Vladimir Baranov

I teach machines
to sleep

Researching biologically-inspired sleep cycles in large language models. A mechanism that lets models consolidate new knowledge directly into weights, the way biological memory separates encoding from consolidation. I come to this from 12 years of building production systems, which gives me a practitioner's instinct for what learning architectures need to do in the real world.

Research & Novel Systems

Sleep-Cycle Continuous Learning in LLMs

2025 – present

An LLM architecture that implements biologically-inspired sleep cycles for continuous learning. The system separates experience accumulation (wake) from weight consolidation (sleep), allowing the model to learn from direct user interaction without catastrophic forgetting.

Drawing on memory consolidation literature, including complementary learning systems theory and hippocampal replay. The sleep phase replays and distills new experiences into stable long-term representations. End-to-end implementation: data pipeline, training loop modifications, evaluation harness.

Exploring applications in personalization, domain adaptation, and lifelong learning for deployed models.

Continuous Learning Memory Consolidation CLS Theory Weight Adaptation

Articles

No Pain, No Gain: A Propositional Architecture for Perception-First Embodied Intelligence
April 2026
The Silicon Split
March 2026
The Bilateral Distraction: Why Left/Right Brain Thinking Is an Evolutionary Accident, Not a Blueprint
March 2026
Your Laptop Can Run a Mind, But Never a Superintelligence
March 2026
There's Only One Way to Sleep
March 2026
The Ghost and The Animal
February 2026
Local Maxima: The Most Important AI Product Doesn't Exist Yet
February 2026

Andre's AI Cube Challenge

A game vibe coded by my 7-year-old son, and then he built an AI that learns to beat it

Andre built a platformer with 11 obstacle rooms: spikes, crushers, lasers, wind zones, and more. Then he trained a neural network to play it using neuroevolution. The AI starts from zero, dies constantly, and gradually figures out each room. Play the game yourself, then watch the AI learn below.

Play the Game
Arrow keys / WASD to move, Space to jump. Press the button in each room, then exit right.
[フレーム]
Watch the AI Learn
A neural network teaches itself to beat each room through trial and error. Use speed controls to fast-forward.
[フレーム]

Research

Per-Fact Graduated Consolidation Resolves the Capacity Ceiling in Weight-Edited Language Models
Zenodo, 2026
Sleep-Wake Memory Convergence in Weight-Edited Language Models
Zenodo, 2026
Sleeping LLM: Two-Phase Memory Consolidation for Lifelong Learning from 3B to 70B Parameters
Zenodo, 2026
Dual-System Memory Consolidation for Lifelong Learning in Language Models: Combining Direct Weight Editing with Sleep-Wake Training
Zenodo, 2026
The Alignment Tax on Continual Learning: Inverse Scaling of Memory Consolidation in Language Models
Zenodo, 2026
Sleep-Wake Consolidation for Lifelong Conversational Memory in Local Language Models
Zenodo, 2026

Engineering & Systems Building

12 years of production software

Founded and ran Tipasoft, a development agency delivering web, mobile, and desktop applications across a dozen industries. Full-stack engineering: architected systems from scratch and integrated into complex existing codebases. Comfortable across frontend, backend, mobile, and infrastructure.

Select clients

Coca-Cola Morgan Stanley KPMG BMW Honda Toyota Mars Wrigley Rogers Sportsnet Concordia University
Python TypeScript React Node.js Mobile Desktop Systems Architecture

Cross-Domain Pattern Recognition

Why breadth matters for AI research

The pattern of entering unfamiliar domains, building accurate mental models quickly, and shipping. It's the same skill continuous learning systems need to acquire.

Music Production

Co-wrote, recorded, mixed, and mastered electronic music. Worked with Grammy-winning producers. Managed live performances at Canada's largest venue (Rogers Centre). Nominated for Toronto Independent Music Awards. Trained intuition for signal processing, temporal patterns, and iterative creative refinement.

Consumer Products & Education

Created a coffee brand and children's robotics kits. Teaches kids robotics. Experience designing physical and digital systems, and thinking about how non-experts learn.

Enterprise Across Industries

Clients across finance, automotive, CPG, media, and education. Each required rapidly building domain models and translating between technical and non-technical contexts.

Education

University of Waterloo York University

AltStyle によって変換されたページ (->オリジナル) /