Tag: Xiaohongshu

A nice deal, Grok $30 for three months

I used to most recommend chatgpt plus, but the other day chatgpt banned my codex. Logging into codex again requires SMS two-factor authentication. Since it was a number from a code-receiving platform, openai does not support self-service phone number rebinding, so the main account is basically dead. I switched to try out supergrok, and the composer it provides…

Tags:
Categories:

If Zhipu's stock price falls, where will the money flow?

Zhipu’s model performance is decent, but its engineering capability is limited, and subscription users can feel the service quality. The advantage of glm5.2 is rigorous thinking, at the cost of low token efficiency, leading to a higher overall price. Assuming most users do appreciate well-performing models, but their feet vote for cost-effective models, then cost-effective vendors will win the users. Assuming the actual profit model of AI model vendors resembles that of factories, valuing scale effects and cost control, then cost-effective vendors are more likely to gain profits…

Tags:
Categories:

opencode go subscription launches glm5.2, genuine glm5.2!

The second screenshot is enough to prove it’s genuine, the familiar 1302 flavor. Stable to use on 6.18, speed is quite good, maybe the supplier no longer uses z.ai. In practice, it’s stronger than deepseekv4pro in coding, script error rate dropped a lot. Of course, the price is also many times higher. Among domestic models, only deepseek is recommended, hoping it soon…

Tags:
Categories:

Learn Product Design from Big Tech

Page 1: As the repayment due date approaches, the most important action is ‘make repayment’ rather than ‘repay with ease’, yet the visual emphasis is on ‘repay with ease’. ‘Repay with ease’ actually refers to installment repayment. Page 2: What matters is ‘how much is the interest’ rather than ‘how much interest is reduced’; the page never shows how much the interest is, and the visual focus is on ’national interest subsidy’. Page 3: The visual focus is on the annualized simple interest rate of 15.16%…

Tags:
Categories:

Huawei token plan is quite expensive

Not cheap, especially the way token consumption is counted—both input and output are counted within the plan. I often use 20 million tokens in a single request, and the 59 yuan tier only covers about 3 of my requests. For coding, roughly 99% of input is input cache, and billing cache hits at full price is pretty absurd. If they want people to actually use it, they shouldn’t count input—only count output toward the quota,…

Tags:
Categories:

github copilot's route choice

From February to July last year, copilot was freely usable in vscode insider. During this period, I witnessed the gradual decline of copilot’s agent capabilities. Hard to imagine, the first month preview of copilot turned out to be the best in effect, and then the agent capability gradually declined. What exactly changed? That is, copi…

Tags:
Categories:

Those worried about being laid off should use AI more

Over the past year and a half, some companies overestimated the role of AI and began layoffs. As time went on, companies gradually realized that AI is not as smart or cheap as expected, and began to reduce AI token consumption and slow down layoffs. A very important reason is that AI is not cheap and cannot take the blame. In the past, some industries joked that AI cannot go to jail, but based on my actual work…

Tags:
Categories:

I understand that we can no longer casually criticize companies

Although some netizens’ posts do have the problem of the blind men feeling the elephant, it shouldn’t go as far as forbidding the squirrel from saying the water is deep. Comparison posts often have losers and winners, and I wonder if losers will be banned in the future. Some manufacturers really overhype, or use obscure qualifiers in their rankings, which easily misleads ordinary consumers. Previously, product criticisms were shared based on personal experience to help others avoid pitfalls, but…

Tags:
Categories:

Xiaomi API ranks among the top in LLM cost-performance

Previously posted criticism that mimo-2.5-pro’s performance couldn’t justify its pricing. Its performance really can’t be compared to gpt5.x and opus4.x. It’s not that I wanted to compare them, but Xiaomi positioned them as competitors, so I pointed out the obvious gap. But now Xiaomi has shifted to benchmarking against deepseek, and I think it might still put up a fight. Selling LLMs…

Tags:
Categories:

