Monday, June 29, 2026

Brothers, I will pick the two strongest signals today.

First, the countdown to AI's own creation has really come. Anthropic's Clark directly threw out the timetable: Recursive Self-Improvement (RSI) will come true before the end of 2028. Moreover, DeepSeek has just published a DSpark paper, which is engaged in speculative decoding to speed up reasoning. To put it bluntly, this thing allows you to learn to "guess" when generating large models. If the guesses are correct, skipping calculations, saving time and computing power. GPT-5.6 was also launched, claiming to be the strongest in history, but it was cheated by itself. It is estimated that after the ability approached the threshold, the output was uncontrollable or the alignment problem exploded. You see, what do these three signals together? It is the AI community that has begun to curl itself up. RSI allows AI to invent better model structures on its own, DSpark allows these models to run faster, and GPT-5.6 reveals that the stronger the ability, the more difficult it is to control. For ordinary developers, what you should be concerned about is that the cost of inference is falling rapidly, but the controllability of model behavior has become a new bottleneck. Don't just stare at benchmark to score points. After you deploy and go online, the model suddenly "wakes up" to complete your work. Can you withstand it? The biggest problem in this field now is not the lack of performance, but how to prevent unpredictable behavior in engineering.

The second story is about the struggle for narrative rights in the AI circle and the tearing of commercial reality. Anthropic accused Ali Qianwen of distilling his Claude. It was an old routine, but he was not using words. But look at a few other items today: mobile phone evaluation bloggers and manufacturers collaborated to fake the fraud and was exposed by CCTV, and Mercedes-Benz cut employee year-end bonuses and asked for extended working hours. What is the common logic behind this? It is a crisis of trust and cost transfer between Party A (financial owner) and Party B (technology/service provider). Distillation technology itself is fine and everyone uses it in engineering, but it is stigmatized as "stealing" by geopolitics and commercial competition. Just like mobile phone reviews, if you pay money, you will praise it, and if you don't pay, you will slander it. What users see is not the real experience at all. Looking at Mercedes-Benz, traditional manufacturing giants directly sacrificed the interests of employees under the impact of new energy. This tells us: No matter how strong AI is, the final commercial closed loop is still a matter of benefit distribution. For teams that make AI products, don't just indulge in technical showmanship, you have to think clearly whether your customers are willing to pay for your so-called "self-research" premium. If you just change the API and put it in a shell, what is the difference between it and fake mobile phone reviews? Sooner or later you will be skinned.

One final tip: Today, a large number of third brothers are pouring into Shenzhen to grab IT jobs, and there is also the open source project Agent-Reach using CLI to view the entire network with zero API fees. These signals all point to the same trend-technological democratization is exposing middle and low-end jobs to global competition, but it is also lowering the threshold for innovation. If you are still struggling with whether skills are important, why not ask yourself: Are your skills the price tag today or the moat tomorrow?

Mercedes-Benz has suspended year-end bonuses to 90,000 employees and extended unpaid working hours. This is not only a pain in the traditional automobile industry, but also a fierce friction when power is handed over from the old mechanical era to the computing era. While Mercedes-Benz saves tens of millions of euros by squeezing manpower, Huawei's Tuling platform is turning the chassis into a computable body, and G7 Easy Flow's wearable hardware digitizes logistics delivery. Traditional giants are cutting benefits to survive, and new forces are using AI to reshape the capillaries of the physical world. Behind this double ice and fire is the fundamental shift in productivity factors: from the accumulation of steel and manpower to the refinement of algorithms and data.

This shift is equally cruel within the AI field. DeepSeek released DSpark to solve the problem of inferential toothpaste squeezing, pulling competition back from arguing parameters to optimizing system engineering, proving that saving money and efficiency are the key. At the same time, Anthropic accused Ali Qianwen of distilling Claude. On the surface, it was a technical ethics controversy, but in fact exposed the strategic anxiety of American AI companies in the face of the approaching low costs of China counterparts. As GLM 5.2 overtakes Claude on benchmark tests, the effectiveness of the West's attempt to maintain hegemony through narrative warfare is fading. The real moat is no longer the closure of the model, but the ability to continuously iterate engineering.

Putting these signals on a macro Timeline, Anthropic predicts that recursive self-improvement will come in 2028. From AI predicting that the World Cup victory rate will surpass humans, to the agent-based video production process, to the double squeeze of basic programming positions by AI and low-cost labor, AI is taking over complex prediction and generation work. When the large model solves the bottleneck of reasoning efficiency, the singularity of AI created AI will accelerate. In the next year, the job values of a large number of basic white-collar workers and junior code farmers will be reassessed, and those who can use AI to transform vertical scenarios will reap the biggest dividends. If traditional companies do not restructure their computing power structures from the bottom, they can only struggle in the quagmire of welfare cuts like Mercedes-Benz.

Anthropic calls for recursive self-improvement in 2028, GPT-5.6 was released with the strongest title but is in trouble, and AI predicts that the World Cup hit rate is only less than eight percentage points higher than humans. This contrast reveals the real situation of the development of large models, and the accumulation of computing power has approached the critical point of diminishing marginal returns. When AI is difficult to predict even a football draw, counting on it to independently build a stronger AI two years later is more like a narrative to maintain valuation. The giants peddling computing power make a lot of money, while the ones paying the bill are small and medium-sized enterprises that blindly access but see no return.

Behind the technological myth is the increasingly serious trust collapse and stock game. CCTV exposed fraud in mobile phone evaluations, and Anthropic accused competitors of distilling and stealing technology without substantial evidence. These are all products of traffic anxiety and fierce competition. When technology dividends cannot support high growth, marketing cheating and narrative suppression become shortcuts to maintain share. This is exactly the same logic as Mercedes-Benz's cancellation of year-end bonuses for 90,000 employees and extension of unpaid working hours. Whether traditional manufacturing or cutting-edge technology, giants are brutally reducing costs and increasing efficiency, trying to whitewash financial reports by squeezing the living space of employees and supply chains.

Before the historical bubble burst, it was always accompanied by the most ambitious vision and bottom-line marketing. Today's technology circle, while depicting the utopia of AI's self-evolution in the cloud, it is also haggling over traffic and layoffs in the quagmire. Now is the time to strip away the myth of general intelligence and examine technologies that can truly reduce reasoning costs and solve specific engineering problems. When the tide recedes, the ones who can survive are never the ones who are the most storytellers, but the ones who have the most cash on their accounts and the most stable systems engineering.