In this exclusive VMblog Q&A, we sit down with TJ Dunham, the founder and CEO of ARC, a deep tech company revolutionizing AI with its groundbreaking Reactor AI.

Focused on sustainability and efficiency, Reactor AI sets itself apart from traditional large language models (LLMs) by drastically reducing energy consumption and GPU requirements. With rapid ontological classification (ROC) at its core, Reactor is changing the landscape of AI development, offering a smarter and more sustainable alternative.

In this interview, Dunham shares insights into the innovations driving Reactor AI and the broader implications for the future of AI technology.

VMblog:  What is ARC and what is ARC's Reactor AI? What does the company do?

TJ Dunham:  We are ARC, a deep tech company dedicated to developing a new generation of super-efficient AI. We started ARC in 2023 with the belief that AI should work in the service of benefiting humanity, while being simple and transparent enough to be accessible to as many people as possible.

We designed Reactor Mk as a purpose-built large language model, to use significantly less energy and resources than the LLMs deployed by OpenAI, Google, Anthropic, and others. We developed a far better and more sustainable way to train AI than any other LLM currently available. With Reactor AI, we've developed a novel, highly performant approach to training AI models.

VMblog:  How does Reactor AI achieve its energy efficiency compared to traditional LLMs? Can you elaborate on the role of rapid ontological classification (ROC) in this process?

Dunham:  Reactor achieves its energy efficiency by taking a fundamentally different approach to model training and data management. Traditional LLMs are built on vast, unstructured data sets that models need to sift through. In contrast, Reactor focuses on concise, highly organized data, from our rapid ontological classification (ROC) system.

While most other models train on massive, unfiltered datasets, Reactor uses data parsed and organized by ROC technology. Our ROC system incorporates elements from open-source models, discarding irrelevant or outdated information as it trains. It's a streamlined approach that allows us to train much more efficiently while consuming far fewer resources and less energy.

ROC technology helps structure the model's data more effectively. Imagine a traditional model with 75 billion parameters trying to self-organize its information; it's like navigating a maze. By contrast, Reactor AI's ROC system organizes data into clear "highways," allowing the model to access relevant information quickly and efficiently. This not only reduces energy consumption but also boosts performance.

VMblog:  What are the environmental implications of Reactor AI's reduced GPU requirements and energy consumption? How might this impact AI industry sustainability efforts?

Dunham:  The environmental implications are enormous. Typically, AI companies and their models are consuming vast amounts of energy, focusing on building bigger, more powerful models at unsustainable rates. Reactor demonstrates a better way: we can build highly efficient models without burning through exorbitant amounts of energy and resources, including water.

Our approach-using just a few cloud GPUs and achieving superior performance-sets a new standard for sustainability in AI. We are a small team outperforming energy-intensive giants on a fraction of the resources. Our whole focus is toward sustainable AI which not only reduces the industry's carbon footprint but also make AI more accessible and responsible.

VMblog:  Can you explain how Reactor's architecture differs from conventional LLMs and why this leads to improved energy efficiency?

Dunham:  Reactor's architecture combines different open-source models and leverages our ROC system, creating a streamlined, highly efficient model. Traditional LLMs must navigate through vast, disorganized datasets for every query, consuming more energy and time. Reactor, however, avoids this by organizing data through an AST-like system, making it much easier to access relevant information fast.

To put it in perspective, in essence Reactor has the most direct highways to its data centers, allowing for quicker, more efficient responses. This simplified "route" to information means Reactor requires significantly less energy, resulting in vastly improved efficiency.

VMblog:  How does ARC's approach to training Reactor with just 8 NVIDIA L4 and 4 NVIDIA A100 GPUs compare to methods used by larger tech companies? What are the implications for AI accessibility and democratization?

Dunham:  ARC's approach changes the game for AI accessibility and democratization. While large companies like OpenAI and xAI use thousands of GPUs to train their models, consuming massive amounts of energy, we achieved Reactor's performance using just 8 NVIDIA L4s and 4 A100 GPUs.

