Signal Wall
Articles, memos, builds, and signals.
Kimi K3: The counterattack of AI in China from closed to open source
Behind Kimi K3 open source is the reverse penetration of China's AI ecosystem into the Silicon Valley open source paradigm
AI writes code, and I stare at the moments when it rolls over
AI writes code so fast, so fast that I really let it write at first-throw the need away and give me a large piece of it, which looks like something. Then I learned after being cheated a few times: Just because it writes quickly doesn't mean you can't help looking at it. During this period of time, I basically came here "staring at AI and writing code". I definitely dare not let it do what tasks I can safely hand over, which types of operations I have to stare at, and which types of operations I have never dared to let it do by myself. They were all overturned by it. This article explains clearly the seven most common rollover moments-start without looking, make small needs into a complete framework, say "completed" before running, forget the previous paragraph after chatting for a long time, do not have a long look at dangerous operations, write up non-existent APIs in a serious manner, and you will be intimidated when you question it-each explains why it is like this, and then gives you a corresponding "Seven Rules to Make it roll less". In the final analysis, AI lowers the threshold of "writing" but raises the threshold of "judgment"; whether it is a helper or a trap depends on whether you will be the one staring at it.
How open source weights can allow AI tools to reverse shape developers
Open weights turn models into customizable tools, forcing developers to re-learn parameters and data governance, and revolutionizing how they use them
Stop writing native apps: the web is the best tool
The tool chain (IDE, SDK, review process) of native apps shapes the developer's "only App" mindset, and web implementation with the same function is often faster and easier to maintain
I received 12 free big models and stepped on them for you
I have a principle for connecting AI to a bunch of projects at hand: I don't spend money if I can. In the past two years, the number of free models has been ridiculously high, and the free quotas of several companies are so large that you can't even spend them on making a small product-buy tokens with real money, and often you haven't done your homework. In the past, I have accepted almost all the large model APIs on the market that can be used for free prostitution, including more than a dozen domestic and foreign companies: some of them smell so fragrant after accepting them, and some want to curse after accepting them. This article will give you this list and the pits I have stepped on once: who to choose for permanent free, how to combine the running volume, where are the overseas companies fast, and the five most annoying pits-especially the "automatic transfer payment when the free quota is used up", which will really deduct money and lock your entire account. If you follow suit, you will save me the detours I took. (Policies are changing quickly, and the numbers in the text may not be accurate in two months. Remember to check it on the official website before accepting it.)
Without opening Xcode, iOS developers have become DevOps
Build without Xcode reduces iOS developers to CI/CD engineers, losing their deep understanding of the platform
There are pictures and truth, and they are officially invalid.
Let's take a look at a picture and guess whether it's true or false: Jianlibao, AD Calcium Milk, and Big Bubble Gum are displayed on the wooden shelves. Even the "red-eye" flaws in old photos are all there-a batch of photos of the "90s Canteen Shop" were swiped on the screen a while ago. Many people were "young again", and then they realized that they were all generated by AI, and there were not a single real shot. What's even more striking is the next step: Are you still divided into halal photos or AI edits the physical picture and bowl of malatang in the merchant's comment area? Something has happened-some college students used AI to convert static photos into dynamic videos that can pass live testing and swiped 50,000 yuan in three months. Someone passed their ID card to the AI and "changed it to Cook's face" and even changed the number. "There is a picture and the truth" is being voided. The first reaction of many people is to practice a stronger "AI counterfeiter" to catch fake pictures-but this road has already lost: in 2023, you can still see flaws, in 2025, the light will be filled in, and in 2026, the human eye and most detectors will not be able to distinguish (the detector accuracy rate is only 70 - 90% and the false alarm rate is 5 - 15%). So the industry changed its head: instead of arresting fakes, it changed to issuing "digital birth certificates"(content certificate C2PA + watermark SynthID) to authentic content-the anti-counterfeiting logic has been flipped from "arresting fakes" to "verifying authenticity." This article explains this turn, the three unblocked pits, and what ordinary people do now.
Is automation that doesn't understand code really more efficient?
Companies deliberately allow automation tools to run on opaque code in exchange for short-term efficiency, but bury long-term technical debt and cognitive disability.