会不会出现。 ⼀位AI⼤模型领域的开发者
Greg
Kamradt
也对此充满了困惑。于是，他设计了⼀个名为“⼤海捞 针”的⼤模型⻓⽂本性能测试⽅法： 在⽂本语料中藏⼊⼀个与⽂本语料不相关的句⼦（可以想象是在整本《西游记》⾥放⼊⼀句只会在 《红楼梦》⾥出现的话），然后看⼤模型能不能通过⾃然语⾔提问的⽅式（Prompt）把这句话准确地 提取出来。 Greg
Kamradt
的“⼤海捞针”实验简述：
y.
Greg
Kamradt
把藏起来的那句话（也就是⼤海捞针的“针”）分别放到了⽂本语料（也就是⼤海捞针 的“⼤海”）从前到后的15处不同位置，然后针对从
1K
到
128K（200K）
等量分布的
15
种不同⻓度 的语料进⾏了
225
次（15×15）实验。为了保证实验的严谨，他还重复进⾏了实验，来取平均值，测 试时调⽤
API
的费⽤总计花了⼏百美元。
Greg
Kamra 来源:
https://twitter.com/GregKamradt/status/1727018183608193393
第⼀次实验 作为⼤模型⻓⽂本技术的重要玩家，Greg
Kamradt
的实验成功引起了我们的注意。由于
Greg
Kamradt
将他的实验所有语料和代码都放在了
Github
上，Moonshot
AI
的⼯程师很容易就可以在
Kimi
智能助⼿
上重复做这个实验。

2003, 2004 Philip Warner. * Portions relating to PNG copyright 1999, 2000, 2001, 2002, 2003, 2004 Greg Roelofs. * Portions relating to gdttf.c copyright 1999, 2000, 2001, 2002, 2003, 2004 John Ellson of Contributing Authors: John Bowler Kevin Bracey Sam Bushell Magnus Holmgren Greg Roelofs Tom Tanner Some files in the "scripts" directory have other copyright owners, but are * Cosmin Truta * Dave Martindale * Eric S. Raymond * Gilles Vollant * Glenn Randers-Pehrson * Greg Roelofs * Guy Eric Schalnat * James Yu * John Bowler * Kevin Bracey * Magnus Holmgren * Mandar

jsonb, tsquery, ltxtquery and query_int (Noah Misch) • Fix handling of DOW and DOY in datetime input (Greg Stark) These tokens aren’t meant to be used in datetime values, but previously they resulted in opaque Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Prevent foreign tables from being rera) • Reduce indentation in rule/view dumps to improve readability and avoid excessive whitespace (Greg Stark, Tom Lane) This change reduces the amount of indentation applied to nested constructs, including

jsonb, tsquery, ltxtquery and query_int (Noah Misch) • Fix handling of DOW and DOY in datetime input (Greg Stark) These tokens aren’t meant to be used in datetime values, but previously they resulted in opaque Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Prevent foreign tables from being Herrera) • Reduce indentation in rule/view dumps to improve readability and avoid excessive whitespace (Greg Stark, Tom Lane) This change reduces the amount of indentation applied to nested constructs, including

jsonb, tsquery, ltxtquery and query_int (Noah Misch) • Fix handling of DOW and DOY in datetime input (Greg Stark) These tokens aren’t meant to be used in datetime values, but previously they resulted in opaque Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Prevent foreign tables from being Notes • Fix possible mis-replay of WAL records when some segments of a relation aren’t full size (Greg Stark, Tom Lane) The WAL update could be applied to the wrong page, potentially many pages past where

jsonb, tsquery, ltxtquery and query_int (Noah Misch) • Fix handling of DOW and DOY in datetime input (Greg Stark) These tokens aren’t meant to be used in datetime values, but previously they resulted in opaque Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Prevent foreign tables from being (CVE-2014-0067) • Fix possible mis-replay of WAL records when some segments of a relation aren’t full size (Greg Stark, Tom Lane) The WAL update could be applied to the wrong page, potentially many pages past where

jsonb, tsquery, ltxtquery and query_int (Noah Misch) • Fix handling of DOW and DOY in datetime input (Greg Stark) These tokens aren’t meant to be used in datetime values, but previously they resulted in opaque Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Prevent foreign tables from being (CVE-2014-0067) • Fix possible mis-replay of WAL records when some segments of a relation aren’t full size (Greg Stark, Tom Lane) The WAL update could be applied to the wrong page, potentially many pages past where

jsonb, tsquery, ltxtquery and query_int (Noah Misch) • Fix handling of DOW and DOY in datetime input (Greg Stark) These tokens aren’t meant to be used in datetime values, but previously they resulted in opaque Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Prevent foreign tables from being (CVE-2014-0067) • Fix possible mis-replay of WAL records when some segments of a relation aren’t full size (Greg Stark, Tom Lane) The WAL update could be applied to the wrong page, potentially many pages past where

Lane) • Fix session-lifespan memory leaks in regular-expression processing (Tom Lane, Arthur O’Dwyer, Greg Stark) • Fix data encoding error in hungarian.stop (Tom Lane) • Fix liveness checks for rows that (CVE-2014-0067) • Fix possible mis-replay of WAL records when some segments of a relation aren’t full size (Greg Stark, Tom Lane) The WAL update could be applied to the wrong page, potentially many pages past where triggers into the right order. • Avoid floating-point underflow while tracking buffer allocation rate (Greg Matthews) While harmless in itself, on certain platforms this would result in annoying kernel log