Laughing to death, I have a group friend who works overtime at NIO every day and still has to increase his workout. A while ago, he suddenly had blood in his urine in the middle of the night, went to the hospital for emergency treatment, and was directly sent to the ICU. His creatine kinase shot up to 1.86 million... (Image 1) Then today I found out that the nurse on duty used him as material to post on Douyin (Image 2) Hahahahahaha.

The Wall Street Journal's report on the AI talent war in Silicon Valley is both smooth and concise, making it highly readable: - Last Friday, hundreds of employees from the AI startup Windsurf gathered in the office to prepare for a party, celebrating their impending acquisition by OpenAI for $3 billion, with the marketing team even ready with promotional videos; - Then an unexpected event occurred: the company's CEO quietly took a small group of "chosen" employees to join Google, where they received seven-figure transfer fees, while the remaining employees were left behind in the now-empty company, and upon hearing the news, many employees began to cry; - By the morning of the first working day, another twist emerged: the remaining Windsurf employees were called into an office again and informed of a second notice that they, as the remaining part of the company, would be acquired by a former competitor; - This was just an ordinary weekend in Silicon Valley in 2025; - The war among the richest companies on Earth for talent is unfolding with unprecedented efficiency, transactions, and betrayals, where those smart researchers, who were never valued so highly, have seen their worth reach the level of top NBA players and Hollywood stars; - At the center of the storm is Meta, where Zuckerberg has started a frantic talent hunt using a time-sensitive bidding strategy to assemble an AI "dream team"; invited researchers must make a decision on the same day, or the offer expires, rendering the traditional negotiation strategies of the companies being poached completely ineffective; - As the main competitor being poached, OpenAI's Sam Altman referred to this behavior as the "battle of missionaries and mercenaries," believing that those researchers who join Meta for money are mercenaries blinded by cash, while those who choose to stay at OpenAI are steadfast missionaries; - Zuckerberg took this comment to heart and wrote a post in response, stating that the key to Meta's ability to attract talent lies in providing researchers with the massive computational resources needed to achieve breakthroughs; when it comes to per capita disposable computing power, no company can match Meta; - However, many have proven that money is not the most important thing, including OpenAI's chief researcher, Mark Chen, who, during a meeting, was offered a total package of $1 billion by Zuckerberg; Chen hesitated for a moment and replied that he was quite happy at OpenAI and would not consider it for now; - At least 10 employees from OpenAI also rejected Zuckerberg's offer of $300 million over four years—of which $100 million would be deposited in the first year; - Chen said it felt like watching a burglary happen and helping to calm the employees, stating that management was not standing by; a new compensation incentive plan would be introduced soon, "If other companies put pressure on you, like absurd offers that expire if not accepted, just tell them not to push; this is an important decision, and pressure is disrespectful to you"; - However, the conversation with Chen indirectly inspired Zuckerberg to consider throwing money at the problem; he subsequently compiled a list titled "Desired Candidates," and the team obtained contact information for these individuals and arranged for them to meet Zuckerberg at his home—rather than at the company; - Ultimately, Zuckerberg selected Alexander Wang, the 28-year-old founder of the data labeling company Scale, to lead Meta's super AI lab, spending $14 billion as a transfer fee; Wang announced to all employees at Scale's headquarters that he would be leaving the company to join Meta, and amidst thunderous applause, the scene resembled "the ending of a Disney movie"; - Almost immediately, OpenAI and Google cut off contracts with Scale, and Scale began layoffs; - Meta also recruited the former CEO of GitHub and the current CEO of SSI, who had previously co-managed a venture capital firm; Zuckerberg generously bought 49% of this company to help investors recover their investments, then cleanly dissolved it; - If you remember, SSI is the new company founded by Ilya Sutskever, the chief scientist who was ousted from OpenAI during its internal power struggles; Zuckerberg invited Sutskever to dinner at his home and proposed buying SSI, but after being rejected, he poached SSI's CEO; Sutskever was "very shocked" to learn that his colleague had left; - Returning to the Windsurf story, the reason the deal with OpenAI failed was due to opposition from OpenAI's largest investor, and after the CEO left the company to join Google, the interim CEO received an email from Cognition (the developer of Davin), another AI programming company, with a simple subject line: "Chat?"; - Over the weekend, Windsurf finalized the acquisition plan, and the interim CEO informed the employees that everyone would benefit from the deal; this time, the long-awaited applause finally erupted.

She looks at me, is it interesting to me?

If a bill can achieve this kind of exponential growth (I have seen that there is a doubling of GDP), then economics really does not exist, and the bill itself is public, and every country will copy a copy and take it home. MAGA complains that Musk has objections but does not choose to communicate privately, and must make a public quarrel with relatives and hatreds is low emotional intelligence, I think he really can't convince himself to pretend to believe these JB eggs, how can he tell fairy tales seriously?

