequious. 

It seems like if LLM companies want to equip users with the latest and greatest they should support a "latest" model that gets any up to date tweaks, but also support named stable versions. 

The whole field is moving so quickly that practices like this are just not established. 

At the same time I was at an industry event this past week hosted by AWS and all of the large media companies in the room were uniform in saying that their strategy is to use frontier models, that local open source models are  more unstable than one might like. 

So I guess it comes down to being very careful in considering your use case, and being extra cautious as we remain in an environment of significant change. 

About Ian Mulvany

Hi, I'm Ian - I work on academic publishing systems. You can find out more about me at mulvany.net. I'm always interested in engaging with folk on these topics, if you have made your way here don't hesitate to reach out if there is anything you want to share, discuss, or ask for help with!