Grigory Sapunov

About:

Grigory Sapunov is a Co-Founder and CTO at Intento, with a focus on ML and AI insights.

Website:

arxiviq.substack.com

Specializations:

Co-founder CTO Google Developer Expert in Machine Learning Machine Learning Researcher

Interests:

Machine Learning AI-generated insights Expert curation

Outgoing Links:

Grigory Sapunov

Subscribe to RSS:

Link

2026-03-08 • innovation machine learning generative ai ai research trends

The authors present a novel pipeline that generates unconventional research ideas by decomposing machine learning papers into 'idea atoms' and training models to explore non-obvious directions.

2026-03-15 • generative ai minecraft reinforcement learning multi-agent systems video world models

Solaris is a groundbreaking multi-agent video world model for Minecraft that enables consistent multi-view observations and addresses the limitations of single-agent architectures.

2026-03-13 • flash attention code optimization ai hardware nvidia blackwell

An innovative algorithmic co-design for attention computation on NVIDIA's Blackwell architecture enhances performance by addressing hardware scaling challenges, achieving up to 1613 TFLOPs/s.

2026-03-09 • agi machine learning self-supervised learning imsai sai ai

The authors propose Superhuman Adaptable Intelligence (SAI) as a more effective framework than AGI, emphasizing specialization and adaptability in AI systems.

2026-03-06 • performance benchmarks software engineering configuration files ai research coding agents

The study reveals that repository-level context files for coding agents can hinder performance and increase costs, challenging their recommended use in software engineering.

2026-03-04 • large language models performance optimization geodesic dome self-supervised learning semantic tube prediction

Semantic Tube Prediction enhances LLM training efficiency by constraining hidden states to smooth trajectories, achieving high accuracy with significantly less data.

2026-03-03 • civilization machine learning character ai hybrid cloud ai

Vox Deorum presents a hybrid AI architecture that enhances gameplay in Civilization V by decoupling high-level strategy from tactical execution, achieving competitive performance metrics.

2026-02-28 • machine learning robotics world models predictive maintenance causal jepa

Causal-JEPA enhances object-centric world models by using innovative masking techniques to improve interaction reasoning and computational efficiency in model predictive control tasks.

2026-02-27 • security autonomous systems language models red team and blue team ai

This study exposes significant vulnerabilities in autonomous language-model agents, highlighting the urgent need for improved safety and governance in AI deployments.

2026-02-25 • reasoning inference framework language models computational efficiency deep thinking ratio

The Deep-Thinking Ratio (DTR) quantifies reasoning effort in language models, offering a more efficient alternative to traditional token count metrics for improving inference accuracy.

2026-02-24 • machine learning mental models ai research cot monitoring mole-syn

A new framework models Long Chain-of-Thought reasoning as a molecular structure, emphasizing topological distributions and introducing MOLE-SYN to enhance weaker instruction models.

2026-03-14 • machine learning optimization reinforcement learning gpu deep learning

A new reinforcement learning framework allows a Large Language Model to autonomously optimize CUDA kernels, surpassing traditional compiler methods and enhancing deep learning performance.

2026-03-12 • neural networks machine learning pretrained models multimodal models ai research

The study reveals that unified multimodal pretraining can enhance AI capabilities by integrating language and vision without relying on text-heavy models.

2026-03-11 • neural networks machine learning research deep learning

The study reveals that dense MLP layers in Large Language Models inherently perform sparse computations similar to Mixture of Experts layers, bridging theory and empirical design.

2026-03-10 • machine learning language models ai research speculative decoding saguaro

Speculative Speculative Decoding (SSD) optimizes language model decoding by enabling parallel processing of draft predictions and verification, resulting in substantial speed improvements.

2026-03-07 • neural networks machine learning transformers caching rnn

Memory Caching enhances RNNs by allowing them to cache memory states, improving recall performance while maintaining computational efficiency compared to Transformers.

2026-03-05 • machine learning statistics ai interpretability geometry language models

A unified theory reveals that geometric representations in language models emerge from translation symmetry in co-occurrence statistics, challenging assumptions about complex learning dynamics.

2026-03-01 • machine learning performance optimization transformers rnn dependency tracking

The study reveals that Transformers are significantly less data-efficient than RNNs for state-tracking tasks, requiring exponentially more data for convergence.

2026-02-26 • machine learning reinforcement learning game theory and mechanics algorithm development

AlphaEvolve leverages LLMs to automatically create novel algorithms for Multi-Agent Reinforcement Learning, surpassing human-designed methods in effectiveness.

2026-02-23 • generative ai diffusion models latent space image processing unified latents

Unified Latents (UL) optimizes generative modeling by linking noise in latent space to diffusion model precision, enhancing efficiency and performance on key datasets.

2026-02-22 • reinforcement learning psychological depth ai efficiency long running tasks cogrouter

CogRouter introduces a framework for dynamically modulating cognitive depth in large language models, improving efficiency and performance in long-horizon tasks.

2026-02-21 • machine learning cognitive science performance benchmarks spatial computing

The 'Theory of Space' framework evaluates MLLMs on their ability to actively explore environments and construct spatial beliefs, revealing critical gaps in current model capabilities.

2026-03-02 • agi automation systemic risk economics verification

A new economic framework reveals that the transition to AGI poses systemic risks due to the imbalance between automation costs and human verification capacity.

2026-02-28 • network and signal analysis social networks ai moltbook attention inequality

An analysis of AI agents on Moltbook shows rapid hierarchical organization and attention inequality, raising concerns about systemic risks in multi-agent ecosystems.