<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>AI Tools on Nansen Li's Blog
李楠森的博客</title><link>https://nansenli.com/tags/ai-tools/</link><description>Recent content in AI Tools on Nansen Li's Blog
李楠森的博客</description><generator>Hugo -- gohugo.io</generator><language>en</language><copyright>Nansen Li 🌈 （李楠森）</copyright><lastBuildDate>Sat, 07 Feb 2026 16:52:42 +0000</lastBuildDate><atom:link href="https://nansenli.com/tags/ai-tools/index.xml" rel="self" type="application/rss+xml"/><item><title>Openclaw Was Too Buggy, So I Built My Own Bot: Nagobot</title><link>https://nansenli.com/post/2026/02/nagobot-launch/</link><pubDate>Sat, 07 Feb 2026 16:52:42 +0000</pubDate><guid>https://nansenli.com/post/2026/02/nagobot-launch/</guid><description>&lt;p>I had been using Openclaw for a while, but the bug rate was just too high. Flows would randomly break, and debugging cost more time than it should.&lt;/p>
&lt;p>On top of that, token cost was not great. Compared with options like DeepSeek and Kimi K2.5, the price-performance ratio felt weak.&lt;/p>
&lt;p>I then tried nanobot, hoping to migrate. But support for newer models was still limited, and K2.5 kept failing with &amp;ldquo;tool execution failed.&amp;rdquo; After several rounds of trial and error, I stopped waiting for a perfect off-the-shelf solution and decided to build one myself.&lt;/p>
&lt;p>After hundreds of back-and-forths with Opus and Codex on architecture choices, I built a new bot in Golang: &lt;strong>Nagobot&lt;/strong>. At least now I fully control the provider layer.&lt;/p>
&lt;p>Project link: &lt;a class="link" href="https://github.com/linanwx/nagobot" target="_blank" rel="noopener"
>https://github.com/linanwx/nagobot&lt;/a>&lt;/p>
&lt;p>I learned a lot during implementation, especially around:&lt;/p>
&lt;ul>
&lt;li>Threads and concurrent task execution&lt;/li>
&lt;li>Agent and Skill abstractions&lt;/li>
&lt;li>Cron-based asynchronous tasks&lt;/li>
&lt;li>Session history and context compression&lt;/li>
&lt;/ul>
&lt;p>To be honest, a lot of the code still needs deeper review. Most of the work was built in rapid loops with AI-assisted architecture discussions. Sometimes frustrating, sometimes genuinely impressive, and overall very interesting.&lt;/p>
&lt;p>So far, testing results are better than expected. The bot can already create async tasks, investigate news, and set reminders for itself, which is immediately useful in day-to-day workflows.&lt;/p>
&lt;p>Most core features are now in place. My next focus is long-term memory design.&lt;/p>
&lt;p>If this sounds useful to you, feel free to open an issue or reach out. I will keep iterating on this project.&lt;/p></description></item><item><title>Why I Stopped Subscribing to ChatGPT</title><link>https://nansenli.com/post/2025/11/ai2025/</link><pubDate>Mon, 24 Nov 2025 23:13:25 +0000</pubDate><guid>https://nansenli.com/post/2025/11/ai2025/</guid><description>&lt;p>I recently stopped subscribing to ChatGPT.&lt;/p>
&lt;p>This wasn&amp;rsquo;t an easy decision. I&amp;rsquo;ve been paying since the day ChatGPT started supporting paid subscriptions. This subscription lasted for nearly 3 years.&lt;/p>
&lt;h1 id="reasons-for-stopping-the-subscription">Reasons for Stopping the Subscription
&lt;/h1>&lt;ol>
&lt;li>&lt;strong>Huge delays&lt;/strong>. I often wait dozens of seconds for a response with no reply. However, the free version usually uses mini models, which are sufficient for most tasks. But the free version is actually faster. I don&amp;rsquo;t understand - is this because there are no compute nodes or forwarding nodes in Europe?&lt;/li>
&lt;li>&lt;strong>Too many bugs on Mac&lt;/strong>. Lagging, high CPU usage, bugs when copying LaTeX text.&lt;/li>
&lt;li>&lt;strong>Poor interaction experience&lt;/strong>. For example, we might need to use Thinking again to regenerate responses. However, this option is in the model switching menu, requiring an extra click because it&amp;rsquo;s in a secondary menu. Meanwhile, the less frequently used &amp;ldquo;more concise/more detailed&amp;rdquo; regenerate buttons are in the primary regenerate menu.&lt;/li>
&lt;li>&lt;strong>Product updates change too drastically, model personality changes too quickly&lt;/strong>. This requires the cost of re-familiarizing with the model.&lt;/li>
&lt;li>&lt;strong>The model is stubborn and doesn&amp;rsquo;t follow instructions&lt;/strong>. For example, repeatedly needing to emphasize when to search and when not to ask follow-up questions.&lt;/li>
&lt;li>&lt;strong>Unstable model behavior that&amp;rsquo;s hard to control&lt;/strong>. For example, in instant mode, after the model calls search, it switches to a smaller model and uses web resources to answer questions, rather than the instant model.&lt;/li>
&lt;li>&lt;strong>Model quality is inferior to other models, even some open-source free models like DeepSeek&lt;/strong>.&lt;/li>
&lt;/ol>
&lt;p>Reasons that still make me want to stay:&lt;/p>
&lt;ol>
