Transcription
Hello everyone. In this issue, the most important fresh news about artificial intelligence and technologies. The E sphere did not destroy jobs after all, but programming has become the only profession where models are truly changing the game. Chat GPT quietly received a powerful update. Enterprite code now works with twelve programming languages. Google paid $68 million after accusations that the assistant secretly recorded user conversations. Vitalik Buterin announced that he is abandoning Google and Telegram. The European Commission is launching an investigation against the Grogk chatbot on X. About this and much more in this issue, watch this video until the end so as not to miss anything. Google released a selection of free AI courses with certificates without payment, without subscriptions. I have collected the 10 best. Go and watch the selection in my Telegram channel via the link in the description under the video. 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 of prompt engineering and working with GNAI in real cases, as well as much more. Everything is in my Telegram via the link in the description under the video. Since 2022, top managers in the field of artificial intelligence have promised a revolution. They were about to destroy jobs. Automation will cover all professions, and people will be left behind. But four years of hype around generative models have passed, and mass unemployment has not happened. It turned out that the vast majority of professions live well without AI, and ordinary people encounter it only sporadically. But there is one industry that has felt the blow more than any other. Programming. Large language models have unexpectedly become excellent at writing code. Now a huge number of developers use AI assistants to speed up their work. And this has turned into one of the fastest-growing segments of the industry. The main star in this market is the company Enispher, which created the Cursor tool. In a couple of years, the company's revenue has soared so high that investors have valued it at $29 billion. But at the same time, the coding AI turned out to be so computationally voracious that startups like Cursor began to incur huge losses. Therefore, the Cursor model became a perfect example of how beautiful growth can hide a heavy economy under the hood. Cursor is an integrated development environment, or IDE, but with the connection of powerful language models. In the interface window, there is a special field where you can simply write a text query, and it will create the necessary code snippet. The user chooses the model themselves. Open Antropic, Google Gemini XY. Everything looks simple. It understands the project structure, navigates through dozens of files, knows how to maintain code style, and works with a large codebase at once. Something that cannot be done through ordinary communication with a chatbot and copy-pasting. By 2025, Cursor had become the favorite tool of millions of developers. In the spring, the company reported $100 million in annual revenue, and in the summer, $500 million. This is the fastest growth among all startups in history. But behind the impressive picture, serious problems were hidden. Cursor does not create its own model; it buys access to the models of major players, which means that every user request turns into real expenses for the company. Models are priced by tokens, small units of text. In 2022, CH GPT3 cost $60 per million tokens. In 2024, cheaper models like Lama 3.2 appeared, which cost 6 cents per million tokens. Theoretically, this should have reduced Cursor's expenses and turned it into a profit machine. But the opposite happened. New generation models turned out to be not just more powerful; they became much more expensive and voracious. The most advanced CH GPT5 Pro costs $21 per million input tokens and $168 per million output tokens. The average price is $94.5, which is higher than the 2022 level. But the main problem is elsewhere. Modern models actively reason, meaning they perform many hidden steps. When an engineer asks AI to add a function to a codebase of 200,000 lines, the model is forced to analyze the architecture, data structure, and connections between files and burns millions of tokens for one request. In tests, chat GPT 5.2 used 135 million tokens for a task, of which 130 million were for reasoning, the model's hidden work. This is how user whales appeared, who run Cursor almost around the clock on huge projects. The company charged them $20 per month, while it paid hundreds or thousands of dollars to power providers for their requests. The unprofitability became so high that in the summer of 2025, immediately after the report of $500 million in revenue and a $90 million investment round, Reuters wrote: "Cursor is operating at a loss." This turned out to be true. Therefore, the company switched to a new model. $20 per month turns into just a credit. And if the user overspends tokens, they are billed. This is how the "Pay for what you burn" system appeared. And then chaos began. Programmers started receiving unpredictable bills. In one case, just a few cents, in another, more than $10 for one request. Screenshots appeared on forums. One team sent a request that required 4.4 million tokens. This is as if the AI read the entire Harry Potter book series four times in a row. This is because the model scanned a huge project, tried to understand the logic, spent reasoning tokens, and, in essence, engaged in intellectual brute force. Working with Cursor turned into Russian roulette. No one knows if the next request will be fast and cheap or if a double-digit bill will suddenly appear. Moreover, AI can make mistakes, misunderstand the task, or produce hallucinations, forcing users to send a new request, and thus spend even more tokens. Despite this, developers continue to pay. In the fall of 2025, Cursor attracted a new round of $2.3 billion and grew to a valuation of $29 billion. The company announced that it had surpassed the $1 billion mark in annual revenue. Perhaps after the change in