It shows that you don't need huge data centers or immense power to train high-performing models. Our method, focused on efficiency and innovation, shows that smaller companies can compete at the highest level without excessive resources. This paves the way for more startups and smaller players to enter the field, fostering a more competitive, diverse, and sustainable AI landscape.

VMblog:  In what ways does Reactor's ontological classification method provide advantages over traditional LLMs that rely on vast training datasets?

Dunham:  Reactor's rapid ontological classification (ROC) offers several key advantages over traditional LLMs that rely on large, disorganized datasets. First, it allows for much more efficient data organization, meaning the model can retrieve relevant information faster. Second, this organization makes the data more accessible, so the model doesn't waste energy parsing through irrelevant information.

Additionally, our models can collaborate more efficiently through agentic interaction, where multiple models work together seamlessly. This is in stark contrast to the typical scaling methods of other companies, which rely on more GPUs to increase power. Reactor's architecture allows us to scale more efficiently, enabling models to be smaller, faster, and more sustainable.

VMblog:  How do you envision Reactor's energy-efficient design influencing the future development of AI technologies, particularly in addressing the growing concerns about AI's resource consumption?

Dunham:  Reactor AI sets a new standard for energy efficiency. This will influence the broader AI industry. Our goal is to create a model that absorbs more carbon than it produces-a "climate-positive" AI. This flips the current narrative that AI is inherently harmful to the environment.

As more companies realize the potential of our approach, they'll be motivated to reduce their energy consumption and adopt sustainable practices.

Instead of competing for the largest, most energy-hungry models, we envision a future where efficiency and sustainability are the key drivers of AI innovation. Ultimately, this shift will result in AI technologies that are not just more powerful but also more responsible.

VMblog:  Can you share some specific data or metrics that demonstrate Reactor's efficiency gains compared to other LLMs in the market?

Dunham:  Here's one example in terms of energy consumption. Our Reactor AI used less than 1 megawatt of energy for training, while other models have consumed upwards of 50,000 megawatts. This difference-50,000 times more efficient-is staggering and represents a leap forward in AI efficiency.

In terms of the speed difference vs. traditional models, this is easily illustrated by the resources it took to train our model. With Reactor Mk, we only used 8 L4 GPUs and 4 A100s, running for less than a day, while GPT-4 is so massive it is believed to have required over 25,000 A100s running for three months for training.

You can also experience Reactor's efficiency for yourself by taking it for a spin. Go to https://reactor.helloarc.ai and give it a whirl.

VMblog:  While the iOS and Android apps you're just announcing are exciting, the focus seems to be on Reactor's efficiency. How do you see this technology potentially reshaping the landscape of mobile AI applications?

Dunham:  Reactor's efficiency will play a pivotal role in reshaping mobile AI. Our goal is to build a full-time assistant that can run on your phone, helping with tasks like drafting emails, organizing work, and even functioning offline. As the technology evolves, we want users to own their assistants fully, with all data encrypted and stored locally on their devices. This would ensure complete privacy and control over the AI.

The implications are profound. Reactor will allow AI to be more deeply integrated into our daily lives without draining device resources. It's not just about having AI on your phone; it's about having an AI that's efficient, private, and truly yours-without compromising on functionality or speed.

##

TJ Dunham, Founder and CEO ARC

TJ Dunham is the Founder and CEO of ARC, a cutting-edge startup at the intersection of AI and blockchain technology. A seasoned entrepreneur, TJ previously led DePo, a multi-market aggregator that achieved significant success with over 50,000 users and an exit valuation of $45m. At ARC, TJ has spearheaded the development of Reactor, an AI model that has claimed the top spot on the MMLU benchmark while using a fraction of the energy typically required for such advanced systems. This achievement underscores TJ's vision and commitment to sustainable innovation in AI. With a proven track record in both the AI and blockchain sectors, TJ continues to drive technological advancements that create value and push industry boundaries. His leadership at ARC reflects a vision for responsible, efficient, and groundbreaking tech solutions.

‍