Calvin Zhou

upenn m&t

[Calvin Zhou @calvinyzhou]

@daniel_mac8 one order of nuclear fusion please

[Calvin Zhou @calvinyzhou]

@ChungMinKim by clipping are you effectively restricting updates to near-on-policy regions? would this hurt creativity? does seem like the tradeoff is worth it

[Calvin Zhou @calvinyzhou]

@cath_zhou @speedrun inspirational stuff

[Calvin Zhou @calvinyzhou]

@ericjang11 thoughts on opus 5.5?

[Calvin Zhou @calvinyzhou]

@theworldlabs @AMD we might finally get hardware that stop people from saying "WMs are too expensive to train/run" like how Nvidia sped up LLMs

[Calvin Zhou @calvinyzhou]

@LarocheUlysse @RemiCadene agreed more rl on cad isnt true spatial intelligence

[Calvin Zhou @calvinyzhou]

@Adamjhung is this more or less efficient than dreamer4 that already doesn’t train on rgb? i assume quality is better which makes it worth it

[Calvin Zhou @calvinyzhou]

@RemiCadene agreed, imo physical agi has a higher probability to bring upon extinction than llms, but with a longer time horizon

[Calvin Zhou @calvinyzhou]

@tongzhou_mu do you envision better scaling laws than egocentric human video? especially when you get to million-hour scale like Dyna-2?

[Calvin Zhou @calvinyzhou]

@ZechenBai woah this is sick. what is the speedup from real-time in the videos? any thoughts on the speed and efficiency of this approach?

[Calvin Zhou @calvinyzhou]

@L42ARO @_wenlixiao one could argue llms with tool calling ability are VLAs...

[Calvin Zhou @calvinyzhou]

@kylifec go kylie!!

[Calvin Zhou @calvinyzhou]

@WLOGSarim @YuXiang_IRVL imo might recover better than current robotics models since they handle context better esp long-running context

[Calvin Zhou @calvinyzhou]

@FangYunhao_X probably a mix of improved sensor tokenization, compression, and time series encoding so we don't waste attention on raw streams

[Calvin Zhou @calvinyzhou]

@MozarellaPesto so WMs become a tool call or a subagent?

[Calvin Zhou @calvinyzhou]

@cath_zhou @speedrun @michael_wu8 congrats!!