pricing policy, users decreased, but each of them began to bring in more money. However, the frequency of new rounds shows that expenses are still huge, and the economy is far from stable. This is a common problem for many AI startups that have not yet learned to turn computation into stable profit. The Cursor story became a perfect illustration of today's AI market. Technologies grow faster than the economy, models become more powerful but more expensive. Users expect magic but get the risks of unpredictable bills. Investors pour in record sums, but profit remains elusive. These are no longer early dreams of a future where AI will free people from labor. This is a reality where companies are trying to survive in the computation race. And so far, it's not the smartest who wins, but the one who can pay for tokens. NVIDIA takes weather forecasting to a new level. A S2 promises to predict climate two weeks in advance. NVIDIA has opened access to S2, a new AI platform that should make weather forecasting significantly more accurate and faster than any existing methods. The company emphasizes that this is not just another model, but a complete system capable of combining different algorithms and data to gather the most complete picture of what is happening in the atmosphere. The main component of S2 is the Medium Range model, which already outperforms the well-known weather engine Deepmant Gencast in more than seventy metrics, showing an advantage in accuracy and forecast stability. Forecastnet 3 works alongside it. It calculates wind, humidity, and temperature at speeds that meteorologists could only dream of before. Up to 60 times faster than classical approaches. And the Coreiv model turns rough global calculations into detailed regional maps, allowing you to see the weather literally at the district level. Nvidia has dedicated a separate direction to emergency situations. The Casting system generates ultra-short-term forecasts for the coming hours. This is exactly what is needed to track storms, downpours, and sharp weather changes. And the Global Data Assimilation module collects data from satellites, weather stations, and balloons, creating a live map of the atmosphere that is updated almost in real-time. Previously, such calculations were performed by supercomputers, half-loaded for a single forecast. Now, on Nvidia graphics cards, this takes mere minutes. R2 combines models from AMCWF, Microsoft, and Google, and is trained through the open Physics Nemo framework, making the system flexible and accessible to researchers worldwide. In essence, Nvidia is offering a new standard for accurate and fast forecasting, something that can change not only meteorology but also industries dependent on climate, from energy to aviation. GPT has unexpectedly received a powerful update that Open is not officially talking about yet, but which is already changing the perception of what an AI tool for developers can be. Simon Wilson noted that the built-in codetepray now supports 12 programming languages at once, not just the usual Python. And this is not partial support, but full functionality within a live environment. In fact, chat GBT has turned into a small universal def server that can be run directly in the chat. It executes bash commands, allows installing packages via PIP or npm, works with files, compiles and runs projects in Go, C++, Swift, Java, and other languages. Users have already shown that the assistant can create files, compile them, and provide results as if you had a full-fledged local machine at your fingertips. Now you can open a terminal directly in the chat GPT window and perform the same actions as in a regular development environment. The new ability to download materials from the internet deserves special attention. Chat GPT can find and download files itself through a secure container download mechanism. This allows you to quickly pull in data or dependencies without leaving the dialogue. The system checks the URL to prevent the download of malicious content. Most notably, all of this is available not only to subscribers of paid plans. The Cд Enterpriteer functions have been expanded even for free accounts if the user is working on model 5.2. This creates a new standard for EI platforms. Now, an ordinary user can run complex programs and experiment with different languages without installing additional tools. While Open has not officially announced this update, the community is already buzzing. Many note that chat GPT is turning into not just an assistant, but a full-fledged development environment accessible to everyone. And if the company offers to develop this functionality, the boundaries between chat and a working IDE may disappear altogether. Google has agreed to pay $68 million after accusations that its voice assistant Google Assistant turned on by itself and secretly recorded user conversations. The lawsuit was heard in the San Jose federal court, and plaintiffs argued that the assistant could be activated without the usual commands like "Hey Google" or "Okay Google." According to them, the collected recordings were then used for advertising targeting, which makes the situation even more sensitive, as it concerns data that people did not intend to share with the company at all. Google officially denied guilt but agreed to the payment, emphasizing that this step is primarily necessary to avoid long and expensive litigation. The agreement applies to users who, since May 18, 2016, noticed that their devices periodically activated the assistant without a command. It was precisely these cases that formed the basis for the class-action lawsuit that led to today's decision. This story has become another reminder that privacy issues remain a serious topic in the digital assistant industry. And Google is not the only company that has fallen into a similar scandal. In December 2024, Apple was already forced to pay $95 million for similar claims related to Siri's operation. Now, courts are increasingly considering such cases not as mere technical glitches, but as real intrusions into