One factor in architecture rot is compromising on development cost

Characteristics of rotten architecture include: lots of special-case handling, compatibility handling, severe coupling, and being too burdened to reform. All architectures gradually rot. A rotten architecture makes it hard to develop new features, fix bugs, and a small change affects everything, making testing and QA very difficult as well. I once saw a cross-industry development enthusiast post that restricting the abstraction behavior of models when writing code makes the code easier to understand. Code that is easier for humans to understand is also easier for LLMs to understand, enabling rapid development and fewer errors. I think cross-industry developers probably work on smaller projects and are used to pipeline-style logic. If there is an expectation of scaling up, the mindset cannot be limited to pipelines but should adopt a construction-worker mindset. Expecting to build a 30-story building means digging a 10-meter foundation; expecting water and electricity means reserving water and electrical shafts. This is goal-clear design. There is also goal-unclear design, such as reserving a socket somewhere in the house for a future appliance without knowing what it will be.

Tags:
Categories:

Moderate forgetting benefits model performance

The GPT-5 series has a high recall rate and can achieve very rigorous logic, making it excellent for debugging. At the same time, it can promptly detect contradictions in the context, offering a unique experience among current models. But recently, in my personal reverse engineering project, I found that GPT-5 has limited reverse engineering capabilities, clearly inferior to the opus model. It often explicitly replies that something won’t work, while switching to opus can quickly achieve functional reverse engineering…

Tags:
Categories:

Does AI-written code have fewer or more bugs

AI writes code so fast that bugs per line are surely lower, but lines of code explode, and requirements we used to skip get cranked out, so bugs keep growing. Also, AI is unbeatable at finding others’ bugs, surfacing hidden ones. Major vendors and popular libraries never blew up with critical bugs every few days like now. Reverse engineering and attacks get easier, and people will gradually realize AI does more than create—future offense and defense may get interesting…

Tags:
Categories:

Recommend using claude code with domestic models

I actually have some aversion to the company anthropic, but it is currently at the height of its popularity, so I have to hold my nose and use it. I subscribed to a claudecode pro plan, but the quota is very small and can only be used for planning. Domestic models perform slightly worse, but the quota is more, offering clear cost-performance advantages. Domestic model providers currently do not offer a response API, so…

Tags:
Categories:

The trick to using the GPT5 series models is to reduce restrictions

Someone shared very long and fancy prompts claiming to improve GPT5’s performance, which should be viewed with caution. Unlike the sonnet or opus4.6 and earlier models in claudecode, GPT5 is more noticeably negatively affected by randomly written custom prompts. GPT5’s system prompts should be carefully considered, otherwise it is easy to degrade the model’s performance.

Tags:
Categories:

Tencent's model is much better than I expected, recommended to try

My main use case is writing code, and my projects are mainly command-line tools. Tencent HY3 Preview performs very well in my use case, with good instruction following and dialectical logic. I also tried Baidu’s CoBuddy coding model recently launched on openrouter, and it’s terrible—just two simple uses reveal its poor capability, with only 131k context…

Tags:
Categories:

Recommended coding-friendly models to try

First, the old friend openrouter. After depositing $10, you get 1000 free calls per day. Most of the time there are a few decent models in the free tier. The official policy reserves the right to set an expiry on the deposited amount, there may be a one-year validity period, but my $10 from over a year ago is still valid. Recently tried on a limited-time basis inclusion…

Tags:
Categories:

codex free auth automatic acquisition flow feasibility

I have two hundred free accounts and one plus account, connected to cliproxyapi, reverse-proxied for use in various CLIs. These accounts are enough for heavy personal use. If you enable 1M context, fast mode, use subagent, high-res image parsing, the tokens of these few hundred accounts will still be stretched thin, and you still have to plan usage. F…

Tags:
Categories:

Sorry but mimo v2.5 pro really doesn't cut it

Some people share that Xiaomi’s model can perform long-horizon tasks, executing hundreds of tool calls without errors. I don’t know who initiated this testing direction, but it’s like saying a student can hold their bladder for 100 minutes in an exam room—meaningless. Normal people understand that for a standardized exam, the only thing that matters is the score. To be more lenient, the score accounts for over 95% of the evaluation system…

