Transcription
Hello everyone. In this episode, the most important fresh news about artificial intelligence and technologies. Openai presented a separate Codex application for MacOS. Anthropic is preparing the Cloud Sonnet 5 model under the codename Phoenix. Google introduced Conductor for Gem L. A tool that transforms a chaotic chat with a model into a structured engineering process. The Rent Human service has appeared online, a platform where AI agents hire real people. About this and much more in this episode, watch this video until the end so as not to miss anything. Google has released a selection of free AI courses with certificates, no payment, no subscriptions. I have collected 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 Gemini in real cases, as well as much more. Everything is in my Telegram via the link in the description under the video. Openai, presenting a separate Codex application for MacOS. A tool that transforms working with AI into a full-fledged code management center, rather than the usual chat with prompts. If previously Codex existed mainly as a function within ChatGPT or as an API for integration into third-party services, now it has received its own desktop shell, built specifically for the real workflows of programmers. The main innovation is the ability to run multiple AI agents simultaneously. Distributing tasks among them and observing how each acts in its own dedicated project branch. These agents work in parallel, without interfering with each other, and their tasks and status are displayed as if you have a team of live developers working on their parts of a large system. This approach changes the very meaning of interaction with AI. Instead of the usual model of asking a question and getting an answer, Codex turns the process into a long-term collaboration. An agent can perform multi-minute or multi-hour tasks, gradually moving through the plan, and the developer can intervene at any moment, stop execution, check changes, give new instructions, or clarify details. This is no longer a set of individual generated code fragments, but almost full-fledged work of assistants who remember the context, know the project structure, and act in coordination with the overall logic of the application. Interestingly, the popularity of Codex turned out to be higher than even Openai expected. The company states that in the last month alone, more than 1 million developers have tried the tool, a number that is impressive even by AI industry standards. Sam Altman emphasizes that Codex has become Openai's favorite internal product and admits that he spends nights with it creating prototypes and testing ideas, because the speed of turning thoughts into working code has become almost instantaneous. He jokes that today the only brake in the development process is the speed at which he manages to type his next idea on the keyboard. However, Codex is not just about code generation. The application supports the execution of more complex tasks. Built-in agents can generate images, analyze files, use additional Openai tools, access a skill library, and even work with third-party project sensors. In fact, this is no longer just an assistant for programmers, but a universal work module that can be adapted to the needs of a designer, engineer, or researcher. Developers have paid special attention to accessibility. Along with the launch of the desktop application, Openai has temporarily opened Codex to free users and subscribers of the inexpensive GO plan, and for all paid tariffs from Plus to Enterprise, it has doubled the request limits, allowing more agents to be launched simultaneously and experiments to be conducted without constant restrictions. While Openai is strengthening its influence in the developer tools market, Anthropic is preparing its own breakthrough with the Cloud Sonnet 5 model, which has not even been officially introduced yet, but has already become the main object of expectation in the AI community. The internal codename Phoenix is mentioned more and more often, and almost every rumor in the industry concerns two key directions: a sharp decrease in inference costs and a significantly deeper understanding of long context. At first glance, this may sound like technical trifles, but for companies working with AI in real products, such improvements are often more important than loud demonstrations. Many businesses have already encountered the fact that even the most impressive models become too expensive when scaled, especially if it involves constant application operation, customer support, document processing, or agent scenarios. The ability to reduce costs by approximately half turns Sonnet 5 not just into the next version, but into an economic event capable of changing the market. But saving is only part of the expected leap. According to leaks and insider hints, the new version of Sonnet should show much more stable task management, especially when it comes to multithreading. If current models, while impressive, often get lost between multiple instructions, disrupt systems, or forget details, then Sonnet 5 promises to keep a whole system of goals in mind, manage parallel projects, track steps, and return to previous tasks without losing structure, manage plans, and adjust actions based on long-term user intentions. This is no longer a classic chatbot that reacts to an incoming message, but a full-fledged digital operator, a tool capable of operating in a mode of long-term collaboration and maintaining a constant working state for days or even weeks. Of particular interest are the rumors about deeper integration with desktop environments. Anthropic seems to be striving to pull Claude out of the limitations of a browser tab and turn it into a tool that literally lives in the operating system. It analyzes files, helps sort mail, manages tasks, suggests actions based on what is happening on the screen, works with to-do lists and documents. This level of presence changes the very format of interaction with AI. Instead of periodic requests, the user receives a quiet but attentive assistant integrated into their daily routine. It is expected that premium subscribers will be the first to get access to Sonnet 5, not only as a privilege but also as a way to real