users' personal lives. Job seekers are increasingly tailoring their resumes to algorithms, and this is already noticeably changing the labor market. Today, not only HR specialists are involved in the initial selection stage. AI bots handle a significant portion of the filtering. People feel this, and more and more candidates are restructuring their resume submission strategy, focusing primarily on machines rather than live recruiters. A survey by the platform Monster, which involved over a thousand US job seekers, showed that the majority fear that their resumes will be automatically rejected and will never reach a human. The fear is so great that many have started stuffing their resumes with keywords and phrases from job descriptions, hoping to pass the algorithmic filter. This is already affecting the structure of documents. If it was previously considered normal to fit everything on one page, now more and more people are exceeding it. For many, one and a half pages or more is becoming the standard. Paradoxically, it takes less time to adapt a resume to a specific vacancy. Most people don't even spend half an hour on it. According to Monster experts, such speed indicates a superficial approach. People simply add keywords, slightly edit the text, and send their application. As a result, resumes become more software-oriented, as candidates try to adapt to algorithms that give weight not only to experience but also to correct phrasing. And this is not surprising. Only a negligible fraction of job seekers are confident that their resumes are read carefully. However, HR specialists warn that an obsession with keywords does not make a resume stronger. If the terminology is not supported by real experience, the document looks unnatural. And it is precisely such resumes that fail most quickly during detailed verification. Algorithms can miss text, but a person will eventually notice the forgery. The labor market is gradually restructuring to a new reality where job seekers have to work for two audiences simultaneously: AI filters and live recruiters. And the balance between these two requirements is becoming increasingly difficult to maintain. Artificial intelligence has once again collided with reality. Beautiful reasoning no longer works when you need to stay within real money, time, and logic. The Alibaba team has introduced a new rigorous benchmark, Deep Planning, which tests AI agents on their ability to plan complex tasks as a whole, from start to finish, with no room for error. This is where many models unexpectedly stumble. The slightest miscalculation, and the entire task fails. The first scenario of the benchmark is dedicated to travel and is intended to act as a personal assistant, creating a multi-day itinerary with precise schedules, considering budget, preferences like a three-star hotel, and coordinating flights, museums, and restaurants. For all this, it is given 9 A, airline tickets, accommodation, food, and attractions. The output should be a clear itinerary with a breakdown of expenses. Any violation of budget or time, and the result is nullified. The task seems simple, only in words: real logistics are constantly changing, and it has to balance on the edge of constraints. The second scenario turns AI into a thrifty shopper who needs to assemble an ideal basket of goods, adhering to many conditions, applying discounts, and skillfully combining promotions. Here, 15 A are involved, and the result should be the minimum final purchase cost. The main enemy is complex discount logic, where an error in one detail changes the entire basket. The results were unexpected. The best accuracy is with GPT 5.2X, which scored 44.6%. Second is Clod 4.5 OPUS, but only in detailed reasoning mode. Then it gets 33.9%, and GPT 5 completes the top three with 31.6%. As soon as the model's thoughtful planning mode is turned off, the scores drop sharply. This indicates that long-term planning without internal thinking is still unavailable even for top systems. The Alibaba experiment shows an important thing. Modern AI is impressive in conversations and text generation, but the real world with its constraints of time, budget, and many variables remains a serious challenge for them. To become full-fledged assistants, models still need to learn not just to reason, but to think strategically and calculate the consequences of their decisions. Clud is a new type of AI assistant that works directly on the user's computer and takes over routine digital tasks. Its creators call the product not just a smart chat, but a full-fledged work tool capable of replacing dozens of services and automation. The main difference is that the work is completely local; no data goes to the cloud, and the user retains full control over their device. The system can already connect to popular services and programs: messengers, calendar, browser, media players, files, and practically everything used in daily work. It has over fifty built-in integrations, and this is what turns clud into a universal assistant that can become a single control center for the digital environment. It not only opens applications but also performs actions within them, analyzes context, and offers solutions. Support for different neural network models is emphasized separately. The search itself chooses which AI is suitable for a specific task, from simple text answers to complex scenarios that include file processing, content work, and automatic material publication. In essence, cloud allows you to build your own digital infrastructure where AI acts as an executor, and a person as a director of processes. The creators emphasize privacy. Everything happens on the device without external servers and risks of data leakage. This architecture makes Cloud similar to a next-generation local digital assistant that takes care of everyday tasks and allows the user to focus on what is truly important. It is the combination of automation, flexibility, and privacy that makes it one of the most discussed AI agents recently. Open neural networks are ceasing to be something complex and