Hello, people, welcome to the regular newsletter of Techcrunch. If you want this in your inbox every Wednesday, register here.
You may have noticed that we were skipping the newsletter last week. the reason? The chaotic artificial intelligence course made more by Deepseek to the lead, and from the industry and the government in practice.
Fortunately, we went back to the right track – not a very early moment, taking into account developments at the end of last week from Openai.
The CEO of Openai Sam Altman stopped in Tokyo to have a stage on stage with Masayoshi Son, CEO of Japanese Softbank. Softbank is an Openai main investor and a partnerAfter I pledged to help finance the huge data infrastructure project in Openai in the United States
Perhaps Al -Taman felt that he owed his son a few hours of his time.
What is the billionaires talked about? A lot of sincere work is far from “agents”, in used reports. Sun said his company will spend $ 3 billion annually on Openai products and will cooperate with Openai to develop a platform, “Cristal (SIC)”, with the aim of automating millions of traditional workflow with white collar.
Softbank said in A. Monday press release.
I ask, though, what is the humble factor to think about all this automation and self -independence?
Like Sebastian Simiakovski, CEO of Fintech Klarna, who often They boast about artificial intelligence solutionIt seems that the son believes that the positions of the position of workers can only nullify the wonderful wealth. It is polished is the cost of abundance. If the wide automation of the jobs comes, Unemployment on a huge scale looks the most likely result.
It is frustrating that those who are at the forefront of the artificial intelligence race – companies like Openai and investors such as SoftBank – you choose to spend press conferences in drawing a picture of automatic companies with fewer workers on salaries. They are works, of course – not charities. And the development of artificial intelligence does not come cheap. But maybe people You will trust artificial intelligence If those who direct it appear a little more anxiety for their well -being.
Food to think.
news
Deep search: Openai “Agent” has launched a new designed to help people conduct complex and in -depth research using Chatgpt, a Chatbot platform on behalf of artificial intelligence.
O3-Mini: In other Openai news, the company launched a new AI model “Thinking”, O3-MINI, after a preview last December. It is not the strongest model for Openai, but O3-MINI is proud of efficiency and improved response speed.
I prohibit the risks there As of Sunday in the European Union, the mass organizers can prohibit the use of artificial intelligence systems that they see as an “unacceptable danger” or harm. This includes artificial intelligence used for social registration and offensive advertising.
Play on “Doomers”: There is a new play about “Domer” culture, which relies loosely on the expulsion of Sam German as the CEO of Openai in November 2023. My colleagues, Dominic and Rebecca, participate in their ideas after watching the first show.
Technology to enhance crop yield: Google X “Moonshot Factory” has announced this week its latest graduate. The only agriculture It is a start -up company that relies on data and machine aimed at improving how to grow crops.
Search paper in the week
Thinking models are better than average artificial intelligence in solving problems, especially science and mathematics queries. But they are not a silver bullet.
A A new study from researchers at the Chinese company Tincent It examines the issue of “thinking” in the models of thinking, where the premature models, incomprehensible, are promising intellectual chains. According to the results of the study, “thinking” patterns are repeatedly tending with more difficult problems, which leads to a switch between thinking chains without accessing answers.
The team suggests a reform that uses a “thinking switching punishment” to encourage models to “develop” every line of thinking before looking at alternatives, which enhances the accuracy of the models.
Week model
A team of researchers supported by the owner of Tiktok bytedance, the Chinese AI Moonshot, and others, has released a new open model capable of generating relatively high -quality music from claims.
The form, called YesA song up to a few minutes can be removed with singing and support paths. It is under APache 2.0 license, which means that the model can be used commercially without restrictions.
There are negatives, however. Yue operation requires a fat graphics processing unit. It takes six -minute a 30 -second song with NVIDIA RTX 4090. Moreover, it is not clear whether the model has been trained using copyright data data; She did not say its creators. If it turns out that copyright -protected songs were already in the model training set, users may face future IP challenges.
Grab a bag
AI Lab Anthropic claims to have developed a more reliable defense technology against “Brucks”, methods that can be used to overcome measures of artificial intelligence system.
Technology, Constitutional worksIt depends on two groups of “classified” models of artificial intelligence: “Enter” and the “output” compiler. The input workbook demands to enroll models with protection with molds describing prison generations and other unanimous content, while the output compiler calculates the possibility of discussing a response from the harmful information form.
Anthropor says that constitutional works can liquidate the “overwhelming majority” from prison operations. However, it comes at a cost. Each healthy query is 25 % more mathematical, the protected model is 0.38 % less likely to answer unpredictable questions.
