Source item

1/ people inbound us all the time about rapid github star growth. I love oss/infra (investor in Docker, LangChain, Baseten etc). I pulled the distribution (all of github, all time). the shortest time to reach the top just went from years to ~an hour! https://t.co/BqGvY2fv5K 2/ record time for a breakout repo to hit ~20k stars: 2013-2022: years AutoGPT (2023): ~13 days Grok-1 (2024): ~1 day DeepSeek Harness (last week): ~1 hour (all AI projects, ofc) 3/ it’s not just the obvious (denominator growth). github grew from 100M to 180M developers in three years, vs ~300x change in velocity 4/ the all-time top 20 barely moved for a decade. today, above the linux kernel (242k stars): - superpowers at 273k (a claude code skills library) - ECC at 239k (agent harness tooling) much of the new T20 is < 1yo 5/ a nice data point is caveman, a token-compression skill. ~600 stars in early april. ~99k today, ahead of bitcoin core! @julius_brussee https://t.co/kagHw2GvHF @julius_brussee 6/ the new popular kids are mostly harnesses, skills libraries, memory layers, context tools @julius_brussee 7/ part of the acceleration is gaming / a known integrity problem. CMU and socket researchers found 4.5M suspected fake stars, with campaigns up ~100x in 2024. separate from the projects above, but relevant for diligence :) @julius_brussee 8/ new question: agents hold github credentials now. should they star? lots of READMEs ask for a star (caveman’s: “star cost zero. fair trade”). when the “reader” is an agent with a gh token, a README is a prompt. documentation converts its own reader! Cool! But a diff signal @julius_brussee 9/ star velocity now measures reach through one dense mixed attention network, human and agent where’s the new signal? contributor retention, forks that receive commits, mentions in other people’s repos, registry downloads net of mirrors, social love. slow/$$$ to fake/prompt @julius_brussee 10/10 GitHub must change (or be replaced)! stars are backlinks in 1997 and waiting for pagerank: maybe rank repos by who depends on them? AEO for agent tech choices? Anonymized private dependency analysis? What other ideas? Long live code reuse