frightening. Now they can be launched in a couple of clicks, just like a regular application. A new tool called Pinocchkiо shows that local AI can be as simple as installing a browser. It takes care of all the routine. No terminals, commands, mysterious errors, or struggles with libraries. Everything works as if you just click install and start using it. The main idea of Pinocchkiо is that it brings together a whole set of popular and powerful models under one roof. Here you can generate images through table diffusion or focus, run local analogues of GPT like i 3, transcribe audio using Visper, create music through Music Gen, and connect dozens of other tools. All of this is deployed on your computer and works without the internet, without subscriptions, and without any queues that stand between the user and cloud neural networks. Pinokki does what previously seemed inaccessible to most. It turns local AI models into understandable applications that are installed almost automatically. And if working with open-source software still required patience, skills, and caution not long ago, now it has become a simple and accessible way to try modern neural networks without limitations and hassle. The European Commission has launched an official investigation into the Grog chatbot operating on the X network. EU authorities have decided to check how the platform complies with rules for reducing user risks, as it concerns the most sensitive issue: the spread of dangerous and harmful content. European regulators want to understand how Grog controls the appearance of materials that could pose a threat to people, including images with inappropriate sexualization and content related to child abuse. The reason for the check was a wave of publications on VK, where users massively posted generated images of real women in nude or sexualized poses without their consent. This caused serious alarm among European authorities. According to The New York Times, in just 9 days at the end of December, at least 1.8 million such images became publicly available, a scale that is difficult for even the largest regulators to ignore. The company X later stated that it had restricted the generation of sexualized images. However, the European Commission wants to ensure that the measures taken are indeed sufficient and that such incidents will not happen again. If the checks reveal a violation of the law, the platform could face hefty fines and additional restrictions within the European Union. Moreover, Europe is not the only region where Grog has come under scrutiny. In early January, similar investigations were launched in Australia, the UK, Germany, and France. Previously, the service was already banned in Indonesia and Malaysia, although Malaysia later partially lifted the ban. All of this shows that the problem has become global. Regulators in different countries are trying to understand how safe the use of such tools is and whether companies are ready to take responsibility for the consequences of their AI systems' operation. Vitalik Buterin has decided to radically overhaul his digital life and has abandoned the familiar Google and Telegram services in favor of tools that offer much more privacy and control over personal data. The creator of Ethereum has seriously engaged in the topic of digital self-sovereignty and has published his own set of services with which he has replaced the familiar products of large corporations. The main motive is to minimize data leaks and use the most transparent and secure technologies. Almost everything on this list is decentralized and open-source. The key element of his new stack is File, a platform for working with documents, where files are stored without the involvement of centralized clouds and are protected by end-to-end encryption. In fact, Buterin has completely abandoned traditional cloud storage, betting on a more independent infrastructure. The same happened with messengers. Instead of Telegram, he now uses Signal as his primary communication channel. And for conversations where increased privacy is needed, he uses Simplex and Session. These services are designed to minimize data transfer to servers and leave control over correspondence with the user. The situation with maps is similar. Buterin has switched to OpenStreetMap and Organic Maps. Both projects work locally, without sending geodata anywhere. According to him, modern maps are too deeply integrated into the ecosystems of large companies, and he wants human movement not to become a commodity. He has also reconsidered email, now using Pro Mail instead of Gmail. Although he admits that ideally, a person could do without email at all if encrypted messengers were more convenient and universal. A separate topic is local artificial intelligence. Vitalik is experimenting with local LLMs on his laptop but is honest about the problems. The models exist. However, interfaces at the level of Google Translate or Photoshop are still lacking, and the system itself requires resources. The same Llama on a laptop consumes about 35W in continuous operation. According to him, there are already many tools for fully replacing cloud AI, but they still lack the convenience and simplicity that the mass user is accustomed to. Buterin's main conclusion is quite harsh. It is not necessary to transfer your data to centralized platforms. The possibility of living in a digital space without constant Big Tech control already exists. It is important only to start using alternatives and help them develop. >> Colleagues, it is necessary to implement artificial intelligence. And this is no longer an option, it is our basic infrastructure. Saw a point where AI can replace something or someone better? Implement AI. Routine processes must be automated with AI. You must transfer all your knowledge and experience, and the future belongs to AI. We need more AI chatbots, bots, robots, and hyobots, and be proactive, get out of your comfort zone, and implement AI. Every employee must become a champion in implementing AI in their area of responsibility, and and and.