DeepSeek - Correcting the Record

Addressing false information about the new DeepSeek V3/R1 model

By Request Only

Inference Compute Scaling

Four inference time compute multipliers that are likely to lead to inference dominating training compute in the near future

Available

NVLink & Beyond Blackwell

NVLink, its optimization in the NVL72, and improvement vectors post-Blackwell

Available

CUDA - Is It a Moat?

A discussion of the consensus view that CUDA is a moat and some dynamics that that consensus is ignoring

Available

Grok 3 Reasoning Confusion

A response to the confusion around the benchmarking of Grok 3 Reasoning and comparisons to o3-mini

Available

Improving Training Data

A detailed discussion of data augmentation techniques that have improved data for training LLMs

Available

Potential Impact of ASICs in LLM Compute

The overstated potential impact of ASICs in LLM compute, and tensor cores as ASIC-like components

Available

Effective LLM Application Design

How to use enterprise data to design LLM systems, and the downsides of chatbots, agents and fine-tuning.

Available