Yubei Chen is co-founder of Aizip inc., an organization that builds the world’s smallest and best AI models. He can be an assistant professor within the ECE Department at University of California, Davis. Chen’s research is on the intersection of computational neuroscience and deep unsupervised (self-supervised) learning, enhancing our understanding of the computational principles governing unsupervised representation learning in each brains and machines, and reshaping our insights into natural signal statistics.
Prior to joining UC Davis, Chen did his postdoc study with Prof. Yann LeCun at NYU Center for Data Science (CDS) and Meta Fundamental AI Research (FAIR). He accomplished his Ph.D. at Redwood Center for Theoretical Neuroscience and Berkeley AI Research (BAIR), UC Berkeley, advised by Prof. Bruno Olshausen.
Aizip develops ultra-efficient AI solutions optimized for edge devices, offering compact models for vision, audio, time-series, language, and sensor fusion applications. Its products enable tasks like face and object recognition, keyword spotting, ECG/EEG evaluation, and on-device chatbots, all powered by TinyML. Through its AI nanofactory platform, Aizipline, the corporate accelerates model development using foundation and generative models to push toward full AI design automation. Aizip’s Gizmo series of small language models (300M–2B parameters) supports a wide selection of devices, bringing intelligent capabilities to the sting.
You probably did your postdoc with Yann LeCun at NYU and Meta FAIR. How did working with him and your research at UC Berkeley shape your approach to constructing real-world AI solutions?
At Berkeley, my work was deeply rooted in scientific inquiry and mathematical rigor. My PhD research, which combined electrical engineering, computer science, and computational neuroscience, focused on understanding AI systems from a “white-box” perspective, or developing methods to disclose the underlying structures of information and learning models. I worked on constructing interpretable, high-performance AI models and visualization techniques that helped open up black-box AI systems.
At Meta FAIR, the main target was on engineering AI systems to realize state-of-the-art performance at scale. With access to world-class computational resources, I explored the boundaries of self-supervised learning and contributed to what we now call “world models” — AI systems that learn from data and picture possible environments. This dual experience — scientific understanding at Berkeley and engineering-driven scaling at Meta — has given me a comprehensive perspective on AI development. It highlighted the importance that each theoretical insight and practical implementation have once you’re developing AI solutions for real-world applications
Your work combines computational neuroscience with AI. How do insights from neuroscience influence the way in which you develop AI models?
In computational neuroscience, we study how the brain processes information by measuring its responses to numerous stimuli, very similar to how we probe AI models to grasp their internal mechanisms. Early in my profession, I developed visualization techniques to research word embeddings — breaking down words like “apple” into their constituent semantic elements, reminiscent of “fruit” and “technology.” In a while, this approach expanded to more complex AI models like transformers and enormous language models which helped reveal how they process and store knowledge.
These methods actually parallel techniques in neuroscience, reminiscent of using electrodes or fMRI to review brain activity. Probing an AI model’s internal representations allows us to grasp its reasoning strategies and detect emergent properties, like concept neurons that activate for specific ideas (reminiscent of the Golden Gate Bridge feature Anthropic found when mapping Claude). This line of research is now widely adopted within the industry since it’s proven to enable each interpretability and practical interventions, removing biases from models. So neuroscience-inspired approaches essentially help us make AI more explainable, trustworthy, and efficient.
What inspired you to co-found Aizip? Are you able to share the journey from concept to company launch?
As a fundamental AI researcher, much of my work was theoretical, but I desired to bridge the gap between research and real-world applications. I co-founded Aizip to bring cutting-edge AI innovations into practical use, particularly in resource-constrained environments. As a substitute of constructing large foundation models, we focused on developing the world’s smallest and best AI models which can be optimized for edge devices.
The journey mainly began with a key remark: While AI advancements were rapidly scaling up, real-world applications often required lightweight and highly efficient models. We then saw a possibility to pioneer a brand new direction that balanced scientific rigor with practical deployment. By leveraging insights from self-supervised learning and compact model architectures, Aizip has been in a position to deliver AI solutions that operate efficiently at the sting and open up latest possibilities for AI in embedded systems, IoT, and beyond.
