Artificial Intelligence

I'll tell you about the ways you can install the Chinese model locally and use it without the internet. And do it in a way that you work not through the console, but through a user-friendly chatbot interface.

The most interesting finance and tech news from Russia and the world for the week: RKN blocked FaceTime, Snapchat, and Roblox, visa-free travel with China and Saudi Arabia, Russia was added to the EU's money laundering blacklist, home surveillance cameras were hacked in South Korea, Musk's Twitter was fined in Europe, and rumors of a 'garlic' model from OpenAI.

I'm sure many Habr readers are familiar with this method, as it's not new. To save you time, I'll just say two words: comss DNS.

Everyone else - welcome under the cut :-)

A collection for those who have tried using neural networks for their tasks but were disappointed: it's unclear how a chatbot can help with anything serious.

Many readers, especially those living in Russia, want to know the best way to use GPT chat. When I'm asked about this, I give a clear answer: if the volumes are not very large, it's better to use Telegram bots.

GPT chatbots on Telegram are analogous to applications that use GPT (Generative Pre-trained Transformer) technology to dialogue with users. Here are the secrets to their popularity:

Now that Forbes Education is dispelling our fears that artificial intelligence will replace artists, and Gizmodo is exposing AI modelsfor simply copying training illustrations, we can relax and continue drawing cats riding dinosaurs. Which is what I'm going to do today!

I tested dozens of Telegram bots and chose the five best services for generating AI images. This will be useful for those who spend a lot of time in the messenger and want to create unique visual content right there.

In my previous Russian-language article on Machine Learning as Alchemy, I discussed the possibility of discovering novel solutions without relying on GPUs or expensive computing clusters. In this article, I will share my experiments with continual learning and the compositionality of thought using micro-neural networks, and explain what the philosopher Lev Vygotsky has to do with it all.

If you are a new developer, you have probably hit the “subscription wall.” You have $20 a month to spend on an AI coding assistant, but you don’t know which one to pick.

The Chinese neural network DeepSeek has recently gained popularity. Today, we'll figure out how to work with its API, how to pay for it from Russia, and whether there are any difficulties.

Image generation by neural networks has become a 'regular button' alongside familiar design tools. Today, you can create an image from a description in Russian, right in your browser, often without registration and, importantly, for free. Such a 'free image generator' is useful not only for designers: entrepreneurs create product cards and hero banners, SMM specialists create ad creatives and stories, journalists and bloggers create illustrations for their materials, and developers create interface prototypes and game mockups.

Why has this topic become so popular?

AI app builders have become very good at generating interfaces. From a technical perspective, that part is mostly solved. The harder problem is what comes next: connecting real data, enforcing user permissions, handling workflows, and deploying the app after the generation.

Much attention is currently focused on the size of neural networks and the gigawatts of power consumed by data centers. However, the future lies not only in giant clusters but also in tiny chips embedded directly into the sensing elements of hardware. When a neural network is placed directly inside a sensor chip, it must be exceptionally efficient.

Through experimentation, I have successfully built a neural network architecture with 380 parameters (with potential for further reduction), capable of operating in conditions considered unsuitable for conventional algorithms.

This guide covers how to set up OpenClaw (formerly Clawdbot) on your local machine and, most importantly, how to secure it so strangers can’t access your computer. If you are ready, then let’s get started! :)

It’s surprising that there’s almost no practical information about Telegram Cocoon beyond what’s on GitHub and the official website. Various media outlets have plenty of general coverage about the network launch, but almost nothing about real user experience.

I decided to spend a bit of time and figure out what’s actually going on in the network, how it works, and, most importantly, whether I, as a developer, can use it today. So in this article I’ll look at Cocoon from a developer’s perspective: how to install it and how to use it.

Привет, Хабр! Меня зовут Юра Петров, я руководитель отдела разработки компании Friflex и автор канала «Мобильный разработчик». OpenAI на днях выпустила крутую статью, где впервые подробно описала работу своего агента для написания и изменения кода — Codex CLI. Сердце системы — «агентский цикл» (agent loop).

Это процесс, в котором модель получает задачу от пользователя, при необходимости вызывает инструменты (например, запускает команды в терминале), анализирует результат и повторяет цикл, пока не получит финальный ответ или не внесет нужные изменения в код. Статья фокусируется на том, как устроен этот цикл, как формируются запросы к модели и как система управляет контекстом.

In modern neural networks, including Transformer-based LLMs, unbounded activation functions—ReLU and GELU—have become the standard. Their main advantage is good gradient flow and the rapid training of deep models.

However, in practice, a problem is observed: when dominant patterns or high-frequency noise appear in the input context (long dialogues, noisy data, repetitive or dominant tokens), models become unstable and prone to generation degradation and hallucinations.

In this article, I attempted to find out if the choice of activation function could be fundamentally linked to LLM hallucinations.

In my previous article, I noted some interesting behavior regarding Weight Decay; here, I examine it in detail.

It is generally accepted in the ML industry that if we take a pre-trained model and fine-tune it on a new task, the old weights are gradually overwritten. Furthermore, if we add Weight Decay (L2 regularization), the process of "forgetting" superfluous information should theoretically happen even faster.

I tested this claim experimentally. The results were counter-intuitive: under specific settings, Weight Decay works in the exact opposite way—it protects the old structure from destruction.

Below is a description of the experiment and conclusions for those involved in model training and AI safety.

In January 2026, Ollama added support for the Anthropic Messages API, enabling Claude Code to connect directly to any Ollama model. This tutorial explains how to install Claude Code, pull and run local models using Ollama, and configure your environment for a seamless local coding experience.

In my previous article, I explored the phenomenon of subliminal learning, but it raised more questions than answers. It is time to dive deeper. Below, you will find the experiments and the code.

In the fields of AI Alignment and LLM Security, a critical question remains: does fine-tuning or Reinforcement Learning from Human Feedback (RLHF) guarantee the removal of unwanted information?

Spoiler: The experiments demonstrated that the well-known Mode Connectivity effect makes the complete erasure of pre-training information practically impossible during standard fine-tuning. Structural Imprinting persists in the weight topology and can be read through a subliminal channel. Even with full weight unfreezing and aggressive L2 regularization (active forgetting), the latent space topology formed during the pre-training stage persists and determines the solution to the new task with an accuracy of 88–99%.