X Posts
X posts in reverse chronological order. This is a backup or draft.
Added Mistral Small 2501 (API) to my MMLU benchmark testing. It achieved an accuracy of 66.0% in the college computer science subject, which is lower than Mistral Large Latest (73.00%). Both are also lower than DeepSeek v3 (78.00%) and DeepSeek R1 (87.14%).
Models without chain-of-thought capabilities are proving difficult to outperform those that have it. Blog: https://lzwjava.github.io/mmlu-en
People want to succeed, get good grades, or earn a decent amount of money. But life is actually like the crypto or stock market. The market often won’t be what you want.
Gmail could be greatly improved with AI. For example, when I send epub files to the Send-to-Kindle service and it reports a conversion error, this user-initiated action should be a high-priority trigger for me. These kinds of filters should be personalized, learning from user behavior like opens, interactions, and active triggers. Information is meant to be consumed. An RAG or LLM-driven mail app will likely become popular.
In a previous post, I wrote, “Sometimes, after sending documents to Kindle via email, it reports a problem with the sent document(s). One issue is that the document lacks a title, and to fix it, we need to pass the –metadata parameter.” However, the condition isn’t actually “sometimes.” Regardless of when, if the document lacks a title, the current send-to-Kindle service will report this error. We can use “sometimes” if the bug is related to time or highly relevant to time.
There are two options in iOS for switching between light and dark mode: sunset to sunrise and a custom schedule. But what if I close the curtains in my bedroom? To me, the whole day is night. iPhones can judge the environment to set the screen’s brightness. They should add a similar option for switching between light and dark mode appearance.
Over the past year, my blog recorded 13,500 users and 50,800 page views. However, I doubt these numbers. I had been using Firefox Focus since December 2024, and just two weeks ago, I disabled its tracking protection. If I exclude the data from December 2024, the blog attracted 9,580 users and 25,000 page views over a one-year period. Never lie to yourself!
Some of my recent blog posts. 最近的一些博客文章。
There have been many updates to my blog recently. Welcome to visit! One main change is the introduction of 7 more languages: Traditional Chinese, French, German, Spanish, Japanese, Arabic, and Hindi. This is heavily powered by the DeepSeek API.
Today, a nearby print shop is closed because the Spring Festival is approaching. I used the Baibu printing machine in my apartment complex, but the print quality is poor because the text is not continuous. I’ve started using my Kindle Scribe again for reading. I used pandoc to convert markdown files to EPUB. The command-line tool kindlegen has been replaced by kindlepreviewer, and the current version is 3.90.0. Sometimes, after sending documents to Kindle via email, it reports a problem with the sent document(s). One issue is that the document lacks a title, and to fix it, we need to pass the –metadata parameter.
Pika, https://apps.apple.com/us/app/pika-ai-video/id6680155400
What a journey. Finally sold out. $ 612 profit, thanks @realDonalodTrump
Test concurrency.
李智维 | Software Engineer (Full-Stack AI) | Read (320+ Books) | Java Spring MySQL Redis JavaScript iOS Android Vue | Azure AWS GCP Alibaba Cloud | PyTorch CUDA