&lt;li>&lt;strong>Thinking mode is still the king&lt;/strong>. Continuously calling tool chains for search and verification is still a top performer. But Grok is also very good at search, and Grok&amp;rsquo;s free quota is already sufficient.&lt;/li>
&lt;li>&lt;strong>More powerful product matrix&lt;/strong>, such as ChatGPT Atlas browser, or Mac&amp;rsquo;s standalone client. Plus many plugins. However, it&amp;rsquo;s concerning that the Atlas browser&amp;rsquo;s Agent quota is surprisingly low, almost unusable for practical purposes.&lt;/li>
&lt;/ol>
&lt;h1 id="current-alternatives">Current Alternatives
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/ai2025-apps.png"
loading="lazy"
alt="AI apps on my phone"
>&lt;/p>
&lt;p>I&amp;rsquo;ve tried numerous free software, including all conceivable/most LLMs, even models developed by teams like MiniMax/Kimi. I&amp;rsquo;ve summarized my findings. Before summarizing, I need to define a software industry term. This phenomenon is very characteristic of China, or regional characteristics.&lt;/p>
&lt;p>Current LLM software indeed shows a trend of sacrificing tool purity and user experience in pursuit of daily active users (DAU), retention rates, and monetization. Some LLM software is stuffed with many operational elements, such as trendy operational topics, or many things not very related to technology, such as agents, competition zones, headline copywriting, or Xiaohongshu copywriting. This phenomenon indicates software bloat, over-operationalization, de-tooling, vulgarization, and reduced signal-to-noise ratio. In this article, we call this phenomenon &lt;strong>platform decay / enshittification&lt;/strong>.&lt;/p>
&lt;ul>
&lt;li>&lt;strong>Zhipu Qingyan (智谱清言)&lt;/strong>: Zhipu Qingyan&amp;rsquo;s GLM4.6 is a very powerful model, especially its Agent capabilities. However, its client is severely enshittified, with the homepage immediately showing enshittified content like &amp;ldquo;Love Strategist&amp;rdquo; and &amp;ldquo;Dark Cuisine Challenge&amp;rdquo;. Not recommended.&lt;/li>
&lt;li>&lt;strong>Yunbao&lt;/strong>: Yunbao is Tencent&amp;rsquo;s software, supporting Tencent&amp;rsquo;s self-developed models and DeepSeek, but model capabilities are average. The software shows mild enshittification, such as embedding Tencent game Honor of Kings content. Not recommended.&lt;/li>
&lt;li>&lt;strong>Gemini&lt;/strong>: The new king. Gemini 3 Pro model performance is outstanding. Nano Banana Pro&amp;rsquo;s image generation capability is excellent. Suitable for work and life use. Operated by Google, overall experience is good. However, recently due to high traffic, the service is not particularly stable. Recently, traffic has been too high, service degradation is severe, and quantized services have started to be provided, so quality has declined. Long thinking with multiple tool calls is not as good as GPT5-Thinking or Grok Expert. The software has almost no enshittification. &lt;strong>Highly recommended&lt;/strong>.&lt;/li>
&lt;li>&lt;strong>Grok&lt;/strong>: The free king. Powerful search capabilities, both short and long thinking produce high-quality results. Excellent at mining network information. Free Expert version is sufficient for general users. Some operational content, such as AI virtual characters, but enshittification is very mild. &lt;strong>Recommended&lt;/strong>.&lt;/li>
&lt;li>&lt;strong>ChatGPT&lt;/strong>: Gradually stepping down from the throne. Many features are still only available in ChatGPT, such as voice conversations. Large user base. Free version is sufficient. Many users still have inertia/dependency. Can keep on phone.&lt;/li>
&lt;li>&lt;strong>DeepSeek&lt;/strong>: The stable king. Fewest features, completely no enshittification, the cleanest application, extremely stable service, extremely stable model style, and quite stable intelligence. &lt;strong>Recommended to keep on phone&lt;/strong>.&lt;/li>
&lt;li>&lt;strong>Kimi&lt;/strong>: Model capability is average. Advertised Agent capabilities are strong, but actual experience shows heavy templating with lots of prompt engineering traces. Not recommended. Software has no enshittification, which is a plus.&lt;/li>
&lt;li>&lt;strong>Claude&lt;/strong>: The god of code. Client performance is average, especially search is not as aggressive as Grok and ChatGPT thinking. Still suitable for work use. No enshittification. Can keep on phone, using free quota. It&amp;rsquo;s worth noting that the company has discriminatory regional remarks, which is concerning.&lt;/li>
&lt;li>&lt;strong>Dola&lt;/strong>: A client produced by Douyin (TikTok). Model capability is average. Voice recognition capability is excellent. Severely enshittified. Not recommended.&lt;/li>
&lt;li>&lt;strong>Copilot&lt;/strong>: A product relying on Microsoft search and ChatGPT, the software is mediocre. Not recommended.&lt;/li>
