Inference Podcast Summaries
Inference on Yedapo: 3 summarized podcast and YouTube episodes. Each includes key takeaways, core concepts and notable quotes with timestamps.
I need to rant about local models
Theo - t3․gg
Jul 7, 2026
While open-weight models like GLM-52 are revolutionary, they are not viable for consumer hardware due to massive VRAM and compute requirements. True performance requires enterprise-grade infrastructure. Instead of chasing local setups, leverage the cloud to benefit from the competitive pricing and efficiency that open-weight models introduce to the hosting ecosystem.
Key insight: Even when open-weight models appear 10x cheaper per token, they often burn 3x more tokens to complete a task, narrowing the actual cost advantage over proprietary models to a mere 2x.
The Story Behind Cerebras’ $63 Billion IPO with Founder and CEO Andrew Feldman
No Priors: AI, Machine Learning, Tech, & Startups
May 21, 2026
Andrew Feldman, CEO of Cerebras, argues that AI's true value isn't incremental improvement, but the creation of entirely new business models. Just as high-speed internet transformed Netflix from a DVD-by-mail service into a global studio, ultra-fast AI inference will fundamentally reorganize how companies operate and drive massive productivity jumps beyond simple automation.
Key insight: The market for slow inference is zero, just as the market for dial-up internet became zero; once AI is used daily, speed becomes a non-negotiable requirement for utility.
TiDAR: Think in Diffusion, Talk in Autoregression (Paper Analysis)
Yannic Kilcher
Dec 27, 2025
Tidar, a hybrid architecture from Nvidia, leverages unused GPU capacity during autoregressive inference to run parallel diffusion-based drafts. By using these drafts as speculative proposals for the autoregressive model, the system achieves a 4-6x speed boost without sacrificing the output quality of standard autoregressive decoding.
Key insight: Tidar closes the performance gap between diffusion and autoregressive models by performing parallel draft computation in the same GPU forward pass used for token verification.