Transcription
Hello everyone. In this episode, the most important fresh news about artificial intelligence and technologies. Open radically updated CH GPT. The depresearch function is now a full-fledged digital analyst. Elon Musk stated that conventional programming may disappear this year. Researchers at Peking University developed Fine R1, a fine-grained recognition system. DeepMind CEO Demis Hassabis shared how he lives in a mode of two working days in one day. About this and much more in this episode, watch this video until the end so as not to miss anything. Google released a selection of free AI courses with certificates, no payment, no subscriptions. I have compiled the 10 best. Follow the link in the description under the video to my Telegram channel to view the selection. In the free courses, you will learn how generative AI works and how it differs from classical machine learning. A foundation for working with artificial intelligence in daily tasks, prompt engineering, and working with GNI in real cases, as well as much more. Everything is in my Telegram via the link in the description under the video. And Open seriously updated part of GPT. The depressor function now works on the flagship GPT 52 model. In essence, the company combined everything most powerful into one brain for paid users. Older models, such as O3 and O4 Mini, were removed from this function, betting on a unified system. The update rollout began on February 11, and now DIP resource increasingly resembles not just a search tool, but a full-fledged digital analyst. The main changes are in the quality and controllability of search. You can now limit research to specific sites and domains to avoid drowning in random sources. Time filtering has appeared, which is especially important for topics where data quickly becomes outdated. In addition, the system supports connecting external applications via MCP, thanks to which research can rely not only on the open internet but also on additional tools. As a result, there is less junk and more relevant information in the results. The work process itself has also changed. The user sees which queries the agent performs during research and can intervene along the way, clarifying the task, adding new sources, or adjusting the search direction. The final report is generated in full-screen mode with a convenient interactive table of contents. The document can be saved in PDF or DOC format, making it suitable for study, work, or presentation. In terms of time, one research takes from 5 to 30 minutes. It all depends on the complexity of the topic. But now it is not just a quick answer in a chat, but a structured analytical material with sources and clear logic. Open is effectively turning Deep Research into a tool that can replace basic analyst work and simplify the preparation of complex reports for business, education, and media. Elon Musk has once again made a loud statement. In his opinion, programming in its usual form may disappear this year. He believes that the code itself is ceasing to be the main element of technological progress. Programming languages, Musk says, are not evolving, they are gradually becoming a thing of the past. And if artificial intelligence continues to develop at the same pace, then by December, Python, C++, or even compilers may become unnecessary. In Musk's vision, the future looks radically different. Artificial intelligence will create machine code directly without intermediate stages and translation from human languages. Compilation will not be required, the usual program structure will not be required. Binary code will become so optimized that a person will physically not be able to come up with a more effective solution. The machine will write for the machine faster and more accurately than any developer. According to him, the problem was never with the code itself. The code was just a necessary intermediary, a kind of tax for the fact that computers did not understand human speech. People had to learn syntax, rules, and system architecture simply because the machine could not perceive ordinary language. Now that AI has learned to understand us much better, this tax can be abolished. Further, according to Musk, the changes will become even more profound. If technologies like Neuralink become widespread, not only programming languages but also familiar work tools will disappear. There will be no keyboard, no screen, not even syntax. A person will simply formulate a thought, imagine the desired result, and the system will design the solution itself and execute it immediately. This is no longer automation of programming, but its complete rethinking. It's not about helping developers write code faster, but about the profession itself gradually dissolving. Years of study, mastering complex languages and architectures may become unnecessary. The gap between idea and its embodiment, according to Musk's design, will become practically zero. In such a worldview, a person no longer creates a program line by line. They imagine the result, and technology immediately turns the concept into a working product. This is a future where imagination becomes the main tool, and software is born almost instantly, without the usual barriers between thought and machine. DeepMind CEO Demis Hassabis has been living for many years in a mode he calls two working days in one day. In an interview with Fortune, he shared in detail how his schedule is structured and why this particular rhythm helps him maintain concentration and conduct research. According to Hassabis, he sleeps about 6 hours a day. This is less than the recommended norm, and scientists have long said that such a regime is not ideal for the brain. However, for himself, he chose this balance, which allows him to combine managerial work and scientific reflection. The first part of the day, the head of DeepMind dedicates to meetings and negotiations. His schedule is almost completely filled. Project discussions, strategic