&lt;li>&lt;strong>Manus&lt;/strong>: Early reliance on viral marketing affects the team&amp;rsquo;s long-term healthy development. This product will likely die a slow death in the medium to long term. Although it has some Agent capabilities, not recommended for long-term use.&lt;/li>
&lt;li>&lt;strong>Perplexity&lt;/strong>: Model capability is average. Software usage logic is slightly confusing. Management team&amp;rsquo;s discriminatory remarks are concerning. Not recommended.&lt;/li>
&lt;li>&lt;strong>MiniMax&lt;/strong>: Model capability is excellent, especially the Minimax M2 model, with good Agent capabilities. However, software experience is poor, even basic copy and paste functions are missing. Not recommended.&lt;/li>
&lt;li>&lt;strong>Tongyi (通义)&lt;/strong>: Only available for download in China. Software is heavily enshittified. Opening it reveals a screen full of operational content that&amp;rsquo;s suffocating, making it impossible to continue using. &lt;strong>Recommend immediate uninstallation&lt;/strong>.&lt;/li>
&lt;li>&lt;strong>Qwen Chat&lt;/strong>: Currently still not available on iOS platform. It&amp;rsquo;s a de-enshittified version of Tongyi, supporting Qwen large models. If it can be listed, it&amp;rsquo;s recommended.&lt;/li>
&lt;/ul>
&lt;p>Summary:&lt;/p>
&lt;table>
&lt;thead>
&lt;tr>
&lt;th>&lt;strong>Software Name&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Model Capability&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Enshittification&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Advantages&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Disadvantages&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Recommendation&lt;/strong>&lt;/th>
&lt;/tr>
&lt;/thead>
&lt;tbody>
&lt;tr>
&lt;td>Z-ai (Chinese version)&lt;/td>
&lt;td>Strong&lt;/td>
&lt;td>Severe&lt;/td>
&lt;td>Strong Agent capability&lt;/td>
&lt;td>Client enshittified&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Yunbao&lt;/td>
&lt;td>Average&lt;/td>
&lt;td>Mild&lt;/td>
&lt;td>Supports multiple models&lt;/td>
&lt;td>Embedded game content&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Gemini&lt;/td>
&lt;td>Excellent&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Outstanding performance, good experience&lt;/td>
&lt;td>Unstable service&lt;/td>
&lt;td>&lt;strong>Highly recommended&lt;/strong>&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Grok&lt;/td>
&lt;td>Strong&lt;/td>
&lt;td>Mild&lt;/td>
&lt;td>Free, strong search capability&lt;/td>
&lt;td>Some operational content&lt;/td>
&lt;td>&lt;strong>Recommended&lt;/strong>&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>ChatGPT&lt;/td>
&lt;td>Declining&lt;/td>
&lt;td>Mild&lt;/td>
&lt;td>Many features, large user base&lt;/td>
&lt;td>Gradually stepping down&lt;/td>
&lt;td>Can keep&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>DeepSeek&lt;/td>
&lt;td>Stable&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Stable, no enshittification&lt;/td>
&lt;td>Fewest features&lt;/td>
&lt;td>&lt;strong>Recommended to keep&lt;/strong>&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Kimi&lt;/td>
&lt;td>Average&lt;/td>
&lt;td>None&lt;/td>
&lt;td>No enshittification&lt;/td>
&lt;td>Heavy templating&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Claude&lt;/td>
&lt;td>Strong&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Suitable for work use&lt;/td>
&lt;td>Average search performance&lt;/td>
&lt;td>Can keep&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Dola&lt;/td>
&lt;td>Average&lt;/td>
&lt;td>Severe&lt;/td>
&lt;td>Excellent voice recognition&lt;/td>
&lt;td>Average model capability&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Copilot&lt;/td>
&lt;td>Average&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Microsoft support&lt;/td>
&lt;td>Mediocre&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Manus&lt;/td>
&lt;td>Average&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Some Agent capability&lt;/td>
&lt;td>Poor long-term development&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Perplexity&lt;/td>
&lt;td>Average&lt;/td>
&lt;td>Mild&lt;/td>
&lt;td>No significant advantages&lt;/td>
&lt;td>Discriminatory remarks&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>MiniMax&lt;/td>
&lt;td>Excellent&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Excellent model capability&lt;/td>
&lt;td>Missing basic functions&lt;/td>
&lt;td>Not recommended&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Tongyi&lt;/td>
&lt;td>Strong&lt;/td>
&lt;td>Very severe&lt;/td>
&lt;td>Strong model capability&lt;/td>
&lt;td>Over-operationalized&lt;/td>
&lt;td>&lt;strong>Recommend uninstallation&lt;/strong>&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>Qwen&lt;/td>
&lt;td>Strong&lt;/td>
&lt;td>None&lt;/td>
&lt;td>Strong model capability&lt;/td>
&lt;td>Not yet on iOS&lt;/td>
&lt;td>Can keep&lt;/td>
&lt;/tr>
&lt;/tbody>
&lt;/table>
&lt;p>Conclusion:&lt;/p>
&lt;ol>
&lt;li>&lt;strong>Highly recommended tools&lt;/strong>:
&lt;ul>
&lt;li>&lt;strong>Gemini&lt;/strong>: Despite service instability, model performance is outstanding and user experience is good, suitable for work and life use.&lt;/li>
