Uploaded June 2026 | Updated November 2026, 2 days ago
AI models are evolving faster than ever. But as models grow, so does the pressure on the infrastructure behind them.
For AI teams, storage isn't just where data lives—it's what determines how quickly researchers can experiment, train, and ship new models.
Join Sidharth Bajaj and Venkatraghavan Srinivasa from Meta as they share how Meta evolved its storage architecture to tackle two of AI's biggest infrastructure challenges: maximizing GPU utilization and accelerating research velocity.
Learn how storage systems are being redesigned for the next generation of AI at scale at Systems & Reliability in Bellevue, WA.
Register today at atscaleconference.com
AI models are evolving faster than ever. But as models grow, so does the pressure on the infrastructure behind them.
For AI teams, storage isn't just where data lives—it's what determines how quickly researchers can experiment, train, and ship new models.
Join Sidharth Bajaj and Venkatraghavan Srinivasa from Meta as they share how Meta evolved its storage architecture to tackle two of AI's biggest infrastructure challenges: maximizing GPU utilization and accelerating research velocity.
Learn how storage systems are being redesigned for the next generation of AI at scale at Systems & Reliability in Bellevue, WA.
Register today at atscaleconference.com




![Ship Fast, Break Nothing: How AI Tests What AI Builds | Sophie Zeng & Sreeja Thummala from Meta
Meta engineers Suja and Sophie share how theyve rebuilt software testing for the AI era, moving beyond traditional landed tests that only catch known bugs to a new system of disposable, AI-generated tests that target specific code changes.
When a risky change appears, their system scores diffs for risk factors like blast radius and intent mismatch, then generates property-based or mutation-based tests designed to fail only when that specific change introduces defects [6:09] [6:45]. Context is critical: providing the generator with code change details and author intent roughly doubled catch rates at each step.
A key innovation is the two-stage architecture: an aggressive generator proposes many hypotheses, while an LLM judge filters out intended behavioral changes to protect engineer attention, discarding the vast majority of generated tests before they reach humans.This reduced engineer time spent chasing bugs by 70% despite not finding fewer actual bugs.
The system has prevented 74 documented safety incidents and bent privacy incident growth to one-fifth of forecasted levels through counterfactual measurement .
The talk closes with three ongoing shifts: tests becoming disposable, solving the long-term trust challenge, and moving testing left from review to co-authorship during code writing. Ship Fast, Break Nothing: How AI Tests What AI Builds | Sophie Zeng & Sreeja Thummala from Meta](https://i.ytimg.com/vi/Pf10kLLcFR0/mqdefault.jpg)


![Evolving GenAI Media Infrastructure Deployments | Rushaan Mahajan, Sima Labs
This presentation by Rushaan Mahajan from Sima Labs explores how to optimize video generation infrastructure to meet the growing demand for personalized, high-quality video experiences.
Video generation is a computationally intensive process that requires significantly more resources than text generation, leading to high latency and costs. [00:35]
The key challenges in video inference include the iterative denoising loop, the memory and bandwidth bottleneck in the VAE decoder, and the need to optimize the entire runtime stack to achieve real-time, high-fidelity, and personalized video generation at scale.
Optimizing the balance between the VAE compression ratio and the denoiser complexity is crucial to reducing the overall computational cost of the video generation pipeline. [09:52]
Techniques like latent space compression, sampling optimization, caching, and pruning can significantly improve the runtime efficiency of video diffusion models without compromising quality. [11:39]
A multi-GPU strategy that utilizes different GPU types for different tasks (base generation, super-resolution, personalization) can help scale video generation without proportional cost increases. [13:35]
The goal is to make personalized video generation truly instant, transitioning it from a compute-bound novelty to a mainstream, interactive, and globally relevant platform. [13:57] Evolving GenAI Media Infrastructure Deployments | Rushaan Mahajan, Sima Labs](https://i.ytimg.com/vi/PrZpl4w1Lxk/mqdefault.jpg)


