InFrom Tokens to AgentsbyVaibhav Kumar·Oct 31How to Design an LLM Inference Engine at Scale: Lessons from vLLMA trained model is not a product. The weights can write fluent text, but the moment 500 people want answers at the same time, on a handful…A response icon1A response icon1
InFrom Tokens to AgentsbyVaibhav Kumar·Oct 28How Attention Works in LLMs: Queries, Keys, Values ExplainedIn Part 3 we gave every token a row in the embedding table, and hit a problem: the row for “bank” is identical whether the sentence is…
InFrom Tokens to AgentsbyVaibhav Kumar·Oct 24How a Model Writes: One Token at a TimeIn Part 4 we saw two odd facts about pricing: output tokens cost three to six times more than input tokens, and cached input can cost as…A response icon1A response icon1
InOperator’s LogbyVaibhav Kumar·Mar 10Episode 3-Informers: The Eyes and Ears of Your OperatorIn the previous episode, we explored controller-runtime, the runtime engine that powers Kubernetes operators. We saw how the Manager…
InFrom Tokens to AgentsbyVaibhav Kumar·Mar 9Tokenization in Practice (Part 4): Why Every Token Costs You MoneyIn Parts 1–3, we built up the theory: what tokens are, how algorithms like BPE split text into subwords, and how embeddings give those…
InFrom Tokens to AgentsbyVaibhav Kumar·Mar 8Embeddings (Part 3): How Words Become NumbersIn Parts 1 and 2, we turned raw text into token IDs — integer codes like [15496, 995, 318]. But a neural network can’t do much with bare…
InFrom Tokens to AgentsbyVaibhav Kumar·Mar 8Tokenization (Part 2): The Algorithms Behind Every TokenIn Part 1, we discovered that words are too messy, characters are too small, and morphemes are too hard to define cross-lingually. The…
InFrom Tokens to AgentsbyVaibhav Kumar·Mar 7Tokenization (Part 1): The Search for the Right TokenWhen you type “unhappiness” into ChatGPT, it doesn’t see one word — it sees three pieces: “un”, “happi”, and “ness”. Why? Because large…
InOperator’s LogbyVaibhav Kumar·Jan 11Episode 2- Controller-Runtime, The Engine Behind Kubernetes OperatorsThe Controller: The Beating Heart of an Operator
Vaibhav Kumar·Jan 7What Psychology Speaks: Motivation Is Not WillpowerWhy discipline, rewards, and success often fail us