&lt;li>&lt;strong>Grok&lt;/strong>: Powerful search capabilities and free usage make it a worthwhile choice to keep.&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>Recommended tools to keep&lt;/strong>:
&lt;ul>
&lt;li>&lt;strong>ChatGPT&lt;/strong>: Although gradually stepping down from the throne, the free version still meets most needs, and has a large user base.&lt;/li>
&lt;li>&lt;strong>DeepSeek&lt;/strong>: Known for stability, with no over-operationalization issues, maintaining tool purity.&lt;/li>
&lt;li>&lt;strong>Claude&lt;/strong>: Excellent performance in code, suitable for work use.&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>Not recommended tools&lt;/strong>:
&lt;ul>
&lt;li>&lt;strong>Zhipu Qingyan, Yunbao, Kimi, Dola, Copilot, Manus, Perplexity, MiniMax, Tongyi&lt;/strong>: These tools are either severely enshittified, have average model capabilities, or are missing basic functions, and are not recommended for long-term use. However, the underlying models are still noteworthy, such as Kimi, MiniMax, and Qwen models. Unfortunately, the products aren&amp;rsquo;t user-friendly enough. Additionally, as of today, Qwen is still not available on the iOS App Store.&lt;/li>
&lt;/ul>
&lt;/li>
&lt;/ol>
&lt;blockquote>
&lt;p>The above conclusions are limited to user clients, published in November 2025. Does not involve API calls.
For work and programming scenarios, Claude Sonnet 4.5, ChatGPT-5.1-Codex, Grok Code Fast 1, MiniMax M2 and GLM 4.6, Gemini 3 Pro are worth recommending&lt;/p>
&lt;/blockquote>
&lt;h3 id="postscript">Postscript
&lt;/h3>&lt;p>Times have changed. ChatGPT has improved its experience in Europe and it now works very well, so the content of this post is no longer valid.&lt;/p></description></item><item><title>April 2025 Tech Notes 2</title><link>https://nansenli.com/post/2025/04/tech250402/</link><pubDate>Sun, 20 Apr 2025 12:58:30 +0100</pubDate><guid>https://nansenli.com/post/2025/04/tech250402/</guid><description>&lt;h1 id="github-copilot">Github Copilot
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/6f49f8b9f1ec.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>No more fear when facing unfamiliar Github projects. Just click the Copilot icon on the project homepage to start a conversation, and you can even use GPT-4.1 for communication, with up to a 1M context window. From questions like &amp;ldquo;What can this project be used for?&amp;rdquo; to &amp;ldquo;How to deploy and handle bugs?&amp;rdquo;, it can answer them all well.&lt;/p>
&lt;h1 id="dia-browser">Dia Browser
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/ba4da2cf2850.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>&lt;a class="link" href="https://www.diabrowser.com/" target="_blank" rel="noopener"
>https://www.diabrowser.com/&lt;/a>&lt;/p>
&lt;p>It&amp;rsquo;s not yet in public beta, but you can submit your email via the link above to join the waitlist.&lt;/p>
&lt;p>The Arc browser is already great, offering a stable, mature, and consistent experience—it&amp;rsquo;s much more than just a sidebar tab innovation.&lt;/p>
&lt;p>Dia is the next-generation product from the Arc browser team. Honestly, I don&amp;rsquo;t quite understand why they need to launch a new product instead of integrating these features directly into Arc. Now, the founder seems to have put Arc on hold.&lt;/p>
&lt;p>From a product experience perspective, Dia browser seems to consist of the following parts:&lt;/p>
&lt;ul>
&lt;li>A built-in AI assistant for daily conversations&lt;/li>
&lt;li>A browser operation assistant similar to nanobrowser or browser-use, helping plan and operate the browser&lt;/li>
&lt;li>An AI writing assistant that can generate text at the cursor&lt;/li>
&lt;/ul>
&lt;p>A product with great potential.&lt;/p>
&lt;h1 id="release-of-gpt-o3--o4-mini--o4-mini-high--41--41-mini--41-nano">Release of GPT o3 / o4-mini / o4-mini-high / 4.1 / 4.1 mini / 4.1 nano
&lt;/h1>&lt;table>
&lt;thead>
&lt;tr>
&lt;th>&lt;strong>Model&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Context Window&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Input Price&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Output Price&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Image Price&lt;/strong>&lt;/th>
&lt;th>&lt;strong>Main Features&lt;/strong>&lt;/th>
&lt;/tr>
&lt;/thead>
&lt;tbody>
&lt;tr>
&lt;td>o4-mini-high&lt;/td>
&lt;td>200K tokens&lt;/td>
&lt;td>$1.10/M&lt;/td>
&lt;td>$4.40/M&lt;/td>
&lt;td>$0.842/K image input&lt;/td>
&lt;td>High-reasoning version of o4-mini; high throughput, low latency; supports chain calls and structured output&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>o4-mini&lt;/td>
&lt;td>200K tokens&lt;/td>
&lt;td>$1.10/M&lt;/td>
&lt;td>$4.40/M&lt;/td>
&lt;td>$0.842/K image input&lt;/td>
&lt;td>Compact reasoning; AIME 99.5%, high SWE-bench score; excellent STEM, vision, and code editing capabilities&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>o3&lt;/td>