Aizip makes a speciality of small AI models for edge devices. What gap available in the market did you see that led to this focus?
The AI industry has largely focused on scaling models up, but real-world applications often demand the alternative — high efficiency, low power consumption, and minimal latency. Many AI models today are too computationally expensive for deployment on small, embedded devices. We saw a spot available in the market for AI solutions that would deliver strong performance while operating inside extreme resource constraints.
We recognized that it shouldn’t be only unnecessary for each AI application to run on massive models, but that it also wouldn’t be scalable to depend on models of that size for every part either. As a substitute, we concentrate on optimizing algorithms to realize maximum efficiency while maintaining accuracy. By designing AI models tailored for edge applications — whether in smart sensors, wearables, or industrial automation — we enable AI to run in places where traditional models can be impractical. Our approach makes AI more accessible, scalable, and energy-efficient, unlocking latest possibilities for AI-driven innovation beyond the cloud.
Aizip has been on the forefront of developing Small Language Models (SLMs). How do you see SLMs competing or complementing larger models like GPT-4?
SLMs and bigger models like GPT-4 aren’t necessarily in direct competition because they serve different needs. Larger models are powerful when it comes to generalization and deep reasoning but require substantial computational resources. SLMs are designed for efficiency and deployment on low-power edge devices. They complement large models by enabling AI capabilities in real-world applications where compute power, latency, and value constraints matter — reminiscent of in IoT devices, wearables, and industrial automation. As AI adoption grows, we see a hybrid approach emerging, where large, cloud-based models handle complex queries while SLMs provide real-time, localized intelligence at the sting.
What are the most important technical challenges in making AI models efficient enough for low-power edge devices?
One in all the basic challenges is the dearth of an entire theoretical understanding of how AI models work. With no clear theoretical foundation, optimization efforts are sometimes empirical, limiting efficiency gains. Moreover, human learning happens in diverse ways in which current machine learning paradigms don’t fully capture, making it difficult to design models that mimic human efficiency.
From an engineering perspective, pushing AI to work inside extreme constraints requires revolutionary solutions in model compression, quantization, and architecture design. One other challenge is creating AI models that may adapt to quite a lot of devices and environments while maintaining robustness. As AI increasingly interacts with the physical world through IoT and sensors, the necessity for natural and efficient interfaces — reminiscent of voice, gesture, and other non-traditional inputs — becomes critical. AI at the sting is about redefining how users interact with the digital world seamlessly.
Are you able to share some details about Aizip’s work with firms like Softbank?
We recently partnered with SoftBank on an aquaculture project that earned a CES Innovation Award — one we’re especially pleased with. We developed an efficient, edge-based AI model for a fish counting application that could be utilized by aquaculture operators for fish farms. This solution addresses a critical challenge in fish farming which may ultimately create sustainability, food waste, and profitability issues. The industry has been slow to adopt AI as an answer because of unreliable power and connectivity at sea, making cloud-based AI solutions impractical.
To unravel this, we developed an answer based on-device. We combined SoftBank’s computer graphics simulations for training data with our compact AI models and created a highly accurate system that runs on smartphones. In underwater field tests, it achieved a 95% recognition rate, dramatically improving fish counting accuracy. This allowed farmers to optimize storage conditions, determine whether fish must be transported live or frozen, and detect potential diseases or other health issues within the fish.
That breakthrough improves efficiency, lowers costs, and reduces reliance on manual labor. More broadly, it shows how AI could make a tangible impact on real-world problems.
Aizip has introduced an “AI Nanofactory” concept. Could you explain what meaning and the way it automates AI model development?
The AI Nanofactory is our internal AI Design Automation pipeline, inspired by Electronic Design Automation (EDA) in semiconductor manufacturing. Early development in any emerging technology field involves loads of manual effort, so automation becomes key to accelerating progress and scaling solutions as the sphere matures.
