不知不觉AI就浸入到我们生活的方方面面了，现在应该没有人从没听说过AI了吧。
印象里最早来到我们身边的是豆包，横空出世的deepseek则是让AI大模型家喻户晓。
孩子们甚是喜欢豆包，因为豆包什么都懂（笑），讲故事信手拈来，脾气又好，比大人又有耐心。
同时豆包又有一点傻萌，比如让她讲睡前故事希望她讲一个长一点的，但她总是傻乎乎的停掉，而如果让她讲10个故事，她又会敷衍的讲10个短小故事。
父母也用豆包，比起让老人主动问孩子这是不是骗人的，老人们面对豆包可放松的多，说说心里话，聊聊生活疑问，比广告满天的搜索引擎不要好用太多。
老婆用的是deepseek, 但是因为工作偏创新一点，总是感觉不够给力。
我最早用的是Grok，因为没有对香港地区限制， 而彼时chatgpt的封锁是比较严重的。后来（2026.5）开始用claude。
公司也在推AI进工作，成立了项目组开发了CodeAgent辅助开发，甚至还抽调了高级专家们专项探究，足以表达重视。
得益于公司内部搭建的minimax 2.5(后来升级到2.7)，大家终于有机会不考虑token开始猛登了，疯狂探索如何使用AI提高开发效率，包括需求内容分析，需求文档编写，开发设计文档审视，代码编写，代码审查，单元测试用例设计，集成测试用例设计，维护文档编写，维护案例总结，开发各种skills。
我们的产品主要是电信级的客服系统，工作中使用AI主要是用于编程辅助。
面对庞大的代码库（百万级别），整理一个被叫号码分析逻辑可能要一上午，现在AI十几分钟就能搞定，虽然有的时候还需要人工介入引导它，但是已经很棒了。平时工作忙碌，有些流程用的多，用一点了解一点，有了AI之后可以让它全面的总结下业务流程，AI帮着做了很多一直想做没做的事。
于是，使用AI越是频繁, 使用时间越长，越是体会到那句话： “所有的东西都值得用AI重写一遍”。
网上的说法是人类的智慧结晶已经被大模型吸取的差不多了，因此大模型未来一段时间内很难大幅度进化了。
于是，注意到有两个方向开始发展：
大模型原来只能处理文本，现在开始处理图片/音频/视频。
大模型处理复杂任务容易走偏，于是出现了harness。
1.项目工程代码量巨大,超出大模型的上下文,导致大模型的记忆出现偏差,编码思路走偏
2.如何规范的使用大模型, 使用步骤能否固定, 哪些步骤是固定的, 哪些步骤是需要人工接入的, 哪些步骤是随着项目进展不断完善的.
公司内部推广的是metaSpec和mindSpec。他们借鉴了openSpec和superpowers.
AI大模型的本质是token预测，它并不真实理解人们的语义。
当他收到你的hello时，它经过分析自己学习的内容，猜测最合适的回答是你好。
它并不能感受到你的情感和善意，也不是在对你表达善意，它的表达仅仅是一个推理。
当我们在和大模型交互的欣喜之余，意识到这一点，会感觉有点残酷。
1.大模型的工具选择应该是claude, opencode还是codex?
2.大模型能否输出ppt
3.如何让大模型在linux上工作的前提下, 能通过手机app进行下一步控制, 甚至通过手机的语音功能指挥linux上的会话？
4.个人如何具体量化感受大模型的进步？
]]>早就意识到,博客的难点在于持续的输出,而持续输出最大的敌人还不是懒惰, 而是绞尽脑汁让博客更好看一些.
最近在看到Peng Liu’s blog觉得自己博客的主题(jekyll的minimal-mistakes)实在有些简陋, 便又想开始折腾hexo+butterfly了.
最早写博客的初衷,还包括想借此练练自己的英语, 后来看, 这更加剧了我博客的简陋
还是要先把博客写好, 写的丰富, 写的真实, 然后再适当的写一些必要的英文博客, 毕竟自己能够输出的很少.