decisions, interaction with teams and partners. Such a tight schedule requires constant engagement and quick decision-making. In the evening, Hassabis tries to spend time with his family and have dinner at home. This is a kind of pause before the second part of the day. And then begins what he calls the second working day. From 10:00 PM to about 4:00 AM, he is alone with his ideas. This is time for reflection, creativity, and research. It is at night, by his own admission, that he feels a special surge of energy. "I don't understand how one can be creative at 4 AM, but around 1 AM I come alive," he says. Hassabis has been living in this mode for about 10 years. His path in science and technology is impressive. He founded Deepmind in 2010, and by 2014 the company was acquired by Google. Over the years, the DeepMind team has created advanced artificial intelligence systems, including Gemini and AlphaFold. In 2024, Hassabis received the Nobel Prize in Chemistry for his contribution to protein structure prediction. An achievement that many call a breakthrough in biology and medicine. The story of Demis Hassabis is an example of how an unusual daily routine can become part of a larger scientific strategy. For him, night is not a time for rest, but a space for ideas that change entire industries over time. Chinese developers have once again made the entire artificial intelligence world talk about them. The new KBIN 1 agent from the startup Feeling AI, in its very first appearance on the prestigious Terminal Bench 2.0 test, scored about 72.9% and immediately took second place in the world. For comparison, many top systems struggle to stay around 60%, meaning every additional percentage here is extremely difficult to achieve. OpenAI remains the leader with a score of around 77%. However, the very fact that a young Chinese team surpassed most major laboratories has become a significant event. Terminal Bench 2.0 is considered a kind of stress test for agents today. It shows whether a system can not only reason but also actually perform tasks within a computer, write code, run it, and fix errors. Brain 1 code focuses on practicality. The model does not waste time on unnecessary information but accesses specific code snippets and documentation that are truly needed to solve the task. If an error occurs, the system analyzes diagnostic data, consults examples, and quickly corrects the solution. This cycle of "write, check, fix" has proven to be effective and economical. In a number of tests, the agent used more than 15% less computational resources compared to competitors. Developers also emphasize flexibility. The agent can change its strategy during task execution and take into account previous experience. In game scenarios, it can break down a goal into steps, gather resources, create tools, build objects, and adapt its behavior depending on the situation. Combined with the previously introduced memory module M Brand, this creates a foundation for more complex and long-lived systems. Equally significant changes are occurring in the field of image generation. Alibaba has introduced a new model, Quen Image 2.0, which focuses not on spectacular but random images, but on precise adherence to the user's intent. One of the main problems with previous systems was that with long and detailed prompts, they lost half of the requirements, ignored individual scene elements, confused styles, or disrupted composition. In Quen Image 2.0, developers have tried to solve precisely this problem. The model can process up to 1,000 tokens of instructions. This means that the author can describe in detail not only the overall plot but also the arrangement of objects, atmosphere, style, lighting, character facial expressions, and even minor details in the background. At the same time, the system maintains the scene structure and does not turn the image into overloaded visual chaos. If it's a series of pictures, for example, a comic strip, the characters remain recognizable from frame to frame. Facial features, clothing, and proportions are preserved, which is especially important for storytelling. Special attention is paid to working with text in images. Previously, neural networks often distorted Chinese characters or created meaningless sets of symbols. Quen Image 2.0 demonstrates more correct rendering of complex inscriptions, including classical texts. This opens up additional possibilities for creating posters, covers, infographics, and advertising materials in Chinese without manual editing. The model also shows confident performance in various styles, from traditional ink painting to 3D stylization and macro photography with pronounced depth of field. In tests with detailed infographics, where each layer and element was described separately, the system correctly built the composition and conveyed textures. When editing, QN Image 2.0 can combine multiple photos, change backgrounds or character clothing, while maintaining natural transitions of light and color. According to international comparative tests, the model is among the strongest solutions in its class and competes with leading Western systems. As a result, Alibaba is strengthening China's position not only in AI agents and video but also in image generation, where not just efficiency but controllability and accuracy of execution become key. The development of Fine R1 by researchers at Peking University seems narrow and academic at first glance, but in reality, it's about a very practical pain that both people and neural networks constantly face. How to distinguish almost identical things when the difference is hidden in the details? If ordinary recognition answers the question: is it a plane or a car, then fine-grained recognition tries to understand what kind of plane it is. Conditionally, not just a Boeing, but a specific model. And among dozens of similar options. In civilian databases, there are hundreds of such types, and the external differences between them can be minimal. The