As a substitute of simply using AI to speed up other industries, we asked, can AI speed up its own development? The AI Nanofactory automates every stage of AI model development from data processing to architecture design, model selection, training, quantization, deployment, and debugging. By leveraging AI to optimize itself, we’ve been in a position to reduce the event time for brand new models by a mean factor of 10. In some cases, by over 1,000 times. This implies a model that when took over a 12 months to develop can now be created in only just a few hours.
One other profit is that this automation also ensures that AI solutions are economically viable for a wide selection of applications, making real-world AI deployment more accessible and scalable.
How do you see the role of edge AI evolving in the subsequent five years?
Edge AI guarantees to rework how we interact with technology, much like how smartphones revolutionized web access. Most AI applications today are cloud-based, but that is beginning to shift as AI moves closer to the sensors and devices that interact with the physical world. This shift emphasizes a critical need for efficient, real-time processing at the sting.
In the subsequent five years we expect edge AI to enable more natural human-computer interactions, reminiscent of voice and gesture recognition and other intuitive interfaces, which might remove reliance on traditional barriers like keyboards and touchscreens. AI can be expected to change into more embedded in on a regular basis environments like smart homes or industrial automation to enable real-time decision-making with minimal latency.
One other key trend can be the increasing autonomy of edge AI systems. AI models will change into more self-optimizing and adaptive due to advancements in AI Nanofactory-style automation, so they may have the option to cut back the necessity for human intervention in deployment and maintenance. That can open latest opportunities across a variety of industries like healthcare, automotive, and agriculture.
What are some upcoming AI-powered devices from Aizip that you just’re most enthusiastic about?
We’re working to expand use cases for our models in latest industries, and one we’re especially enthusiastic about is an AI Agent for the automotive sector. There’s growing momentum, particularly amongst Chinese automakers, to develop voice assistants powered by language models that feel more like ChatGPT contained in the cabin. The challenge is that almost all current assistants still depend on the cloud, especially for natural, flexible dialogue. Only basic command-and-control tasks (like “activate the AC” or “open the trunk”) typically run locally on the vehicle, and the rigid nature of those commands can change into a distraction for drivers in the event that they do not need them memorized with total accuracy.
We’ve developed a series of ultra-efficient, SLM-powered AI agents called Gizmo which might be currently utilized in a variety of applications for various industries, and we’re working to deploy them as in-cabin “co-pilots” for vehicles too. Gizmo is trained to grasp intent in a more nuanced way, and when serving as a vehicle’s AI Agent, could execute commands through conversational, freeform language. For instance, the agent could adjust the cabin’s temperature if a driver simply said, “I’m cold,” or reply to a prompt like, “I’m driving to Boston tomorrow, what should I wear?” by checking the weather and offering a suggestion.
Because they run locally and don’t rely upon the cloud, these agents proceed functioning in dead zones or areas with poor connectivity, like tunnels, mountains, or rural roads. In addition they enhance safety by giving drivers complete voice-based control without taking their attention off the road. And, on a separate and lighter note, I assumed I’d also mention that we’re also currently within the strategy of putting an AI-powered karaoke model for vehicles and bluetooth speakers into production, which runs locally just like the co-pilot. Mainly, it takes any input audio and removes human voices from it, which lets you create a karaoke version of any song in real-time. So except for helping customers more safely manage controls within the automobile, we’re also searching for ways to make the experience more fun.
These sorts of solutions, those that make a meaningful difference in people’s on a regular basis lives, are those we’re most pleased with.
Aizip develops ultra-efficient AI solutions optimized for edge devices, offering compact models for vision, audio, time-series, language, and sensor fusion applications. Its products enable tasks like face and object recognition, keyword spotting, ECG/EEG evaluation, and on-device chatbots, all powered by TinyML. Through its AI nanofactory platform, Aizipline, the corporate accelerates model development using foundation and generative models to push toward full AI design automation. Aizip’s Gizmo series of small language models (300M–2B parameters) supports a wide selection of devices, bringing intelligent capabilities to the sting.
