Engineering brief
Krea built infra to train K2 from scratch, prioritizing metrics, checkpointing, and
At a glance
- Relevance
- Practical value
- Warnings
- None
Krea built infra to train K2 from scratch, prioritizing metrics, checkpointing, and hybrid GPU scheduling.
Potential fit for engineering leaders.
Summary
Gabriel from Krea describes the infrastructure behind training K2, a diffusion transformer trained from scratch. The core challenge wasn't model
architecture but operational reliability: as GPU count scaled, silent failures became the norm. The team learned to accept crashes as
inevitable, relying on aggressive checkpointing (every 20 minutes) and custom metrics to survive runs that rarely lasted 8 hours.
Watch the video
This video is blocked due to your privacy settings. To watch this video, please accept YouTube marketing cookies.
Related breakdowns
Agents face the same operational debt as microservices—prepare now
Navan shares hard lessons from running agents in production: runtime is solved, but cost, testing, and debugging gaps threaten every team scaling agentic AI…
Better data is the cheapest compute multiplier you're ignoring
Compute scarcity is real, but data quality is the overlooked multiplier. DatologyAI shows 100x training efficiency gains through smart curation. Engineering…
Your LLM Bill Is a Model Selection Problem
Most features don't need frontier models; evaluate SLMs with a golden dataset and prompt engineering to match quality and eliminate inference costs.
Get TL;DW
Too Long; Didn't Watch.
A concise breakdowns of the AI and devtools videos that actually matter for engineering leaders.
Free. Weekly. No hype.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.