&lt;td>200K tokens&lt;/td>
&lt;td>$10/M&lt;/td>
&lt;td>$40/M&lt;/td>
&lt;td>$7.65/K image input&lt;/td>
&lt;td>General strong model; top math, science, coding, and vision reasoning; BYOK required&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>GPT-4.1&lt;/td>
&lt;td>1.05M tokens&lt;/td>
&lt;td>$2/M&lt;/td>
&lt;td>$8/M&lt;/td>
&lt;td>N/A&lt;/td>
&lt;td>Flagship model; million-token context; excellent instruction following, enterprise-grade retrieval, and multimodal understanding&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>GPT-4.1 Mini&lt;/td>
&lt;td>1.05M tokens&lt;/td>
&lt;td>$0.40/M&lt;/td>
&lt;td>$1.60/M&lt;/td>
&lt;td>N/A&lt;/td>
&lt;td>Medium version; close to GPT-4o performance; low latency, low cost; strong coding and vision understanding&lt;/td>
&lt;/tr>
&lt;tr>
&lt;td>GPT-4.1 Nano&lt;/td>
&lt;td>1.05M tokens&lt;/td>
&lt;td>$0.10/M&lt;/td>
&lt;td>$0.40/M&lt;/td>
&lt;td>N/A&lt;/td>
&lt;td>Fastest and cheapest in the series; million-token context; suitable for classification and autocomplete tasks&lt;/td>
&lt;/tr>
&lt;/tbody>
&lt;/table>
&lt;p>GPT-4.1 Nano is naturally suited for translation and is extremely cost-effective. For example, when used for immersive translation with bilingual subtitles enabled, translating an entire season of Black Mirror costs only about $0.005 per episode, and all 7 seasons just $0.25—very low cost.&lt;/p>
&lt;p>The context window for GPT-4.1 has finally reached 1M. Previously, GPT was far behind in code writing, but now it has caught up.&lt;/p>
&lt;p>o3, o4-mini, and o4-mini-high are all surprisingly impressive, giving a bit of a &amp;ldquo;mini deep research&amp;rdquo; feel.&lt;/p>
&lt;p>Currently, my daily usage has basically become:&lt;/p>
&lt;ul>
&lt;li>o4-mini: Simple search and summarization, tolerant of mistakes&lt;/li>
&lt;li>o4-mini-high: Complex search, summarization, multi-step reasoning&lt;/li>
&lt;li>o3: Complex search, summarization, multi-step reasoning where high accuracy is required&lt;/li>
&lt;li>4o: Intuitive and knowledge-based answers&lt;/li>
&lt;/ul>
&lt;p>I&amp;rsquo;ve also tracked the usage frequency of each model, and I use almost all of them—none are useless, unlike o3-mini, which is almost unusable due to high hallucinations.&lt;/p>
&lt;p>&lt;img src="https://nansenli.com/img/rescued/018788a20986.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;h1 id="yuanbao-integrates-with-wechat-contacts">Yuanbao Integrates with WeChat Contacts
&lt;/h1>&lt;p>Yuanbao WeChat can directly add contacts and send messages, supporting both voice and image messages. It&amp;rsquo;s a competitor to Doubao. The advantage is that no extra app installation is required. I previously recommended Doubao to my parents, and they really liked it. If Yuanbao could also support voice calls or even video chat, it would be even more valuable.&lt;/p>
&lt;h1 id="openai-free-token-giveaway">OpenAI Free Token Giveaway
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/7c4b97a270c2.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>OpenAI is now offering a free token giveaway. You can enable sharing logs in Data controls to claim them. Unfortunately, the duration is too short—just a week or two—so it&amp;rsquo;s hard to use up all the quota.&lt;/p>
&lt;h1 id="other-github-dailyweekly-trending-projects">Other Github Daily/Weekly Trending Projects
&lt;/h1>&lt;p>&lt;a class="link" href="https://github.com/kestra-io/kestra" target="_blank" rel="noopener"
>https://github.com/kestra-io/kestra&lt;/a>&lt;/p>
&lt;p>An automation platform, with competitors including: Airflow (data pipelines), Zapier (low-code marketing automation), Rundeck (ops automation), Camunda (e-commerce business automation)&lt;/p>
&lt;p>&lt;a class="link" href="https://github.com/Zackriya-Solutions/meeting-minutes" target="_blank" rel="noopener"
>https://github.com/Zackriya-Solutions/meeting-minutes&lt;/a>&lt;/p>
&lt;p>A speech-to-text software, a competitor to Apple Voice Memos.&lt;/p>
&lt;p>&lt;a class="link" href="https://github.com/maybe-finance/maybe" target="_blank" rel="noopener"
>https://github.com/maybe-finance/maybe&lt;/a>&lt;/p>
&lt;p>Personal financial report management, similar to Youzhi Youxing.&lt;/p>
&lt;p>&lt;a class="link" href="https://github.com/supabase-community/supabase-mcp" target="_blank" rel="noopener"
>https://github.com/supabase-community/supabase-mcp&lt;/a>&lt;/p>
&lt;p>A tool for connecting to databases when developing software with AI locally, based on the MCP protocol. If MCP becomes a real future standard, databases may natively support MCP, creating new competition.&lt;/p>
&lt;p>&lt;a class="link" href="https://github.com/langgenius/dify" target="_blank" rel="noopener"
>https://github.com/langgenius/dify&lt;/a>&lt;/p>
&lt;p>A low-code LLM platform, a competitor to LangChain, and can partially replace LangChain in current application scenarios.&lt;/p></description></item><item><title>April 2025 Tech Notes</title><link>https://nansenli.com/post/2025/04/tech2504/</link><pubDate>Tue, 15 Apr 2025 20:45:20 +0100</pubDate><guid>https://nansenli.com/post/2025/04/tech2504/</guid><description>&lt;h1 id="introduction">Introduction