shape of the nose, the number and arrangement of windows, the geometry of the tail fin, wing proportions, small elements on the engines. For AI, this is difficult for two reasons. Firstly, such classes strongly overlap in appearance. If the frame is taken from afar, in fog, or at an angle, the signal of differences becomes almost imperceptible. Secondly, in real life, there is rarely a huge dataset for each rare model. For many categories, you can collect not thousands, but literally a few photos. And at this point, most neural networks start to hallucinate or confidently confuse one with another. Fine R1 tries to reverse the usual logic. Not first train on a mountain of data, then guess, but learn to reason by features and rely on comparison. The system is designed so that it does not make a decision in one leap. It goes through several steps. First, it captures visual details that may be important for distinguishing subtypes. Then it forms a set of possible options, after which it compares them and only then selects the final category. In essence, it's an attempt to make the model behave like an attentive expert, not just looking at overall similarity, but checking key differences, even if they are very small. Especially impressive is how much data Fine R1 needs for training. The approach description emphasizes that the model can get by with approximately four training images per category. This is a very small amount for a classification task. Usually, when there are so few examples, models either overfit or memorize individual pictures and fail on new angles. Here, the authors achieve stability through a special training mode on similar triplets of images. The model sees an image of an object, then an image from the same category, and another image from a different category, but as similar as possible. This setup forces the system to literally look for what distinguishes almost identical variants and to reinforce precisely these subtle features. In tests, Ferr 1 was compared with well-known universal systems that are often used as a base for visual tasks. And although such models are strong on average, in fine-grained recognition, they often lose because they are used to catching large, obvious features and are poor at grasping micro-differences. FA1 on a number of datasets showed higher results precisely where this jewelry precision is required. Moreover, an important point is that the system can name the category directly, not just choose the correct option from a list of suggestions. This increases its value in real applications where the model cannot be offered five choices in advance. Why is all this needed beyond airplanes? There are more such tasks than you might think. In medicine, fine-grained recognition is distinguishing similar conditions in images, where an error can be costly. In industry, quality control, where a defect looks like a barely noticeable deformation or microcrack. In security and logistics, recognizing specific types of equipment, in biology, distinguishing species that look almost identical. And everywhere the problem is the same: there is little data, classes are similar, and the cost of error is high. Against the backdrop of general progress in video and image generation, Fine R1 stands out because it's not about making things beautiful, but about seeing precisely. This is another side of evolution. AI systems are gradually becoming not only creative but also more attentive to details, more evidence-based in their choice logic, and less dependent on giant datasets. And if such an approach scales, then in the future, we will get models that not only recognize an object but can explain exactly which features led them to their conclusion. And this is precisely what was lacking in many black boxes of the previous generation. Open announced a new phase of cooperation with the US government. CH GPT is being integrated into the corporate platform GEN and AMIL, used by the Department of Defense. This concerns working with unclassified information for approximately 3 million civilian and military employees. The project is positioned as part of a broader program to use advanced technologies and artificial intelligence in government structures. According to company representatives, the integration is carried out in a secure environment and is intended for everyday tasks. Chat GPT will be able to help specialists analyze large volumes of documents, prepare memos and reports, check procurement and contract materials, and simplify the preparation of internal documentation. The idea is not to replace people, but to speed up routine work and reduce the burden on employees. This new initiative continues the already established cooperation between Open and the Pentagon. Previously, the company participated in projects related to cybersecurity and experimental AI programs. Now, the technologies are moving to a broader level. They are becoming part of the department's overall digital infrastructure. Special emphasis is placed on security. Chat GPT operates in a government cloud with complete isolation from external systems. This means that user data does not leave the protected environment. Furthermore, the information that the Department of Defense employees work with is not used for training or improving other company models. This approach should alleviate concerns about leaks and unauthorized access. VPA emphasizes that the goal of the project is to increase the efficiency of administrative work and make processes more transparent and manageable. Integration into GenAMIL is seen as a step towards broader application of artificial intelligence in government structures, where reliability, control, and adherence to security standards are particularly important. Thus, CHG GPT is gradually becoming a tool not only for business and private users but also for large government organizations. If the project proves successful, similar solutions may become widespread in other departments where