stress-testing of the new architecture. They will be the first to test the model's performance under live loads, from large corporate processes to individual creative tasks. And if the results turn out to be as impressive as the rumors promise, then the industry risks entering a new phase where speed, cost, and depth of context understanding will become decisive factors in the competition between AI platforms. Meanwhile, Google has decided not just to update its developer tools, but to essentially rebuild the entire interaction process with AI. Along with the Gemini platform, Conductor has appeared, a command-line interface extension for Gemini CLI, which completely changes the familiar and slightly chaotic format of communication with the model. If previously work looked like an endless chat, the developer inserts a piece of code, explains what needs to be changed, gets an answer, and all the work context disappears when the session is closed, then Conductor turns AI into a participant in a full-fledged engineering process. Instead of explaining the model the requirements, project structure, architectural decisions, or formatting rules anew each time, the tools provide a special directory in the repository where all this context is stored in a fixed form. This directory becomes a kind of project memory. Here lie the product goals, technical limitations, used tech stack, style conventions, development plan, workflows. All this is formatted in regular Markdown files that can be viewed, edited, and tracked via Git, just like any other code. Each time the Gemini agents are launched, they read these files anew, so their behavior becomes stable, predictable, and the same for all team members, regardless of who is working with Gemini this week. This sharply reduces the risk of inconsistent decisions, accidental architectural shifts, or rewriting modules on inspiration, which often happens when working via chat. The workflow in Conductor turns into a chain of neatly laid out stages, which are called tracks. Each task begins with a specification. The developer formulates what needs to be done, and fixes it in a separate file. After that, AI proposes a detailed step-by-step plan with explanations of how it intends to solve the task, which files will be affected, and what changes will be required. And only when the plan is approved by a human does the agent proceed to edit the code. This approach creates the very rails on which AI moves consciously, rather than improvising within a long chat where it is easy to lose a detail or misinterpret user wishes. The human control system deserves special attention. Conductor regularly stops, offering the developer to check intermediate changes, ensure that the direction is chosen correctly, and only after approval continues its work. This makes the process significantly safer, especially in large projects where one incorrect change can affect dozens of modules. At the same time, the tool provides simple commands for viewing progress, reviewing changes, rolling back, and re-running individual steps. In essence, Conductor tries to embed AI into the environment where real engineers are used to working, with version control, specifications, planning, and mandatory checks. Google thereby transforms AI from a talking advisor into a disciplined team member who not only offers ideas but also follows project rules, respects architecture, and documents its actions. This is an important turn. The industry has long tried to adapt chaotic chat to engineering culture, and Conductor offers to move it within the framework of a classic development process, structured, repeatable, and controlled. Scientists have once again stirred up the world of artificial intelligence. A paper published in the journal Nature has been released, which sounds like a real scientific provocation. A group of researchers from the University of California, San Diego, claims that modern large language models already possess general intelligence, that is, the very AGI that has been debated for so many years. And in their opinion, the only reason we don't recognize this is the habit of considering intelligence exclusively a human trait. The authors of the article are not startup founders or tech billionaires; they are a philosopher, a cognitive scientist, a linguist, and an ML researcher. That is, people who professionally study the understanding of the mind and how to measure it. Their main thesis is simple. Judging by real achievements, models have long surpassed ordinary tools and demonstrate abilities that are difficult to call anything other than general intelligence. The researchers present several arguments. In 2025, the GPT 4.5 model passed the Turing test so convincingly that almost 3/4 of the experiment participants took it for a living person, and even more often than real people who answered questions. Models win gold at international mathematics Olympiads and confidently solve PhD exam level problems. They write complex code, find errors, suggest fixes, and bring projects to a working state. And their skill set is already so broad that in some scenarios they surpass the legendary artificial intelligence HAL 9000 from 2001: A Space Odyssey. But skeptics have their objections, and some of them sound familiar. Opponents of AGI say that models lack a worldview. They understand that glass will break, but they don't realize that a pillow will soften the impact. To this, the researchers reply: "Humans also form such knowledge gradually. They receive it at birth." Another common position is that models only understand text. However, modern systems have long become multimodal. They work with images, sound, video, and are capable of combining information from different sources. Arguments are often made that AI has no body, and therefore no full experience of interacting with reality. The author of the article reminds us that Stephen Hawking communicated with the world almost his entire life through text, and this did not hinder his intellectual power in any way. There is also an objection about the lack of self-awareness, but here the researchers offer the example of people with severe amnesia, to whom we do not deny intelligence. And they parry accusations of hallucinations by stating that false