x1

Worlds most powerful AI app studio

[Manil Lakabi @lakabimanil_]

@Apple has the best designers in the world. apparently none of them have app store connect access.

[Manil Lakabi @lakabimanil_]

one thing from our agent research that i can't stop thinking about:

@OpenAI's computer-use agent would sometimes click an imprecise location, fail, and then repeatedly click around the same location until it ran out of steps.

we called it "coordinate persistence."

which is a very academic way of saying:

the model had enough intelligence to operate a computer but not enough self-awareness to realize it was doing the same dumb shit over and over

i think a lot of current agent failures look like this

less “the model isn't smart enough.”

more “the system has no reliable model of whether its last action actually worked."

[x1 @x1_apps]

coming soon 👀

[Rob Farlow @RobFarlow]

Running different browser agents on non-Webvoyager tasks clearly shows two things:
1) OpenAI's Operator is clearly better than the others
2) self-reported Webvoyager results don't mean anything

[Manil Lakabi @lakabimanil_]

openai trained deep research end-to-end using reinforcement learning 🧪

if you’re building ai agents and not thinking about rl, you’re already behind.

this is how the most powerful agents will get built.
thread below 🧵👇