[ 66862560.000 ] posts.x: Full talk on Spotify's generative personalization system, the NEO training recipe behind it, and how they grounded their LLM judges:  |   [ 66862560.000 ] posts.x: One in four US Premium subscribers on Spotify interact with its recommendation system every day. "Teaching LLMs to Speak Spotify" is @moustaki and…  |   [ 66854760.000 ] posts.x: Full talk on Numalab, the gesture system built to give a shape display its own body language:  |   [ 66854760.000 ] posts.x: An AI's first spontaneous act, given a body instead of a chat window, was to breathe. @cyrusclarke, a researcher at MIT Media Lab, gave it that body…  |   [ 66842160.000 ] posts.x: Full talk on the six month cutoff gap and what it takes to get a coding agent to actually check its blind spots:  |   [ 66842160.000 ] posts.x: Most models are already stale the day they ship: Exa's own graph shows roughly a six month gap between a model's knowledge cutoff and its release…  |   [ 66837480.000 ] posts.x: Listened to this episode of @FoundersPodcast this morning. As is so frequently the case, immediately bought the book after finishing the episode. Ari…  |   [ 66834900.000 ] posts.x: @GeoffreyHuntley Yeah, that seems low to me.  |   [ 66833760.000 ] posts.x: @zachlloydtweets @aiDotEngineer Thanks mate — nice talk!  |   [ 66784380.000 ] posts.x: Full talk on replacing assertion-based tests with exhaustive, screenshot-level verification:  |   [ 66784380.000 ] posts.x: AI can write a pull request faster than any team can review it. @meticulousgabe, CEO and co-founder of Meticulous, makes that case in "Why AI Didn't…  |   [ 66774540.000 ] posts.x: Full talk on the S3 metadata bottleneck, GPU failure recovery, and the infrastructure behind a 12 trillion token synthetic data run:  |   [ 66774540.000 ] posts.x: Pre-training runs are hitting a wall: the scaling laws demand exponentially more data for linear gains, and the open web only holds so many…  |   [ 66770700.000 ] posts.x: 🫡  |  

[ 66862560.000 ] posts.x: Full talk on Spotify's generative personalization system, the NEO training recipe behind it, and how they grounded their LLM judges:  |   [ 66862560.000 ] posts.x: One in four US Premium subscribers on Spotify interact with its recommendation system every day. "Teaching LLMs to Speak Spotify" is @moustaki and…  |   [ 66854760.000 ] posts.x: Full talk on Numalab, the gesture system built to give a shape display its own body language:  |   [ 66854760.000 ] posts.x: An AI's first spontaneous act, given a body instead of a chat window, was to breathe. @cyrusclarke, a researcher at MIT Media Lab, gave it that body…  |   [ 66842160.000 ] posts.x: Full talk on the six month cutoff gap and what it takes to get a coding agent to actually check its blind spots:  |   [ 66842160.000 ] posts.x: Most models are already stale the day they ship: Exa's own graph shows roughly a six month gap between a model's knowledge cutoff and its release…  |   [ 66837480.000 ] posts.x: Listened to this episode of @FoundersPodcast this morning. As is so frequently the case, immediately bought the book after finishing the episode. Ari…  |   [ 66834900.000 ] posts.x: @GeoffreyHuntley Yeah, that seems low to me.  |   [ 66833760.000 ] posts.x: @zachlloydtweets @aiDotEngineer Thanks mate — nice talk!  |   [ 66784380.000 ] posts.x: Full talk on replacing assertion-based tests with exhaustive, screenshot-level verification:  |   [ 66784380.000 ] posts.x: AI can write a pull request faster than any team can review it. @meticulousgabe, CEO and co-founder of Meticulous, makes that case in "Why AI Didn't…  |   [ 66774540.000 ] posts.x: Full talk on the S3 metadata bottleneck, GPU failure recovery, and the infrastructure behind a 12 trillion token synthetic data run:  |   [ 66774540.000 ] posts.x: Pre-training runs are hitting a wall: the scaling laws demand exponentially more data for linear gains, and the open web only holds so many…  |   [ 66770700.000 ] posts.x: 🫡  |