By the way, as a result of Meta's crazy spending money and poaching people in the circle of AI researchers, there is also a very subtle structural change, that is: Yann LeCun, Meta's chief AI scientist, one of the Turing Triumvirate, and creator of convolutional neural networks, 64-year-old Da Niu, now, in fact, is reporting to Meta's newly appointed AI leader, Alexandr Wang, a 28-year-old MIT dropout who previously worked on data annotation. At the moment when the big wave hit, qualifications, seniority, and honors were all set aside. Many people are pushing the feelings of "caring" about Yann LeCun, who is an old man who is calm and calm, feeling that he should have stayed in an academic position, and it doesn't matter who the boss is. There is also speculation that Yann LeCun will resign within a year, in addition to the reason for the heart, it is also related to the fact that Yann LeCun himself has actually been negative about the development direction of this wave of large models, and if he firmly believes that Meta's gambling architecture is wrong, then it will be even more uncomfortable to work.

It's really fun to get to know everyone, but something has happened recently due to personal reasons, and this should be the last day of my first half of the year on Twitter, thank you all for being with me for so long.

Following the last time many media took the large model to do the college entrance examination math questions, this year's college entrance examination real test results have also come out, but this time it is the byte Seed team that is evaluating it, it is obvious that they are full of surprises and surprises for the last college entrance examination math results of the bean bag, and they are also gearing up to try to see how it will perform in the whole subject test questions. As a rule, the 5 large models participating in this competition are still among the top players in the market in terms of various running scores: Gemini 2.5Pro, DeepSeek R1, Claude 4, OpenAI o3, and Seed1.6-Thinking. The college entrance examination is actually a very typical test scenario, which can not only achieve the purpose of testing the generalization ability of large models, but also have high practicability, and it is not difficult to imagine similar use scenarios, which should play a greater role in teaching, scientific research and other aspects. In order to ensure fairness, not only is the evaluation method completely aligned with the college entrance examination paper method, and the open questions are evaluated by two high school teachers with senior experience in the joint entrance examination, and the large model has not introduced any prompt word engineering, and all inputs are the original questions of the college entrance examination. To put it simply, this is almost no different from the environment in which a real candidate makes a paper. Let's talk about the conclusion first, the overall ability of the large model has been able to get a fairly high score, among which Doubao and Gemini, which rank first in arts and science, have scored 683 in liberal arts and 655 in science respectively - this score can even impact Qingbei - intercept some key information in the evaluation results for you to see: - Unsurprisingly, the subject with the best performance in the large model was English, and the scores of several assessments were very close, making it difficult to close the gap; - The subjects with the lowest scores are chemistry and biology, but this is related to the test questions themselves, these two subjects involve a lot of picture reading questions, because this set of test questions is not officially released, so some of the pictures are more vague, which directly causes the loss of marks for the large model; - Doubao got the highest scores in the six subjects of Chinese, English, Physics, Geography, History, and Politics, and its advantages in geography, history, and politics in the Chinese subjects were obvious. - Contrary to the bean bag, Gemini's science performance is very strong, even in the case of unclear figures, chemistry and biology still get the highest scores, and the difference between its route and the bean bag is very valuable to observe; - There is a slight difference from the results of the last college entrance examination mathematics assessment, this time the top of the mathematics subject is DeepSeek, but it is also very small behind other large models; - GPT o3 started to work again,Last time it was the only large model that lost points on mathematical objective problems,As a result, this time the Chinese composition was written directly off topic,This led to the o3 language score becoming all the large model language and mathematics English main subject 15 results,The only one that did not exceed 100... - After discovering the problem of reading pictures in chemistry and biology, the test team found a higher definition version of the test paper, and redid the two subjects in the way of interweaving pictures and texts, and found that the total score of the two subjects could be improved by about 30 points in this way, which means that the full-modal reasoning of graphic and text synchronization can stimulate the potential of the model to a greater extent, which is worth further study. - AT PRESENT, IT SEEMS THAT THE VISUAL SOLUTION OF THE LARGE MODEL IS PROGRESSING RAPIDLY, BUT AFTER ALL, THE VISUAL TOKENS CONSUMPTION IS MUCH HIGHER THAN THAT OF ORDINARY TASKS, SO THE MAIN PROBLEM IT NEEDS TO SOLVE AT THE MOMENT IS HOW TO REDUCE COSTS; - In addition to the real questions of the college entrance examination, the test team also conducted another Indian Institute of Technology JEE Advanced exam assessment, the questions are all input by pictures, the total score is still Gemini and Doubao leading, and even the results of the two models can already enter the top 10 in India. For example, at the beginning of this year, when the well-known HLE benchmark in the AI industry first appeared, the scores of major mainstream models were generally lower than 10%, but the HLE development team also said that according to historical rules, this number may be 50% by the end of this year. What I mean by this is that no matter how much human beings rack their brains to get AI to do the questions, it will be a matter of time, but this does not mean that the results are meaningless, the important thing is the process of continuous learning, and it is also a necessary step to explore the extent to which AI can be used by humans. We are already looking forward to next year's college entrance examination, what kind of vision AI will give us.

3 months ago, the byte essay actually had a follow-up...

Coatue, a hedge fund company, has released its annual investment market report, with a total of 102 pages of PPT, which is very informative. From a macro point of view, the cycle of technological innovation is getting faster, and new opportunities are gradually changing from once in 10-20 years to once in 3-5 years, and the influence is also greater from generation to generation; （1/n）

Keling's Orange Cat Theater, that is, the pixel worker of dreams, and Veo 3's Everything Can Cut ASMR, the three major video models have each produced short video hit templates on the creative side.