AI training projects routinely cost millions of dollars and years of work. People wonder: why is it so expensive? The usual answer is that it’s not particularly expensive—especially considering that most teams don’t own their own hardware yet. Training AI has always been about massive data centers; that’s just the reality of the field. When you’re not immersed in it, the sheer scale can be hard to visualize. ...
In traditional companies, everything is built on a clear hierarchy: decisions are made at the top and executed at the bottom. This approach might work in a stable environment, but for an IT startup, especially in AI, it stifles growth. IT companies need adhocracy: a management model where competence and results are valued more than titles. It’s about flexibility over bureaucracy and speed over approvals. The value of an idea is judged by its effectiveness, not by the position of its author. ...
From my experience working with neural networks, it’s become obvious how two people can interact with the same model and get radically different outcomes. This isn’t a minor variation—it signals a fundamental departure from the search engine paradigm. We have entered a new era of interacting with artificial intelligence. AI doesn’t just aggregate data; it selects and synthesizes relevant information in response to specific requests. This changes the very nature of how we engage with information. Where the key skill was once finding data in search engines, it is now the ability to correctly formulate requests to an AI. ...
I’m often asked why international AI models, like those from OpenAI, consistently outperform Russian counterparts such as GigaChat. To understand the gap, we need to look beyond the code and analyze the foundational, structural challenges. Here are the key factors limiting Russia’s position in the global AI race. 1. The Compute Bottleneck Effective AI development at scale depends on raw computational power. Since 2022, access to essential high-performance NVIDIA chips (like the A100 and H100) has been severed. Training a model on the scale of GPT-4 requires a cluster of over 10,000 GPUs—a resource capacity that simply doesn’t exist in Russia. For context, Sber’s most powerful supercomputer, Christofari Neo, operates at around 12 petaflops, making it 50 to 100 times less powerful than the world’s leading AI research centers. ...
Users see a chat window. Sometimes voice, sometimes images. But behind this familiar interface lie radically different architectures and capabilities. Here are five key parameters that distinguish the top AI models in 2025: 1. Memory (Context Window) This defines how much information a model can retain within a single conversation. GPT-4o: 128k tokens (~300 pages of text) Claude 3 Opus & Gemini 2.5 Pro: Up to 1 million tokens (~2,000 pages) DeepSeek-VL Mini: ~8k tokens (~20 pages) More memory enables greater context and reduces hallucinations, but it also demands more powerful hardware. ...
OpenAI has released Codex CLI, an open-source AI agent for developers. This marks a quiet but significant victory for the open-source community. The tool allows developers to use natural language directly in the terminal—the agent interprets the request, then writes, executes, and tests the code. Most importantly, this entire process runs locally, without sending data to the cloud. With this release, the industry moves one step closer to a system that can independently understand, build, and deploy solutions. It underscores a critical point: the future isn’t just about choosing the right model, but about engineering the right architecture that connects thought → action. ...
Yesterday, OpenAI opened access to the GPT-4.1 API. It’s a refined version of their flagship model—faster and architecturally closer to the concept of ‘agents.’ In parallel, the company officially announced it is winding down GPT-4.5, its most resource-intensive model, due to its excessive complexity and support challenges. With GPT-4.5, it seems they hit an architectural dead end. We are at a point where models appear and disappear rapidly. They are becoming what they should be: tools, not landmark events. We have a growing catalog of specialized AIs: some calculate, others write code, plan tasks, or generate video. But the average user should not be expected to know and choose between every AI in existence. That paradigm defies the logic of good user experience. ...
Your Thought Process, Packaged Deep Research isn’t just another AI feature; it’s a fundamental shift toward an agent-based architecture. In this model, the LLM stops being a simple chatbot and becomes a co-author—an agent that independently searches, filters, validates, and structures information. What does this change? If you’re designing a business, a startup, or a product, you don’t have time to personally read 200 sources. Now, an AI agent does it for you. This frees you up to do the high-value work: to think, not just to search. ...
Scientists recently digitized a single cubic millimeter of a mouse’s visual cortex, a project that generated 1.6 petabytes of data to map 84,000 neurons and half a billion synapses. To put that into perspective, the number of synapses in that tiny piece of brain tissue is comparable to the number of parameters in large-scale AI like DeepSeek or GPT models. It’s significantly more than the 29 billion parameters in a model like GigaChat. This comparison is a useful analogy for scale and complexity: just as synapses determine a brain’s processing capacity, parameters define the “power” of an AI. ...
DeepSeek has once again followed its “quiet release” strategy, making its new DeepSeek-V3-0324 model available on Hugging Face without any major announcements. Instead of marketing hype, they’ve simply delivered a solution for the community to evaluate. I tested the model locally on a Mac Studio equipped with an M3 Ultra chip and saw impressive performance, generating over 20 tokens per second. This marks a significant acceleration for running capable models on local hardware, making it a viable option for developers. ...