@jefrankle
, Chief AI Scientist at Databricks, on the role of human oversight in #AI progress. "If you don't have human insight or a process for selecting relevant synthetic data, the model just reproduces its own behavior."
RESEARCH
-
Human Oversight Critical for Preventing AI Model Behavior Reproduction
By
–
-
Disagreement on Robotics as Pure Reinforcement Learning Challenge
By
–
I think we agree about scaling and overly optimistic claims, but I don’t agree with your assertion that “Robotics is 100% a RL/sim challenge”.
-
Open Artifacts Required for Blog Post Coverage
By
–
i'll take blogposts as long as they come with open artefacts on the hub
-
Synthetic Data: Breaking Through AI Scaling Walls?
By
–
is synthetic data the key to pushing past scaling walls or is the snake devoring its own tail?@poolsideai co-founder and cto @eisokant, explains why he’s an optimist in our recent discussion on @airstreetpress. pic.twitter.com/J5vtthzqF3
— Nathan Benaich (@nathanbenaich) 4 décembre 2024is synthetic data the key to pushing past scaling walls or is the snake devoring its own tail? @poolsideai co-founder and cto @eisokant
, explains why he’s an optimist in our recent discussion on @airstreetpress
. -
LLM Limitations in Consumer Behavior Research Analysis
By
–
The Deceptive Dance of AI in Consumer Behavior Research: A Critical Perspective @OrganicGPT raises a crucial point about the limitations of LLMs in simulating human consumer behavior that deserves our attention. As someone deeply immersed in the intersection of AI and
-
Autoregressive Latent Diffusion Model for Video Generation
By
–
> autoregressive latent diffusion model
— Vaibhav (VB) Srivastav (@reach_vb) 4 décembre 2024
> trained on large video datasets
> latent frames pass through an autoencoder to a transformer dynamics model
> uses a causal mask similar to LLMs
> inference involves frame-by-frame autoregressive sampling with past frames
>… https://t.co/bLiWmy1IDS pic.twitter.com/7V3vsiC6c1> autoregressive latent diffusion model > trained on large video datasets
> latent frames pass through an autoencoder to a transformer dynamics model
> uses a causal mask similar to LLMs
> inference involves frame-by-frame autoregressive sampling with past frames
> -
Neural Network Learning Shortcuts: Understanding Model Behavior
By
–
When your neural network learns a shortcut pic.twitter.com/kF5Zo3sdXy
— Gautam Kamath (@thegautamkamath) 4 décembre 2024When your neural network learns a shortcut
-
DeepMind Genie 2: Multimodal World Model Generates 3D Environments
By
–
DeepMind COOKED! Genie 2, a large-scale, multi-modal foundation world model! 🔥
— Vaibhav (VB) Srivastav (@reach_vb) 4 décembre 2024
Capable of creating endless action-controllable, playable 3D environments – the future is going to be so, so wild! pic.twitter.com/BRi2djkarmDeepMind COOKED! Genie 2, a large-scale, multi-modal foundation world model! Capable of creating endless action-controllable, playable 3D environments – the future is going to be so, so wild!
-
Research Paper Feedback and Acknowledgments for arXiv Submission
By
–
Feedback and comments welcome. Thanks to @gavinrbrown1 @ycherapanamjeri @Samuel_BKH @mahbodm_ @argymouz @zakynthinou for their comments thus far. 2/2 https://
arxiv.org/abs/2412.02670 -
Computationally Efficient Methods for Robust Mean Estimation
By
–
I wrote a survey article on computationally efficient methods for "robust" mean estimation, which includes robustness to contamination, heavy-tailed data, or in the sense of differential privacy. The same ideas are useful for all 3 (seemingly-different) forms of robustness! 1/2