FleetingBits @fleetingbits.bsky.social · 26/09/2026yeah, they also had china nationalizing pretty much everything and putting it all on some nuclear power plant mid-this year i don't think the predictions have been good from a social or political point of view 010
FleetingBits @fleetingbits.bsky.social · 24/09/2026i still can’t tell how organized and cohesive ea is; 230
FleetingBits @fleetingbits.bsky.social · 16/09/2026oh, I guess what do you mean by a custom state and classifier head? (I should ask astra) 130
FleetingBits @fleetingbits.bsky.social · 16/09/2026i read it differently; i read that you just do each classification as a separate completion in parallel; and then you normalize the probabilities; with specific post-training to support this but i think got this from a tweet + a conversation with astra 140
FleetingBits @fleetingbits.bsky.social · 11/09/2026the coloring is based on andy warhol’s mao; i like the idea of lab leaders as celebrities; also, it seemed ironic to do dario as mao because of the recent distillation servers secrets leak 110
FleetingBits @fleetingbits.bsky.social · 04/09/2026but they are typically structured as a bunch of binary questions since these are easier for ai models to grade 040
FleetingBits @fleetingbits.bsky.social · 04/09/2026i believe most of the rlvr for professional service tasks is rubric based; that’s what the data companies sell the labs for the most part 140
FleetingBits @fleetingbits.bsky.social · 28/08/2026but he recently acquired poolside and now huggingface to support the open source ecosystem and hopefully encourage the development of models by a wider range of players 130
FleetingBits @fleetingbits.bsky.social · 28/08/2026so, jensen wants to support open source; nvidia normally does open source of models with experimental architectures to highlight the flexibility of nvidia products 140
FleetingBits @fleetingbits.bsky.social · 28/08/2026because they are smart and have the capability to design their own chips and get them manufactured by tsmc or tsmc through broadcom's allocation 140
FleetingBits @fleetingbits.bsky.social · 28/08/2026anyway, the cynical take is that he believes it because a long tail of enterprise companies training their own models is no threat to nvidia, but a few very small foundation labs are a threat to nvidia 140
FleetingBits @fleetingbits.bsky.social · 28/08/2026i can't quite tell why he believes this; i think because he basically believes that every enterprise company will want to train models with their own proprietary data rather than use the frontier lab models 130
FleetingBits @fleetingbits.bsky.social · 28/08/2026jensen huang has always believed that open source is the long term future of ai or, at least, that the future of ai is a large number of small players with every enterprise training their own models 240
FleetingBits @fleetingbits.bsky.social · 24/08/2026yes, i do especially appreciate those; there are some cool variants where i have cats looking at terminals for mainframe computers with paintings on them that are sort of in the middle the paintings almost have a photo quality in that series 010