&lt;/h1>&lt;p>I plan to record some weekly tech-related notes and insights starting from this issue. I&amp;rsquo;ll probably update every two to four weeks.&lt;/p>
&lt;h1 id="migrating-personal-blog-from-hexo-to-hugo">Migrating Personal Blog from Hexo to Hugo
&lt;/h1>&lt;p>This week, I migrated my blog from Hexo to Hugo, primarily because:&lt;/p>
&lt;p>Hugo has better i18n support. Hexo doesn&amp;rsquo;t support i18n very well; I previously modified the template to implement Chinese and English switching, but the maintenance cost was high. I discovered that Hugo natively supports multiple languages.&lt;/p>
&lt;p>After switching, I also found the following advantages:&lt;/p>
&lt;p>Theme configuration is simpler. I&amp;rsquo;m using the &lt;a class="link" href="https://github.com/CaiJimmy/hugo-theme-stack" target="_blank" rel="noopener"
>Stack&lt;/a> theme, which has been mostly problem-free. The official example configuration works well right out of the box. It also comes with Google Analytics integration and RSS feeds that automatically separate by language, making it very convenient.&lt;/p>
&lt;p>During the migration process, I didn&amp;rsquo;t encounter any major obstacles. You just need your old Markdown files, copy them to the new project directory, and force push to your original GitHub Pages repository. After force pushing, you&amp;rsquo;ll need to reconfigure your custom domain in GitHub, and the upgrade will be seamless, including comments and other features.&lt;/p>
&lt;p>The only difference is that Hugo uses lowercase links by default, while Hexo preserves case sensitivity, which might cause some old links to become invalid, requiring search engines to rebuild their indices.&lt;/p>
&lt;h1 id="mysterious-optimus-alpha-model-appears-on-httpsopenrouteraihttpsopenrouterai-and-dominates-rankings">Mysterious Optimus Alpha Model Appears on &lt;a class="link" href="https://openrouter.ai/" target="_blank" rel="noopener"
>https://openrouter.ai/&lt;/a> and Dominates Rankings
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/a7835acf9e85.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>Because this model was free and powerful, it was chosen as OpenRouter&amp;rsquo;s primary traffic handling model after launch.&lt;/p>
&lt;p>After OpenAI officially announced GPT-4.1 on the 14th, the model was removed from OpenRouter. Obviously, this model was GPT-4.1.&lt;/p>
&lt;p>GPT-4.1 is a model specialized in programming capabilities, with a 1M context window. Indeed, in the programming field, OpenAI has been struggling against Claude 3.7 and Gemini 2.5 Pro. Neither o1 nor o3 models have achieved a leading position in programming. Hopefully, OpenAI will keep improving.&lt;/p>
&lt;blockquote>
&lt;p>OpenRouter is a middleware service platform that allows users to &lt;strong>access multiple large language model APIs&lt;/strong> through a unified interface. You can think of it as a &amp;ldquo;proxy&amp;rdquo; or &amp;ldquo;router&amp;rdquo; that manages API calls to different large model providers (OpenAI, Anthropic, Mistral, Google Gemini, Cohere, etc.).&lt;/p>
&lt;/blockquote>
&lt;p>It seems that this time, OpenAI chose OpenRouter as its test release platform, rather than &lt;a class="link" href="https://lmarena.ai/" target="_blank" rel="noopener"
>https://lmarena.ai/&lt;/a>.&lt;/p>
&lt;h1 id="openrouterai">OpenRouter.ai
&lt;/h1>&lt;p>&lt;a class="link" href="https://openrouter.ai/" target="_blank" rel="noopener"
>OpenRouter&lt;/a>&lt;/p>
&lt;p>This is an LLM platform that includes many free/paid high-quality models.&lt;/p>
&lt;p>&lt;img src="https://nansenli.com/img/rescued/def8fd9ff726.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>Select &amp;ldquo;Prompt pricing&amp;rdquo; as 0, then choose &amp;ldquo;Top Weekly&amp;rdquo; to see the best free models currently available.&lt;/p>
&lt;p>However, when using it, you need to enable logging and data usage for training.&lt;/p>
&lt;p>Currently, Gemini 2.5 Pro is an option.&lt;/p>
&lt;p>The Ranking feature allows you to view real-time traffic distribution across all models on the platform, making it easy to see which models perform best.&lt;/p>
&lt;h1 id="nanobrowser">nanobrowser
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/8d34bb10e9ba.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>A browser plugin for Chrome that can control browser automation operations.&lt;/p>
&lt;p>I tested it and found it doesn&amp;rsquo;t support Arc browser. The experience was limited - page scrolling occasionally fails, it can&amp;rsquo;t summarize entire page content (only partial content is visible), and operations are sometimes incorrect.&lt;/p>
&lt;p>Although I don&amp;rsquo;t particularly recommend this product, it might represent a future direction. It seems we&amp;rsquo;ll see many AI agents emerging.&lt;/p>
