GPU Scheduling, Self-Critique, and Code for AI Gadgets

AI21’s 10,000-GPU deployment of Kueue offers a concrete infrastructure example, while research on multimodal self-critique explores model improvement. Elsewhere, Meta releases gadget-building code and Example.com updates its design.

Scheduling Work Across 10,000 GPUs

AI21 has implemented Kueue to manage a fleet of 10,000 GPUs, using it to allocate resources and distribute workloads for training and inference. The operational focus is on coordinating a large pool of compute. AI21 describes improved performance from that allocation work, making scheduling a concrete part of its approach to running AI workloads at scale.

Models Learn From Their Critiques

Research on multimodal models learning from their own critiques reports performance improvements across multiple tasks. The approach gives self-assessment a role in refining model outputs, suggesting a route to improvement through the models’ own feedback. Its scope remains a research result: the reported gains concern the tasks studied.

Code for Building Muse Gadgets

Meta has released open-source code for creating Muse AI gadgets, giving developers a starting point for building their own devices. The concrete development is access to the code and the ability to build with it. A wider range of applications is a possible consequence of that access, rather than an outcome established by the release.

Example.com Refreshes Its Web Design

Example.com has launched its most substantial redesign in decades, with the stated aim of improving user experience and engagement. That makes it a redesign to examine for its usability choices; whether those changes improve engagement remains an intended benefit.