Top all-time · #12570
Fergus Finn
Personal blog · fergusfinn.com/blogaisystems
Fergus Finn writes Fergus's blog, a personal blog at fergusfinn.com about ai and systems. Since 2025, 6 different Hacker News readers have linked it 27 times, most often “70x faster cold(ish) starts for SGLang”. It is still writing: the last post was on 2026-08-28, and you can follow it by feed.
Too few sampled comments (1) to say how often HN recommends it.?
Why HN links it
HN readers link this post on the economics of speculative decoding, an LLM inference technique. The only comment is a bare link with a short note on DeepSeek's shared expert, so no judgement of the writing is given.
Summarised by Claude from a sample of 1 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
A blog about LLM inference
Most-linked posts?
- 70x faster cold(ish) starts for SGLang3 people · 3 mentions · first linked 2026-04-24
- InfiniBand, RoCE, and All That3 people · 3 mentions · first linked 2026-06-19
- How fast can an LLM go?3 people · 3 mentions · first linked 2025-10-30
- Redundant Information in LLM Weights2 people · 2 mentions · first linked 2026-05-05
- Speculative KV coding: ~4× losslessly compressed KV cache using a small model2 people · 2 mentions · first linked 2026-05-12
- Parallel Primitives for Multi-Agent Workflows2 people · 2 mentions · first linked 2026-01-05
What Hacker News says?
https://fergusfinn.com/blog/economics-of-speculative-decodin... good point tho - plus for Deepseek the shared expert increases the overlap slightly
Comment linking /blog/economics-of-speculative-decoding/
on “The Economics of Speculative Decoding”Submitted What happens when you run a CUDA kernel? as a story · 9 points
HN discussionSubmitted Bringing Up DeepSeek-V4-Flash on AMD MI300X as a story · 3 points
HN discussionSubmitted Redundant Information in LLM Weights as a story · 3 points
HN discussionSubmitted InfiniBand, RoCE, and All That as a story · 2 points
HN discussionSubmitted InfiniBand, RoCE, and All That as a story · 2 points
HN discussionSubmitted UCCL-EP: DeepEP-style expert parallelism on any NIC, no GPU-initiated comms as a story · 2 points
HN discussionSubmitted Anatomy of a high-performance EP kernel as a story · 2 points
HN discussionSubmitted Speculative KV coding: losslessly compressing KV cache by up to ~4× as a story · 2 points
HN discussionSubmitted Speculative KV coding: ~4× losslessly compressed KV cache using a small model as a story · 2 points
HN discussionSubmitted How fast can an LLM go? as a story · 2 points
HN discussionSubmitted How fast can an LLM go? as a story · 2 points
HN discussionSubmitted How fast can an LLM go? as a story · 2 points
HN discussionSubmitted Adaptive speculative decoding: picking draft lengths at runtime as a story · 1 points
HN discussionSubmitted InfiniBand, RoCE, and All That as a story · 1 points
HN discussionSubmitted The Economics of Speculative Decoding as a story · 1 points
HN discussionSubmitted 70x faster cold(ish) starts for SGLang as a story · 1 points
HN discussionSubmitted Pushing memory bound CUDA kernels past the speed of light with data compression as a story · 1 points
HN discussionSubmitted In search of wasted bits: how much information do LLM weights carry? as a story · 1 points
HN discussionSubmitted Tans: Precomputing RANS as a story · 1 points
HN discussionSubmitted Also-RANS: Asymmetric Numeral Systems for Entropy Coding as a story · 1 points
HN discussionSubmitted 70x faster cold(ish) starts for SGLang as a story · 1 points
HN discussionSubmitted 70x faster cold(ish) starts for SGLang as a story · 1 points
HN discussionSubmitted Parallel Primitives for Multi-Agent Workflows as a story · 1 points
HN discussionSubmitted LLM powered data structures: A lock-free binary search tree as a story · 1 points
HN discussionSubmitted Parallel Primitives for Multi-Agent Workflows as a story · 1 points
HN discussionSubmitted Scheduling in LLM Inference as a story · 1 points
HN discussion