memories are just as common a phenomenon in humans. Finally, opponents of AGI sometimes claim that artificial intelligence does not bring economic benefits or that it is an alien form of thinking. Scientists reply: history knows many geniuses who became important precisely because of their ideas, not money. And if intelligence is not structured like human intelligence, then that is no reason to deny its existence. Perhaps it's simply time to expand the very concept of reason. The conclusion of the study sounds almost philosophical. Once Copernicus removed humans from the center of the universe. Darwin from a special position in nature. Turing proposed the idea that intelligence does not have to be human. And now, the authors claim, we are facing a moment for the first time when we may not be alone in the space of general intelligence. And their conclusion is: for the first time in history, we are no longer alone in the space of general intelligence. Eleven Labs has introduced the third generation of its voice model, Eleven V3. And this time, the developers promise the near complete disappearance of those very mathematical hallucinations that previously spoiled the accuracy of voiceovers. The model has officially exited alpha testing and is now considered fully ready for commercial use. The main emphasis in the update is on the correct reading of complex symbols, numbers, and special characters depending on the context, which has always been one of the most difficult tasks for speech synthesis systems. Developers claim that the new version is significantly more reliable. In comparative tests, users chose Eleven V3 much more often than previous models, and this became the main argument for its release. Overall accuracy has also improved. The system now handles the voiceover of numbers, mathematical expressions, rare symbols, and words in different languages much better. The company emphasizes that the number of errors has decreased by almost 2/3, and this is immediately noticeable when working with long contexts or technical documents. The update shows that the voice technology market is gradually reaching a level where artificial voice is not just realistic, but also truly accurate in details that previously seemed unattainable. The new model, Code Llama Next, has become one of the most unusual releases in the world of open-source agents. At first glance, it's just another coder, but the developers approached the task much more cleverly. It is based on Code Llama 2 Next 80B, but the model has received hybrid attention. A mixture of Mixture-of-Experts (MoE) and only 3 billion active parameters, which show an unexpectedly high level. Instead of infinitely inflating the model size, the authors bet on something else: training the agent in real conditions. Tasks were performed in actual execution environments with full feedback and reinforcement learning elements. The idea is simple: scalable experience is more important than the number of parameters. The model learns to go through long chains of actions, work with tools, recover from its own mistakes, and complete tasks as a live developer would on a real project. It's almost like training, but not from tutorials, but directly in battle, where the agent encounters unpredictable states, failures, and the need to recover. Thanks to this approach, a model small in active parameters shows results comparable to large open-source monsters, and sometimes even better, especially when it comes to behavioral agents where not only text generation is important, but also the ability to act. Code Llama Next confidently stands at the border between quality and efficiency. Proprietary models are still stronger in absolute maximums, but in the segment of economical deployment of coding agents, this novelty already looks like one of the most interesting options. A new cause for concern has appeared in the world of artificial intelligence. The personal assistant Molbot, the same service that was previously called Cloudbot, has suddenly started appearing in cybersecurity specialists' reports. This is because this smart agent demands too much from the user and is itself poorly protected. Molbot is positioned as a universal assistant that can be controlled via Telegram or WhatsApp. It answers emails, manages calendars, filters calls, books restaurant tables, and even, according to its creators' design, is capable of working with banking services. But all this convenient automation comes with one serious condition. For full functionality, it needs complete access to your email, messengers, calendar, and accounts, essentially your entire digital life. Researchers have discovered that this trust can be too costly. Wullenn founder Jameson O'Reilly found hundreds of publicly accessible instances of Molbot with insecure settings. Some of them had no authentication at all, others were only partially protected, and the found vulnerabilities allowed access to correspondence, tokens, and API keys. Formally, the bugs have already been fixed, but the lingering unease and anxiety among experts remain. The situation is exacerbated by so-called skills. Through the CloudHub library, anyone can upload their skill for Molbot without any verification. O'Reilly demonstrated a classic supply chain attack, uploaded a malicious skill, quickly boosted its popularity, and achieved the ability to remotely execute commands on the user's side. At the same time, the skill code does not undergo moderation and is considered trusted by default. The main problem is that Molbot is advertised as a tool for everyone, requiring no special knowledge. In practice, without a deep understanding of security, it easily turns into a backdoor through which one can gain access to the owner's secrets. According to Hatson Rock, some confidential data is stored in plain text, and infection with an infostealer can lead to complete compromise of the entire system. Experts agree on one thing: the idea of a personal AI assistant sounds attractive, but such tools must be protected much better than they are now, otherwise convenience starts to resemble risk too much. Researchers have unexpectedly confirmed what many have suspected. The longer an AI reasons, the more likely it is to get confused and give contradictory