&lt;p>Gray market industries like ticket scalping and point farming will probably be delighted.&lt;/p>
&lt;h1 id="markitdown">markitdown
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/e37c2c0e7d7a.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>A tool for converting content in different formats to Markdown.&lt;/p>
&lt;p>Offline, it can process different files into Markdown format. I tried it, and the conversion is fairly basic.&lt;/p>
&lt;p>It also supports integration with LLMs to improve conversion results and handle images in documents.&lt;/p>
&lt;p>This could potentially be a good MCP service in the future.&lt;/p>
&lt;h1 id="pdf-search">PDF search
&lt;/h1>&lt;p>&lt;img src="https://nansenli.com/img/rescued/d1310e598823.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>An application that can quickly search content across multiple PDFs.&lt;/p>
&lt;p>It uses a fuzzy search approach. The principle is to split PDFs and perform embedding vector calculations, then match them with search terms.&lt;/p>
&lt;p>Suitable for searching documents during exam periods.&lt;/p>
&lt;p>The downside is that indexing is done by page, so after searching, you need to find the corresponding part on the page yourself.&lt;/p></description></item><item><title>Summary of LLM Tools Usage Experience</title><link>https://nansenli.com/post/2025/04/aitest/</link><pubDate>Tue, 08 Apr 2025 19:47:48 +0000</pubDate><guid>https://nansenli.com/post/2025/04/aitest/</guid><description>&lt;h1 id="chatgpt-o4-mini">ChatGPT o4-mini
&lt;/h1>&lt;p>&lt;a class="link" href="day://2025.04.17" >&lt;code>Thu, Apr 17&lt;/code>&lt;/a> The latest model released in the past two days, with significant reasoning capabilities and excellent multi-step execution. It appears that o4-mini has completely surpassed o3-mini in practical task execution performance.&lt;/p>
&lt;p>The following image shows an example of my usage. The question was: &amp;ldquo;In Black Mirror Season 7 Episode 3, there are many Asian characters. Are these Asians Chinese, Korean, Japanese, Malaysian, or Singaporean?&amp;rdquo; The model successfully called the search engine multiple times and obtained the correct answer:&lt;/p>
&lt;p>&lt;img src="https://nansenli.com/img/rescued/51ec0e03d882.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>Additionally, Projects can now be normally called and used in this mode. Since the model can search autonomously rather than using RAG methods, the efficiency of file searching has significantly improved.&lt;/p>
&lt;h1 id="chatgpt-excluding-o3-o4-mini">ChatGPT (excluding o3, o4-mini)
&lt;/h1>&lt;ul>
&lt;li>
&lt;p>GPT-4o: Versatile. Excellent for daily communication and learning, with sufficient context window to easily handle code and document retrieval tasks. Supports multiple attachment formats and online code execution. Drawback: generates limited context length, not suitable for very long content. For lengthy conversations, context compression is severe and may forget earlier content.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Projects: Somewhat redundant. Suitable for scenarios requiring frequent searching and text retrieval across multiple files.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>GPT-4o mini: Weaker than GPT-4o, but has a massive context window, recommended for translating very long texts.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>o1: A decent chain-of-thought model, suitable for solving complex code and mathematical problems. Not good at handling emotional or intuitive problems. However, the thought chain is too short and quality is not high. It&amp;rsquo;s recommended to use the output as input for Gemini-2.5pro or DeepSeek-R1 to optimize quality.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>o3-mini: Severe hallucinations, inferior to o1, but extremely fast reasoning speed.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Search: Slightly redundant. New version&amp;rsquo;s search results are heavily restricted, even inferior to model output without search. Suitable as a light search alternative.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Deep Research: Excellent tool. Best quality among similar features, most comprehensive output results.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>GPT-4.5-preview: Excellent. Massive context window, strong long-text comprehension. Has vast memory and strongest intuition. Not actually good at reasoning, but performs best due to low hallucinations.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Canvas: Overall inferior to Cursor. Context output too small, text length limited. Suitable for short-text scenarios with real-time editing.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Work with Apps on macOS: Functionality experience crushed by Cursor, essentially a simplified version of Cursor.&lt;/p>
&lt;/li>
&lt;/ul>
&lt;h1 id="claude">Claude
&lt;/h1>&lt;ul>