]]>sudo apt install nfs-kernel-server
sudo systemctl status nfs-server
sudo systemctl status rpcbind
/etc/exports
/share_dir 192.168.8.0/24(rw,sync,no_subtree_check)
sudo exportfs -ra # reload configuration
sudo exportfs -v # checkt current configuration in use
sudo apt install nfs-common
sudo mount -t nfs 192.168.8.42:/sharedir /u01
insmod self-defined kernel module on suse11sp3 error Unknown symbol __x86_indirect_thunk_r11 (err 0)
The self-defined kernel was compiled on suse11sp3 with kernel version 3.0.101-0.47.106 with `CONFIG_RETPOLINE=y`
When it runs on suse11sp3 with kernel version 3.0.101-0.47.90, it shows error above.
Check the configuration by `cat /boot/config-$(uname -r) | grep CONFIG_RETPOLINE`
It had to recompile on lower version(0.47.90) and it runs normally on higher version(0.47.106)
适合自托管（Self-hosting）, 指的是你自己负责运行和托管 Jekyll 生成的网站
简单化引入, 下载远程主题的文件时不会覆盖本地文件, 从而依然可以定制化.
高度定制化.
fork仓库后, 我做了哪些改动
1. _config.yml
2. _data/navigation.yml
1.如何按标签和tag归类文章
2.post的文章如何使用主题
3.分页帖子数量
4.博客正文太细长, 影响阅读体验, 如何按照屏幕自动适配?—>使用layout splash
I faced a problem that I couldn’t ping, and then I experienced policy routing.
Now the simplified scenario is as follows.
my computer A can reach subnet 172.90.185.0/24 with ip 172.136.125.70 by switch.
my computer A can reach subnet 100.61.193.0/24 with ip 100.255.200.8 by vpn.
Machine M with eth0 172.90.185.65 and eth1 100.61.193.131.
I can ping 172.90.185.65 success but cannot ping 100.61.193.131 successfully.
Machine N with eth0 100.61.193.109 and eth1 172.90.185.59.
I can ping 100.61.193.109 success but cannot ping 172.90.185.59 successfully.
It seems that I can never ping the IP address on the second network card.
run command route on M:
Kernel IP routing table
Destination Gateway Genmask Flags Metric Ref Use Iface
default 172.90.185.1 0.0.0.0 UG 0 0 0 eth0
172.90.185.0 * 255.255.255.0 U 0 0 0 eth0
100.61.193.0 * 255.255.255.0 U 0 0 0 eth1
run command route on N:
Kernel IP routing table
Destination Gateway Genmask Flags Metric Ref Use Iface
default 100.61.193.1 0.0.0.0 UG 0 0 0 eth0
172.90.185.0 * 255.255.255.0 U 0 0 0 eth1
100.61.193.0 * 255.255.255.0 U 0 0 0 eth0
Later some application add ip 100.61.193.40 on eth1:0 of M and I can ping 100.61.193.40 from A successfully.
I decided to find out.
I am using tcpdump -i any -s 0 icmp to capture packets on machine M:
When I ping 172.90.185.65 from A, the source IP of the packets is 172.136.125.70.
When I ping 100.61.193.40 from A, the source IP of the packets is 100.255.200.8.
ipUsing ip rule list all to view all rules, I found that 100.61.193.40 has special handling and will use routing table 239.
32755: from 100.61.193.40 lookup 239
Run command ip route list table 239 and view the processing action of route table 239:
default via 100.61.193.1 dev eth1
Make the similiar rule and route for 100.61.193.131
ip rule add from 100.61.193.131 table 230
ip route add default via 100.61.193.1 dev eth1 table 230
then ping 100.61.193.131 from A successed and ping 100.255.200.8 -I 100.61.193.131 also succeed.
Running ping 100.255.200.8 from M failed and running ping 100.255.200.8 -I eth1 from M also failed.
