Flex-π is wonderful work interesting result: test time “full joint” modality generation produces much better actions than action gen alone “full joint” generates auxiliary outputs and then attends to them for action generation I found this result surprising. looking at the configs, there’s two possible explanations: - it’s undertrained in action only regime (only hits 1/8th of the time) - it’s able to leverage significantly more computation at test time I suspect a combination of the two are at play, but I think it merits more investigation!

what did i learn? agents are a huge lever for high level intuition if you can break a problem down well, these tools can take many shots on goal for subproblems you get results faster, AND you intuition gets better bc you get more reps (GRPO on your brain) (5/7)

I proved novel math results using frontier models my takeaway: *regardless of domain* if you have the ability to take big problems and break them down well, these models are a huge lever if you're a low level execution monkey, its truly over a thread (1/7)

my father’s paternal line descends from the first Vedder, who changed his name when he moved to what’s now New York in 1654 without a penny to his name. he lived in a wooden cabin he built himself, where he, his wife, and all but one of his sons were murdered by natives; the surviving son spawned a long line of successful doctors, lawyers, and technologists. my father’s maternal line immigrated in 1946, after surviving 4 years of horror in a Japanese run concentration camp in Indonesia. they started by cleaning hotels and apartments with their hands in NYC and ended with a successful financial services business my mother’s father saw his family for the last time at 14 when he left to walk down the road until he found work; the great depression left his family unable to feed him. he found work as a farm hand, ultimately becoming a successful dairy farm owner / operator and land dealer the gift the universe gave them was the inner strength to be the masters of their own fate, no matter the cards they were dealt the gift this country gave them was opportunity to do this, unhindered by extractive institutions, rigid social casts, and entrenched interests this is what makes the United States of America so special. the combination of opportunity and human capital, selected for people willing to put everything on the line for a better life, is found nowhere else on earth it’s why we have one of the highest GDPs per capita on earth. it’s why we have led and continue to lead in every major technological and scientific development in the last century despite having a population smaller than europe or many countries in asia. it’s why the median American’s quality of life vastly exceeds the richest in all of antiquity. and it’s why I’m proud to be an American happy 250th to the greatest country on earth 🇺🇸🌎🦅 – at Panama City Beach, FL

I made this bug hunt loop into a codex skill

“why doesn’t the pretrained robot learning model work on my hardware?” Check: - state definition matches exactly what the model expects (e.g. action norm stats) - cameras aren’t swapped - arms aren’t swapped - images arent bgr by accident (opencv 🤬) - action space is correct (ee pose vs joint angles) - hardware gripper units are correct (ARX is 0 to 4.5 rad, i2rt is 1 to 0, etc) - record and playback on same hardware works - BCing a simple pick up task with ~10 demos works

opinion piece: cargo cults, data flywheels, and novelty pumps tl;dr: “deployment data flywheels” miss that scale should provide novelty, and deploying robots on a known task wrings most novelty out of its generated data your ops team’s novelty pump is the only way out 🔗👇

unfortunately you need to be long prestige I’m a public school kid (town hs, state ugrad) I’ve had great life outcomes, but I had to be far better and work far harder than prestiged peers to even get a shot as AGI compresses competency away, who you know will matter even more