working with large volumes of information is required with strict data protection requirements. In one of the Munich offices, a regular workday was suddenly interrupted by a fire alarm. Employees heard the alarm and began discussing what was happening in the corporate chat. One of the colleagues wrote a message about a possible fire alarm to gather information and understand what was happening. The company uses an AI assistant built into Slack named GN. And it intervened in the dialogue almost immediately. The bot replied to the employee in private messages, stating that there was nothing to worry about and that it was supposedly a planned drill. The problem is that no drill was planned for that day. The artificial intelligence made a conclusion on its own, without access to the real schedule and without verifying the information. In essence, it simply assumed the most likely scenario and presented it as fact. Despite the bot's message, employees did not rely on the automatic response. People organizedly left the building and gathered at the designated assembly point outside. Later, firefighters arrived at the office. Fortunately, there was no serious threat, and everything ended without incident. However, the situation itself caused concern. In fact, the digital assistant advised ignoring the alarm signal, which directly contradicts basic safety rules. In emergencies, every minute counts, and such confidence from the system can play a cruel trick. If the fire had been real, the advice about a planned drill could have misled people and slowed down the evacuation. This case became a clear example of how even modern AI systems can make mistakes, especially when they draw conclusions without a complete set of data. Technologies are increasingly being integrated into work processes, but the responsibility for making critical decisions still rests with humans. And as this incident showed, sometimes it is the common sense of employees that proves more reliable than algorithms. YouTube has learned to create playlists for users with the help of artificial intelligence. Google has launched a new AI playlist feature for YouTube Premium and YouTube Music Premium subscribers. Now, you just need to write a few words about your mood or genre, and the service will select suitable videos or music tracks. The idea is simple: the user formulates a request in free form, and the system analyzes it and offers a ready-made selection. This could be music for an evening run, calm tracks for work, or, for example, energetic pop for a trip. The algorithm takes into account the user's preferences and tries to create a playlist that matches the description. This new feature is available in the YouTube Music app on iOS and Android. In the library section, there is an option to create an AI playlist. After entering a text request, the service takes about a minute to generate a list of songs. Then, the user can review the result and make changes if necessary, delete individual tracks, change the title or cover. Thus, the process of creating playlists becomes faster and easier. Instead of manually searching and adding each song, a short description is enough. Google is betting that such tools with AI elements will help people spend less time searching and more time listening to their favorite music. In Hollywood, an unusual way to combat neural networks that use third-party intellectual property has been found. Major studios are increasingly hiring private teams, essentially hunters, to track down services trained on protected content. This refers to characters, scripts, and visual imagery belonging to giants like Disney and Marvel. One notable player in this area is the startup Light Bar. The company is looking for specialists who will analyze the work of chatbots and image generators. Their task is to find cases where artificial intelligence reproduces familiar characters or styles, bypassing built-in restrictions. For example, if a system masquerades as a Disney princess or creates images of superheroes despite prohibitions, this is recorded as a potential violation. The work requires attention to detail and technical expertise. Employees test various services, gather evidence, take screenshots, and compile reports. If it is confirmed that the model was trained on protected material without the copyright holder's permission, the information is passed on to lawyers. In some cases, it can lead to a lawsuit. Financial motivation is also provided. A fixed reward is paid for each confirmed case of violation. And if the process ends in a lawsuit and compensation, investigation participants may receive a percentage of the recovered funds. Thus, the fight for copyright in the era of AI is gradually turning into a separate profession at the intersection of technology, law, and digital security. >> Colleagues, I want to raise an important issue. In fact, many of us regularly have overtime. Since this is the case, it must be officially registered and paid. Yerlan, I understand you, of course, but for example, yesterday I myself stayed up until one in the morning, didn't notice the time at all, and then at night there was a call. Guys from a neighboring department are panicking. There's a critical error in the report, but everyone understands that without me, nothing can be done. I got involved. Fixed everything. And you know, I didn't think about money for a second. I was also planning to go somewhere with my family on the weekend, but it's like, no. Nikita's help is needed. So I'm getting involved, I'm for the common cause. Arkady, and I think that when you're for the common cause, >> Nikita, you're great, of course, but I'm addressing the manager more now. Yerlan, this is a general meeting. I expressed my opinion, it's not forbidden. I just shared my story with colleagues. What kind of toxicity is this? Well, Yerlan, you see, everyone is coping, everyone is passionate about the project, and you are constantly the only one with some complaints.