Tags:
Categories:

Escape from Cursor

It may be a bit late to talk about escaping from Cursor now. Actually, over the past six months I have gradually reduced my use of Cursor, because my coding habits have changed, and Cursor’s strengths are no longer the reason I choose it. The feature with which Cursor still firmly ranks first is its unique completion model, supporting cross-line and cross-file completion in all directions. Whether in response speed or acceptance rate, no rival can match it…

Tags:
Categories:

Anthropic's well-known advantage is common sense.

Anthropic’s well-known advantage is common sense. Some development work relies on common sense, some on proprietary logic. Opus’s reasoning style is to guess many possibilities and then quickly verify. GPT5 tends to reason based on code implementation, repeatedly sorting out the business. The style is similar to the difference between breadth-first and depth-first. As…

Tags:
Categories:

Zhipu's terrible service is unmatched

This is the third time I’ve been fooled by Zhipu. They said they were retaining old users by giving unlimited weekly quota, so I came back to give it another try, and got tricked again. It has no weekly limit, but it’s completely unusable when I need it. It only comes back to life right when I start and finish work, which is like meeting at the peak. In February I said Zhipu lite had single concurrency, and Zhipu used its business certificate to report me to xhs…

Tags:
Categories:

Winner-take-all market, losers merely scrape by

The cost narrative does not fit the large model industry; everyone is losing money, even the winners. The losers feel this cannot go on, but lack the ability to persuade the winners to raise prices, so they decide to raise prices on their own outdated products first—this is reckless self-sabotage and digging their own graves. Domestic open-source model vendors make a name for themselves, and after their models gain some competitiveness they all go closed-source; qwen3.6/minimax2.7/glm5.1turbo, etc. have not been open-sourced. Other small vendors just tinker with the base models, distill the winners’ models, and then expect to charge money—they must be dreaming.

Tags:
Categories:

This feeling is addictive

Half a year after the release of the GPT5 series, its logical rigor leads the field, but the most criticized aspect is its slow speed, especially GPT5.2, which is absurdly slow. For many ordinary requests, opus may start making changes after 20 seconds of context gathering, while GPT5.2 might analyze for 20 minutes before starting to make changes. The concurrent GPT5.2 codex model…

Tags:
Categories:

Ten Times the Price, Six Times the Usage

On Claude’s side, max 5x costs five times the pro price with five times the usage, and max 20x costs ten times the price with 20 times the usage. On OpenAI’s side, pro is six times plus, and review is ten times, making the cost-performance hard to judge. Of course it also has the privilege of prioritized request processing, and I’m not sure how much faster it can be. I use chat for daily conversations, with extend thinking on, and it feels like an average of about five minutes per question. Plus has a weekly limit of 3000 conversations, which is far from being used up. The codex cli weekly quota is heavily enough for two days (without fast mode on), so you only need to prepare two or three plus accounts to meet weekly needs. If you’re impatient and turn on fast mode, having four accounts ready is also enough.

Tags:
Categories:

Daily Coding Plan Auto-Scraping Summary

Only curated domestic vendors with some reputation, not resellers, with a certain level of security assurance. In actual use, first-token latency, token rate, concurrency, and whether quantized are important but rarely mentioned by vendors. Comparatively, speed matters more, really. #jqknono.github.io/coding-plans-for-…

Tags:
Categories:

Thanks to OpenAI the Great Benefactor

At peak I subscribed to several plans at the same time, but couldn’t use them all up, wasting at least half of the few hundred bucks every month. OpenAI recently opened codex cli to free users with a certain free weekly quota. Those familiar with codex know its quota is very different from claudecode’s, and is extremely durable. I myself have 6 accounts connected behind a cli proxy, and using only the free tier already basically meets my needs. Some share that free accounts may have model quantization and reduced intelligence; I also have an account in an enterprise team, and later I’ll carefully test whether free GPT-5.3-codex differs from the paid one.