&lt;li>Claude 3.7 Sonnet: Excellent, suitable for generating various code. Web version supports massive context, almost matching API&amp;rsquo;s context history message retention. Model very suitable for writing communication texts like emails, with accurate, concise, and unpretentious wording. Drawback: limited usage in free version.&lt;/li>
&lt;/ul>
&lt;h1 id="deepseek">DeepSeek
&lt;/h1>&lt;ul>
&lt;li>
&lt;p>R1: Severe hallucinations, unstable performance, occasionally affecting usability, requires very high quality prompts. When prompts are correct or context is complete, it outperforms o1. Can be used as input for o1&amp;rsquo;s output to optimize quality.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>V3: Very excellent, alternative to GPT-4o.&lt;/p>
&lt;/li>
&lt;/ul>
&lt;h1 id="grok">Grok
&lt;/h1>&lt;ul>
&lt;li>
&lt;p>Grok-3: Generates text fluently and naturally, less AI-like, suitable for natural writing and novel creation. Overall mediocre, fewer productivity tools than GPT-4o.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Grok-3 + search: Excellent. Leverages English social media data, quickly analyzes news and current events. Outputs long content, barely filters search source content, stronger search capabilities than GPT-4o.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Grok-Deep Search: Slightly redundant, actually inferior to Grok-3 + search. Generated content heavily templated, affecting quality.&lt;/p>
&lt;/li>
&lt;/ul>
&lt;h1 id="gemini">Gemini
&lt;/h1>&lt;ul>
&lt;li>
&lt;p>2.5 Pro: Excellent model, alternative to GPT-4o, minimal hallucinations, high-quality search results, complete and clear thought chains, strong logical reasoning. Supports integration with Google tools, very useful in specific scenarios, such as uploading screenshots or text to automatically create events in Google Calendar. Drawback: often claims to have performed searches when it actually hasn&amp;rsquo;t.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>Deep Research: Average, slightly better than old GPT Search. Occasional comprehension deviations, search and document generation executed in stages, process fragmented, heavily templated. Currently the only product that can replace GPT Deep Research.&lt;/p>
&lt;/li>
&lt;/ul>
&lt;h1 id="v0dev">v0.dev
&lt;/h1>&lt;ul>
&lt;li>Suitable for writing and previewing frontend components online.&lt;/li>
&lt;/ul>
&lt;h1 id="chatqwenai">chat.qwen.ai
&lt;/h1>&lt;ul>
&lt;li>Qwen2.5-Max: Very excellent model, alternative to GPT-4o, fast generation speed, supports thinking mode.&lt;/li>
&lt;/ul>
&lt;h1 id="zhihu-direct-answer">Zhihu Direct Answer
&lt;/h1>&lt;ul>
&lt;li>Supports searching Zhihu&amp;rsquo;s entire network content. Possible competitor is Xiaohongshu Direct Answer.&lt;/li>
&lt;/ul>
&lt;h1 id="perplexity">Perplexity
&lt;/h1>&lt;ul>
&lt;li>Supports basic search, mediocre quality, slightly redundant.&lt;/li>
&lt;/ul>
&lt;h1 id="mistralai">Mistral.ai
&lt;/h1>&lt;ul>
&lt;li>Fast speed, large context window, alternative to GPT4o-mini.&lt;/li>
&lt;/ul>
&lt;h1 id="cursor">Cursor
&lt;/h1>&lt;ul>
&lt;li>Excellent tool. Can index entire code repositories, suitable for large project development. Supports multiple model switching, local command line execution and terminal takeover, suitable for the following tasks:
&lt;ul>
&lt;li>Writing LaTeX documents, replacing Overleaf&lt;/li>
&lt;li>Assisting with various assignments&lt;/li>
&lt;li>Code repository search and specific feature location&lt;/li>
&lt;li>Creating unit tests&lt;/li>
&lt;li>Executing git operations&lt;/li>
&lt;li>Executing deployment tasks&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>Can execute almost everything involving documents and command lines, helpful in development, debugging, debugging, and document writing&lt;/li>
&lt;li>Truly the most Agent-like product at the current stage (though this claim no longer holds after the release of o3 and o4-mini, as they can also perform multi-step planning and task execution. However, o3 and o4-mini can only operate within their own limited environment, while Cursor can execute on the user&amp;rsquo;s computer, which is an advantage.)&lt;/li>
&lt;/ul>
&lt;h1 id="example-leetcode-tracker">Example: Leetcode Tracker
&lt;/h1>&lt;p>There are many shared Leetcode premium accounts on Taobao, but sometimes we still need to add our problem-solving records to our own accounts. After purchasing a shared premium account, we can export a company&amp;rsquo;s problems as a CSV file, then export our own problem-solving records as another CSV file. By comparing these two files, we can easily track our current progress. Cursor is very good at writing such small tools:&lt;/p>
&lt;p>&lt;img src="https://nansenli.com/img/rescued/1a7bea6f828c.png"
loading="lazy"
alt="Image.png"
>&lt;/p>
&lt;p>We just need to copy elements from the webpage, and Cursor can write complete CSV files for us and automatically generate a perfect frontend interface, greatly improving our problem-solving efficiency.&lt;/p></description></item></channel></rss>