Neural Networks Enthusiast

Danil Kutny

I love science, AI research, and startups. Intelligence is the last great mystery in humanity’s most significant pillars of knowledge — and we are on the verge of solving it.

“What I cannot create, I do not understand.” — Richard P. Feynman

Telegram · @DanilKutny danil.kutny@gmail.com +7 926 618 55 55

Contact me any time. GPUs are appreciated 🥰

Research projects

Neuroevolution

Neural networks that can evolve like living organisms. This mechanism of evolution is inspired by real-world biology and is heavily focused on biochemistry. Much like real living organisms, these neural networks consist of cells, each with their own genome and proteins. Proteins can express and repress genes, manipulate their own genetic code and other proteins, regulate neural-network connections, facilitate gene splicing, and manage the flow of proteins between cells — all of which contribute to a complex gene-regulatory network and an indirect encoding mechanism for neural networks that enables advanced evolutionary processes.

View on GitHub →

A backpropagation alternative

A novel approach to learning that replaces traditional backpropagation with neural-network-based learning rules. Typically, backpropagation computes gradients by matrix multiplication between the error at each layer and the corresponding weights, then propagates those errors back through the network. In contrast, this method concatenates key information — backpropagated error, weights, inputs, pre-activations, and activations — and uses a separate neural network to predict gradients from it. This allows the learning rule’s own weights to evolve, enabling the discovery of optimization strategies potentially superior to conventional gradient descent. Preliminary results suggest a flexible, scalable alternative that can adapt to complex learning environments.

Recurrent RL agent outperforming the S&P 500 thesis

This thesis investigates using neural networks to simulate stock-market behavior and make trading decisions. Using 20 years of data — S&P 500 prices, news articles, and macroeconomic indicators — the network makes daily buy/sell decisions for individual stocks. The model outperformed the S&P 500 in 2019, achieving better overall results, demonstrating the potential of neural networks for improving trading strategies in financial markets.

View on GitHub →

Other explorations

Neural loss-function evolution

Local learning rule

Multithreaded GPT

An experimental extension of GPT models that explores parallel processing in transformers by applying a unique set of weights to each token in a sequence.

View on GitHub →

AI startups

Wordify

Wordify offers a highly intuitive interface with seamless access to a variety of AI models, focused on user experience and efficiency. Whether for research, development, or everyday use, it simplifies the process of interacting with powerful language models — streamlining innovation for all users.

wordify.ru →

Sensei Solutions

Simplifies the creation of large texts — theses, books, and extensive writing projects. With an intuitive interface, users can easily draft and structure their content, streamlining long-form writing. Ideal for academics, authors, and content creators who need a straightforward tool to manage their writing efficiently.

sensei-solutions.ru →

Telegram Swarm

A network of Telegram channels powered by a large language model. The system creates unique content across a wide range of topics — economics, art, travel, and many more. Using a sophisticated tree-based prompt design, the LLM generates engaging, diverse content, ensuring a continuous flow of fresh material for years to come.

📜 History 🕰️ 🍳 Gourmet 🥗 💪 Motivation ✨

Check the whole folder pack →

Writing

Philosophy · Consciousness

The Boundaries of Cognitive Closure: An Argument for Mysterianism in the Philosophy of Consciousness

This paper introduces a new argument for mysterianism, drawing on insights from artificial neural networks. Using a simple multilayer neural network trained to classify images, it shows that even such information processing can lie beyond our cognitive capabilities. This raises questions about the feasibility of understanding consciousness, suggesting that our cognitive limitations extend to the fundamental principles of interpreting complex systems.

Download the paper (PDF)

Get in touch

I live in Moscow and cooperate with people and companies all over the world. Contact me for cooperation.

Telegram · @DanilKutny WhatsApp danil.kutny@gmail.com +7 926 618 55 55