Tags:
Categories:

GPT-5.3-Codex-Spark Worth a Try

Those familiar with LLM benchmarks know that Opus often tops GPT in various leaderboards, but in actual use, GPT’s problem-solving ability is unmatched. As of posting, the latest major models in the GPT-5 series include gpt-5.2-mini, gpt-5.2, gpt-5.3-codex, gpt-5.3-codex-sp…

Tags:
Categories:

Zhipu Refund Now Available, Refund Window Closes on 3.6

The refund window will open within one week after the holiday, and close on March 6, 2026. At that time, go to the My Plans page and click Cancel and Refund to complete it with one click. Note that the refund window period ends on 3.6. Programmer friends, recall your experience using Zhipu in the past two months, trust your own feelings, and make a rational decision.

Tags:
Categories:

First Time Encountering Zhipu Model's Review Gate

First, the task was merely to translate Chinese to English for a nodejs app that converts a mermaid into an image. This is a pure tool, and the task was only translation. So far, I have encountered the following issues while using the Zhipu model: 1. Extremely low concurrency limits, the more you need to use it, the lower the concurrency 2. Low token generation speed 3….

Tags:
Categories:

Zhipu Plan Concurrency Too Low

Many people say the lite plan has low concurrency, but they may not have made clear exactly how low it is. Let me explain it more clearly. Only one concurrency. This means in tools with a subagent mechanism, lite simply cannot run. I have been buying various AI products for four or five years, and Zhipu’s plan is the biggest rip-off I have encountered; it is almost unusable for coding. Now it can only be used for single-thread…

Tags:
Categories:

Notes on using OpenRouter's free models

After recharging $10 on OpenRouter, you get 1000 free model calls per day, which is more than enough for me to use for validation. But for a long time, many of its models kept returning 429 errors, such as openai/gpt-oss-120b:free. Today I manually tested various methods and finally confirmed…

Tags:
Categories:

Open-source HarmonyOS App I Made -- Dev Cheatsheet

Open-source repo github:jqknono/reference-Harmony Main features see image, displays shortcuts and common commands for various popular dev tools, which help improve work efficiency. “Dev Cheatsheet” has clear architecture, simple and focused features, will never show ads in the future, and fully complies with Huawei’s app review policy. Submitted in Dec…

Tags:
Categories:

GPT-5.3-Codex becomes my new main model

I was stunned for a moment when I saw it, then realized “plan” was translated as “package” GPT5.2high is ridiculously slow, GPT5.3 hasn’t been released yet, but 5.3-codex came out first. In actual testing, the speed improvement is significant, and the logic capability drop is not as obvious as the drop from GPT-5.2-Codex compared to GPT5.2. It is a specialized coding model worth switching to…

Tags:
Categories:

Review of Qwen App

The Qwen app uses four-direction swipe gestures. On the left are chat history and agents, and on the right is the agent marketplace. When the swipe trajectory on the screen is not straight up and down, but rather like a left-falling stroke (丿) or right-falling stroke, the side content will be triggered. To return to the chat interface, you don’t swipe the screen, but perform a back gesture, sliding in from the screen edge to return. If the directions get mixed up, the app will freeze. The Qwen app…

Tags:
Categories:

Review of Doubao's App

Inconvenient to start a new conversation; after starting one, only a popup notification appears while the old messages remain in place. Even after switching topics, the previous answer is still visible, giving a feeling that the context hasn’t switched—which is uncommon among AI chat apps. However, Doubao’s lower barrier to entry does help it grow its user base.

Tags:
Categories:

Is it wise for model vendors to stockpile GPUs?

When some model developers first release a new version, the response speed is ideal and there is no intelligence degradation. But if too many people start using it, both speed and quality decline. This is the behavior of a leader—once performance falls behind, users churn and the stockpiled cards sit idle. To avoid this, one could consider decoupling heavy assets from intellectual property assets, selling model usage rights, or renting hardware resources.

Tags:
Categories:
All Tags