RAG doesn’t require the model to memorize an enterprise’s knowledge; it retrieves the right knowledge at inference time, while embeddings and rerankers determine whether the model gets the right context in the first place. Today we move up the stack and answer a question that appears constantly in enterprise AI. Link to post in comments 👇

I’ve started thinking about quantization less as model compression and more as a capacity lever. If you can reduce the memory footprint while preserving acceptable quality, that potentially changes how much expensive GPU infrastructure the workload requires. link to post in comments 👇

One observations about inference economics is that some of the biggest gains aren’t necessarily about cheaper GPUs. They’re about eliminating unnecessary computation and extracting more useful work from the compute you’re already paying for. Batching gets more useful work from each GPU; caching avoids making the GPU repeat work it has already done. Both turn serving architecture directly into inference economics. What if you could make a model require substantially less expensive compute, but doing so might slightly change its quality? full post in comments 👇