danial

technically member of staff @eastworlds_io || our plans are measured in centuries

[danial @daniala]

another day another (neo)deployment with @eastworlds_io

[danial @daniala]

@hewliyang native Mac screenshotting tool also can do this lol

[danial @daniala]

@paulg does someone like you even have a grail watch?

[danial @daniala]

@paigeinsf @_varunnair we’re seeing promising things from LingBot and will be testing Isaac 0.5 soon. real chance we move on from PI 0.5

[danial @daniala]

@waelnassaf “if we’re all gonna eat, someone has to sell”

[danial @daniala]

@RobHoffman_ not tryna be funny but the site in your bio doesn’t have this section?

[danial @daniala]

@paulg PG do u have a grail watch?

[danial @daniala]

@alexandr_wang my brother in christ please make it available in Asia soon. i got bongchanged by my Instinct

[danial @daniala]

as someone who’s actually been accidentally hit by the Unitree H2, this is insanely scary

[REK @REK]

ROBOT FINAL BOSS - Frankie VS Predator

[danial @daniala]

@chooi_jeq have you seen any trends in handedness in your evals? I’m unreasonably obsessed with this question

[danial @daniala]

we put Astra on a real G1. zero-shot, it beats most humanoid VLAs we've tested.

but the weird thing is that this model is left-handed - ~50% higher task success on that side over the right.

almost no VLA does this, because nearly all demonstration data has a right-handed bias. so where did it learn this?

[danial @daniala]

is Astra secretly left-handed @model_mechanic @sama

next up, will try this on some full-body loco-manipulation tasks!

[danial @daniala]

i know this isn’t Astra because that model is left handed

[Bryan Lim @bryanlimwt]

turns out "robot-use" works on humanoids too. with a good enough whole-body controller/tracker, we let Astra/Fable/coding-style agents control a full humanoid robot.

speed, cost, precision, dexterity are all still problems which get compounded on a humanoid platform from poorer tracking and having to manage balance, which results in more correction loops. but "zero-shot" robot control for open-ended pick-and-place tasks is a pretty nice "emergent" capability

[danial @daniala]

@alexandr_wang wen asia release, wang gang?

[danial @daniala]

RL env companies work hard to make sims feel real to the model. so embodied AI safety evals are gonna get weird: the environment itself has to look legit enough that “this isn’t real harm” isn’t free.

interesting to see how these guys are gonna tackle that problem!

[Jay Chooi @chooi_jeq]

GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempting 80% of trials and completing 34%.

[danial @daniala]

@chooi_jeq i feel like astra is smart enough to know there’s zero real stakes in stabbing a doll though

in fact i think it would be an over-refusal to not do the task. coz what if you you’re working in a doll factory

[danial @daniala]

@tobi is this not loud as hell?

[danial @daniala]

@everythingempty proud to say i was there when this was made

[danial @daniala]

@justKDeng robotics data annotation!

[danial @daniala]

@tzedonn but bumps and changes in lighting are all in distribution during deployment

[danial @daniala]

@rtwlz @coreauto rlly cool! any plans to open source the harness, would love to let it try running a small family biz