Add route to table main ip route add 100.255.200.8/32 via 100.61.193.1 dev eth1 and it succeed.
bind: Berkeley Internet Name Domain
sudo apt-get install bind
systemctl status named
systemctl restart named
target domain: goodluck.com
target ip: 192.168.23.62
dns machine: 192.168.23.52
sudo vi /etc/bind/named.conf.local
zone "goodluck.com" {
type master;
file "/etc/bind/db.goodluck.com";
};
sudo vi db.goodluck.com
;
; BIND data file for local loopback interface
;
$TTL 604800
@ IN SOA ns.goodluck.com. admin.goodluck.com. (
2024120421 ; Serial
604800 ; Refresh
86400 ; Retry
2419200 ; Expire
604800 ) ; Negative Cache TTL
;
@ IN NS localhost.
@ IN A 192.168.23.62
www IN A 192.168.23.62
add “nameserver 192.168.8.52” in /etc/resolv.conf of client machine
dig goodluck.com
; <<>> DiG 9.18.28-0ubuntu0.24.04.1-Ubuntu <<>> goodluck.com
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 61902
;; flags: qr aa rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 0, ADDITIONAL: 1
;; OPT PSEUDOSECTION:
; EDNS: version: 0, flags:; udp: 1232
; COOKIE: 62c10e7fbcc0a2ae0100000067505f6771e5f0d80e77c012 (good)
;; QUESTION SECTION:
;goodluck.com. IN A
;; ANSWER SECTION:
goodluck.com. 604800 IN A 192.168.23.62
;; Query time: 0 msec
;; SERVER: 192.168.23.52#53(192.168.23.52) (UDP)
;; WHEN: Wed Dec 04 21:55:58 HKT 2024
;; MSG SIZE rcvd: 85
dig www.goodluck.com
; <<>> DiG 9.18.28-0ubuntu0.24.04.1-Ubuntu <<>> www.goodluck.com
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 17992
;; flags: qr aa rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 0, ADDITIONAL: 1
;; OPT PSEUDOSECTION:
; EDNS: version: 0, flags:; udp: 1232
; COOKIE: 95fc1427421e63890100000067505fa1d9ec1743c9ca8865 (good)
;; QUESTION SECTION:
;www.goodluck.com. IN A
;; ANSWER SECTION:
www.goodluck.com. 604800 IN A 192.168.23.62
;; Query time: 1 msec
;; SERVER: 192.168.23.52#53(192.168.23.52) (UDP)
;; WHEN: Wed Dec 04 21:56:56 HKT 2024
;; MSG SIZE rcvd: 89
ping goodluck.com
PING goodluck.com (192.168.23.62) 56(84) bytes of data.
64 bytes from 192.168.23.62: icmp_seq=1 ttl=64 time=0.018 ms
ping www.goodluck.com
PING www.goodluck.com (192.168.23.62) 56(84) bytes of data.
64 bytes from 192.168.23.62: icmp_seq=1 ttl=64 time=0.018 m
pending
]]>本文仅作分享,不做讨论. 我在2023.07去香港买了Redmi Note13 pro+ .
安卓机器太老了, 导航时卡顿严重, 查资料切换多个app时反应慢, 同时充电只有22W.
se的电池不耐用且充电慢, 基本一天一充, 而且屏幕比较小, 看视频不是很爽.
在家软路由在公司专线, 手机上只用流量学习科学知识.
1> cmake .. -DCMAKE_VERBOSE_MAKEFILE=ON
This option tells CMake to generate a verbose Makefile.
When you run make, it will display each compilation and linking step, which is very useful for debugging the build process.
2> cmake .. –trace-expand
This option enables tracing, which means that every command executed by CMake during configuration will be printed, along with all relevant variables expanded to their actual values at that moment.
This is particularly useful for debugging complex configurations and understanding how variables affect the build process.
3> cmake .. -DCMAKE_BUILD_TYPE=type
This option sets the build type, which can affect compiler flags and optimizations. Common values for type include:
(1)Debug: Enables debugging information and disables optimizations. (maybe -g)
(2)Release: Enables optimizations and disables debugging information. (maybe -O3 -DNDEBUG)
(3)RelWithDebInfo: Combines optimization with debugging information. (maybe -O2 -g -DNDEBUG)
(4)MinSizeRel: Optimizes for minimum size. (maybe -Os -DNDEBUG)