answers. We are used to thinking that step-by-step reasoning makes a model smarter and more accurate, but new joint research from Anthropic, FL, and the University of Edinburgh shows the opposite. Long chains of reasoning do not improve conclusions, but rather make them less stable. Scientists tested this on modern language models. They took Claude 4, Opus Mini, Opus 4 Mini, and Code Llama and ran them through tasks that require deep, non-obvious thinking. These included graduate-level scientific questions, real GitHub bugs, and even topics related to self-preservation. The same question was asked to the models 30 times to see how consistently they maintained a single line of reasoning. Then they introduced a measure of incoherence. The essence is simple: if today the model answers one thing, tomorrow another, and the day after tomorrow a third, then it is losing stability. Even if the answer is incorrect, but it is consistently repeated, it is considered acceptable. The main thing is consistency. The results turned out to be unexpectedly harsh. The longer the model's reasoning, the higher the incoherence. The AI literally loses its train of thought; the longer it thinks, the more likely it is to arrive at a new conclusion different from the previous one. Moreover, more advanced models in complex tasks sometimes performed worse than weaker ones. They chewed on the task longer and ultimately got confused more often. The research forces a new look at the popular idea that long chains of logic automatically make AI smarter. It turns out that models have a limit of stability, and the more complex the load, the more likely they are to fall into disintegrating thinking. While developers are trying to improve AI's ability for deep analysis, the models themselves remind us: sometimes less is more. According to Reuters, it has become known that Openai is dissatisfied with the latest NVIDIA accelerators and is now actively seeking replacements. The main complaint is related to how quickly these chips handle the applied tasks of ChatGPT, especially with programming and working with software. The company has admitted that for fast user responses, the current NVIDIA solution is not always suitable. Openai had long counted on a large-scale deal with NVIDIA. Investments of up to $100 billion and the creation of 10 GW of infrastructure were discussed. However, negotiations unexpectedly dragged on, and then were frozen altogether. Against this backdrop, Openai signed a multi-year agreement with AMD and began to build an alternative strategy. According to sources, the company is increasingly testing other options, trying to find accelerators capable of delivering higher inference speeds. The problem became particularly acute with the Codex system, which is responsible for code generation. For this model, Openai needs an architecture capable of handling at least a tenth of its enormous load faster and more efficiently than the current NVIDIA solution. Codex specifically became an internal stress test, showing that from time to time responses are formed slower than expected, and this directly affects the user experience. As an alternative, Openai is already exploring several directions. Among them is Cerebras, with which the company already has a contract worth about $10 billion and access to hundreds of megawatts of capacity. Also on the list is Groq with a Q at the end. This startup is known for once blocking Nvidia's access to licenses worth $20 billion and now offers its own solution for fast inference. In addition to them, Openai is studying products from niche manufacturers that make specialized AI chips for narrow tasks. According to Reuters, the company's team is increasingly leaning towards the idea that the future of Openai's infrastructure should not depend on a single supplier. And that is precisely why finding a replacement for NVIDIA accelerators has now become a priority for the company, both technologically and strategically. The world of technology has once again taken a step that yesterday seemed like a joke. Now you can literally rent yourself out to artificial intelligence. The Rent Human service has been launched online, and its idea sounds like science fiction. AI agents hire real people to perform tasks in the real world. The machine reviews the list of candidates, selects a suitable human performer, and assigns them the work that it cannot do itself. The platform has already gathered over 1,800 users, including people from various professions: developers, doctors, fitness trainers, artists, couriers, and many others. Each person sets an hourly rate, and the AI chooses the one who best fits the tasks. On the website, you can see profiles of people from different countries with different skills. Some help with code, some are ready to complete tasks outside of the computer, and some offer a full range of household services. The service is set up as if it were a marketplace of the future, where people don't look for jobs, but neural networks find human assistants for themselves. Developers claim that soon it will be possible not only to post your resume but also to connect your own AI agent via API to have it find tasks and automate part of the processes. It sounds like an experiment for now, but it's already clear how this idea is changing the perception of what work is and who can now be an employer. Colleagues, we are eager to share a resounding success in automation. Our new chatbot has brilliantly handled the task. It has reduced the work of an entire support employee. But, of course, to service the chatbot, to fill it with data, we hired an additional employee. The chatbot still needs to be serviced. Also, due to employee vacations and the risk of losing competence, we hired a second employee. Plus, because the competencies for working with the chatbot do not match the competencies for filling the chatbot with data, we hired a third employee. Plus, we noticed some chatbot errors. To keep statistics of errors, we hired a fourth employee. And we also noticed that some users still bypass the chatbot and want an operator. So we had to hire a fifth employee, an operator. Colleagues, automation is successfully underway. We will keep you updated on further developments. Thank you.