mark

19 | berkeley

[mark @markshi07]

@thesimaximalist @JensenHuang 100,000 gpus synchronized and the keynote audio still ends up in someone's earbuds. the real last mile problem

[mark @markshi07]

@willdepue @joshuawolk @LukeFarritor amateur astronomy might be the best possible domain for this. decades of public survey data nobody has time to dig through and a community that already knows how to verify a finding

[mark @markshi07]

@Alfred_Lin the sprint out of the gate point cuts both ways though. teams that sprint before they know what to build just burn their best months. speed compounds once the direction is right, before that it's mostly motion

[mark @markshi07]

@runwayml open weights under apache 2.0 for a world action model is a real commitment. does the video pretraining actually transfer to robot morphologies it never saw, or does every new embodiment still need its own fine-tuning data

[mark @markshi07]

@C_A_G0101 most people still treat prompt injection as a text problem. if the model can read instructions that are invisible in the image itself, eyeballing an upload before sending it stops being any kind of defense

[mark @markshi07]

@dhh sharing any local model across your tailnet is underrated. a dgx spark as the house inference box with every other machine just borrowing it is a pretty compelling setup, assuming it can actually serve a big enough model to be worth sharing

[mark @markshi07]

@Polymarket the privacy argument is probably real but the FAA rules part feels like the first excuse that stuck. airlines have wanted to kill third-party access to booking for years, agents just forced the issue

[mark @markshi07]

@chris_j_paxton the towel folding demos are getting weirdly reliable lately. curious how much of that is the model and how much is the setup being carefully chosen, though i guess that's exactly what the coliseum setup is for

[mark @markshi07]

@KJHMiao @karinanguyen @UCBerkeley sounds cool, how would one enroll?

[mark @markshi07]

@justsamyyyy_ Super cool! Would it, say, be able to keep adjusting positioning based off of calculated win% to find the optimal placements during the 30s planning phase? Paired with aug like Future Sight this could be super broken...

[mark @markshi07]

@aakashgupta how much of the vertical moat is knowing the job vs having permission to actually do it?

[mark @markshi07]

@CurseMuse flagg making the all-time team already is a very efficient rebuild

[mark @markshi07]

@TheTuringPost "perceptive humanoid motion from simulation" is a title precision-engineered to get a click out of me

[mark @markshi07]

@warriorsworld sengun, wemby, zion AND flagg on one road trip is a war crime against draymond's lower back

[mark @markshi07]

@bvsbharat is the 6-9x mostly from bigger stable timesteps or a cheaper narrowphase? catching a 0.7mm card at 16ms is wild either way

[mark @markshi07]

@VadimmmO "almost profitable" is just "losing money" in a nicer outfit

[mark @markshi07]

@kevin_zakka mjlab collecting 'proudly powered by' posts like infinity stones

[mark @markshi07]

@swann_bessa crazy speed-up. nice to see significant improvement to a very fundamental algo that everyone learns in intro classes

[mark @markshi07]

i think the recent shift to glp-1s and protein craze has also been slightly amplified by recent looksmaxxing/blackpill content. before, being fit was attractive as it was (largely) representative of proper health, now, people are sacrificing their health for the sake of aesthetics

[mark @markshi07]

@VokabreRobotics @hisashi_is affetto is still the most unsettling robot face ever made and nothing is taking that crown