Arjun

[Charlie Deane @KicksCharlie]

Prompting is temporary. We believe the future of computer use is completely unprompted.

Introducing our latest research: a browser agent that predicts your next action from only the current page and your recent history.

[Arjun @arjunsriram2002]

@KicksCharlie So cool! Just waiting for it to start paying for me too 😃

[Arjun @arjunsriram2002]

@OwainEvans_UK Sports psychologists have known a version of this for decades. Never say "don't miss this forehand," say "hit a forehand winner."

This sorta maps to your Pink Elephant result and it's interesting to see the parallels between humans and models!

[Arjun @arjunsriram2002]

Full write-up: https://t.co/GX57W0JIYC

[Arjun @arjunsriram2002]

I asked a model to choose between two options completely at random, 10,000 times. It picked the one it liked 69% of the time.

By steering its activations, I found I can move that pick from 0% to 100%.

Somehow, it still always claimed that it chose at random! What's going on?

[Arjun @arjunsriram2002]

Limitations: one model (Gemma 3-27B), narrow task set

Curious what this looks like in reasoning models, where there's a visible chain of thought that could either expose the bias or rationalize around it.