Nebius — searchable YouTube transcripts & AI Q&A

Why a 10x Smaller Open Model Beats the Frontier

Open-source models are closing the gap to closed-source frontier models faster than expected, now within three to six months, argues Ryan Henry Wong. Inference optimization, such as pruning unimportant KV cache entries, can generate four to five times more tokens at the same GPU cost, making smaller open models competitive. He recommends startups start with an inference provider and only build in-house when scale justifies the investment.

S1E3: From Data Analysis to Dog Games, Can Our Agents Pass the Test?

The speaker argues that for simple tasks, smaller and cheaper models can perform as well as larger ones, based on a benchmark comparing three Nvidia NIM models on a data analysis task. The demo also reveals that agents can debug code effectively, as one agent identified and provided a workaround for a bug in the LangChain adapter for NIM, which was later fixed upstream. The speaker emphasizes the importance of iterating and using vision models to evaluate and improve generated games.

Getting More Runs from the GPUs You Already Own

SkyPilot's CEO argues that in the current compute-scarce environment, AI teams must adopt a multicloud strategy from day one to secure GPU capacity, and that the real challenge is efficiently utilizing that capacity once obtained. He reports that SkyPilot helps teams standardize across diverse cloud providers and improve GPU utilization by 10-30% through better scheduling and bin packing, citing a customer that achieved a 10x increase in experimentation velocity. He also advises planning for a future of abundant compute by maintaining flexibility across the entire AI stack, including model, platform, and compute layers.

How to Optimize the RL Training Loop | Alexander Golubev, Research Lead at Nebius

Asynchronous RL with in-flight updates and KV cache reuse can cut trainer step time by 43% without hurting model quality, the speaker reports. Switching from DAPO to CISPO-like objectives makes training robust to high async levels, allowing async 8 or higher without performance drops. Using FP32 precision for the language model head on the inference side also boosts benchmark scores by several points almost for free.

Can a LangChain Deep Agent direct a live race broadcast? | Nebius Token Factory

This video demonstrates building a LangChain Deep Agent to control a live race broadcast, using Nebius Token Factory and Multiviewer to show all drivers' qualifying laps.