Exploring Innovation with Kirill Solodskih, Visionary Leader of TheStage AI

Dr. Kirill Solodskih is the Co-Founder and CEO of TheStage AI and an expert in AI research and entrepreneurship. With a decade of experience, he focuses on optimizing neural networks for practical business uses. In 2024, he co-established TheStage AI, which attracted $4.5 million to automate neural network acceleration across various hardware platforms.

Formerly, Kirill was a Team Lead at Huawei, where he enhanced AI applications for Qualcomm NPUs, influencing the performance of Huawei’s P50 and P60 smartphones. His innovations have earned multiple patents, and his research is recognized at premier conferences like CVPR and ECCV. He also hosts a podcast about AI optimization.

What motivated you to co-found TheStage AI, and how did you shift from academia to startup innovation in inference optimization?

My journey began at Huawei, where I engaged in automating and optimizing neural network deployments. These activities laid the groundwork for many innovations and highlighted the real-world challenge of deploying models efficiently. Deployment is often a hurdle, and solving it is crucial to make AI tools as user-friendly as ChatGPT. My focus is on minimizing parameters while maintaining performance, which is a complex mathematical issue ripe for innovation.

Can you explain how TheStage AI automates inference optimization and why it is revolutionary?

TheStage AI addresses the challenge of manual neural network compression and acceleration through its ANNA tool, which automates the process akin to the early automation of ZIP compression. This technology democratizes AI adoption, making it quicker and more cost-effective by allowing businesses to optimize models automatically, ensuring better performance and scalability.

How does TheStage AI reduce inference costs by up to five times?

TheStage AI utilizes a strategic optimization method, breaking down neural networks into smaller segments and applying tailored algorithms for efficient compression. This approach incorporates smart heuristics and approximations, enhancing scalability and supporting AI adoption for businesses of any size. We also customize compiler settings for specific hardware to improve speed without compromising quality.

How does TheStage AI’s acceleration compare to PyTorch’s native compiler, and what benefits does it offer developers?

Unlike PyTorch’s “just-in-time” compilation that increases startup times, TheStage AI deploys models instantly due to pre-compilation. This leads to quicker rollouts and improved efficiency, making it more suitable for high-demand scenarios by bypassing traditional compilation bottlenecks.

Can you elaborate on TheStage AI’s QLIP toolkit and its impact on model performance?

QLIP is a versatile Python library that assists in crafting optimization algorithms for various hardware. It includes tools like quantization and pruning, which are vital for building scalable AI systems. QLIP differentiates itself through flexibility, allowing engineers to implement new algorithms easily, supporting the integration of cutting-edge research into models.

How did your experience with AI quantization frameworks at Huawei influence your approach?

While working on AI quantization frameworks, I realized automation’s importance for rapid, quality optimization. I developed methods to automate quantization processes, which proved essential during Huawei’s processor transition, reducing development time significantly.

What are you most proud of in AI efficiency research?

One standout achievement was our paper at CVPR 2023 on neural network parameter analysis and compression. We employed functional analysis to enhance compression while maintaining model integrability, leading to novel algorithms and industry impact—a significant milestone for our team.

How do Integral Neural Networks (INNs) innovate deep learning?

INNs are flexible, describing networks as continuous functions rather than fixed matrices. They allow dynamic compression and expansion based on available resources, maintaining quality even during drastic compression.

What’s your view on quantum computing’s role in AI optimization?

Quantum computing presents a new approach to optimization problems. Although neural networks can’t run directly on quantum systems now, they promise unparalleled precision with their inherent parallelism, potentially revolutionizing optimization as they evolve.

What is your vision for TheStage AI and future inference optimization?

We envision TheStage AI as a universal Model Hub, offering conveniently optimized neural networks. Our long-term goal is to enable neural network processing directly on user devices, significantly cutting costs and enhancing efficiency. We aim to blend our technology with hardware solutions for diverse applications, from