In Conversation with Moshe Tanach CEO and Co-Founder of NeuReality
With the unprecedented investments that are pouring into artificial intelligence, countries and companies are racing to harness the transformative potential of technology. Amid this rapid advancement, the challenges of balancing innovation with security and regulation become increasingly crucial.
AIWire had the opportunity to interview Moshe Tanach, the CEO and Co-Founder of NeuReality. We discussed the critical role of both government and private sector in advancing AI technology and security.
NeuReality is a venture capital-backed AI compute and network infrastructure startup with a mission to revolutionize AI inferencing to reduce cost and power consumption while unleashing the full potential of GPUs and all AI accelerators worldwide. The startup offers a unique AI architecture that can deliver 90 percent cost savings and a 10-15x boost in energy efficiency compared with today’s host CPU and NIC-centric architectures.
Commenting on the US-China geopolitical implications of the recently announced $32 billion annual funding for AI research and development, Tanach stressed the importance of the U.S. taking the lead in innovation and supporting the AI industry, highlighting that resource allocation is a vital component of this strategy.
Tanach stated that China is advancing rapidly in computer vision technology but with concerns about data privacy due to government access to information. In contrast, the U.S. emphasizes data protection, which may lead to stricter AI regulations. Without more investment and regulatory clarity in the U.S., particularly in AI innovation and adoption, the country risks remaining #2 and falling further behind China.
We asked Tanach about his views on how the European Union (EU) is handling the AI revolution. Tanach said that while Europe leads in areas like sustainability, green technology, and regulation, this focus may hinder its growth in AI innovation and may not allow it to be at the forefront of cutting-edge technology. Tanach emphasized that Europe is a key market for AI.
“If you want to expand and grow globally, you need to sell in Europe,” Tanach said. “Europe still holds significant commercial power. It's definitely a region that NeuReality is investing in, and they in us – referring to the EU’s capital investment in the AI start-up earlier this year. There are many customers across sectors like telecom, enterprise, government, healthcare, and finance.”
The need to not only sell but partner with major hardware and software companies in Europe could potentially have a ripple effect, requiring companies to comply with European regulations and standards.
Tanach thinks that trying to prevent sales of more GPUs to China or vice versa is not the way to go, as there is “always a bypass”. We are already seeing China starting to develop more hardware capabilities to bypass such restrictions.
“You can't stop technology,” Tanach said. “Instead, you need to learn how to work with it, protect it, and, ideally, reduce regulation. More alignment between nations is necessary, and it's crucial to safeguard dual-purpose technology from falling into the wrong hands, such as terrorist organizations or other potential misuses.”
Tanach is passionate about building security deep into the hardware and infrastructure layer of AI technology, explaining that the deeper the security protocols, the harder they are to obstruct. We asked him how AI systems can be designed with security embedded from the ground up to mitigate risks like digital fakes, cyber threats, and election interference.
According to Tanach, there are several strategies that can be employed to achieve this. It starts with data protection between servers. NeuReality is improving data protection and operational efficiency by redesigning server architecture and utilizing advanced hardware. Instead of combining GPUs with CPUs, NeuReality is directly integrating AI accelerators with its NR1 AI Inference Solution – a silicon-to-software system that boosts the output of any AI accelerator, while improving cost and energy efficiency by orders of magnitude.
While that addresses one layer of security, Tanach shared that NeuReality also addresses the challenges of deploying AI models at scale. The company invented a hardware-based AI Hypervisor layer to manage different processing stages. Instead of relying on CPU-hosted software, which is vulnerable to errors and hacks, NeuReality takes a hardware-based approach to fortify its platforms. Additionally, NeuReality’s approach is designed to protect each client’s work in isolation, adding multiple layers of security to the AI processing pipeline.
Tanach believes that as cloud solutions often share resources among multiple customers, there are concerns about data privacy and security. Large AI models, which are trained on diverse data sources, can pose risks if not adequately protected.
“I think hackers will eventually find ways to exploit prompt engineering,” Tanach said. “For instance, they might develop a long sequence prompt to access data that was intended to remain hidden or not to connect to other data. These are critical issues that warrant further research.”
Tanach continued: “As we address this new type of technology, which is complex and continually evolving like an onion with many layers, it is crucial for countries to invest in protecting multi-modal and large language models. I would like to see the US, China, and the EU allocate funds specifically for research into safeguarding these models.”
Regarding the involvement of private sector companies in shaping global AI governance frameworks, Tanach emphasized that both hardware and software infrastructure companies must ensure their products meet essential standards and security measures.
“We’re shifting the paradigm of CPU-centric architecture to NeuReality architecture,” Tanach said. “We are already working with Qualcomm and Cirrascale with their accelerators and cloud solutions, respectively, and we are taking it to market, bringing four to eleven times improvement on the cost per token or the cost of frame processing, in any modality—audio, video, text recommendation engines.”