Normal Computing, the startup building full-stack probabilistic compute infrastructure enabling artificial intelligence (AI) for the most critical and complex applications, announced today that it has raised $8.5M in a Seed funding round led by Celesta Capital and First Spark Ventures, with participation from Micron Ventures.
The funding from the investment round will “advance Normal Computing‘s commitment to helping large companies use technologies like Generative Artificial Intelligence in intricate and high-stakes real-world contexts.”
It will also “support the research and development of Normal Computing’s application development platform and Probabilistic AI technology.”
Despite reliability issues like unpredictable factual errors or “hallucinations”, large general-purpose models like OpenAI’s GPT-4 continue to fascinate the globe.
While these limitations are acceptable for early consumer applications, according to Faris Sbahi, the CEO and Co-Founder at Normal Computing, “they pose key challenges for advancing core enterprise workflows where AI’s transformative value creation potential has yet to be unlocked.”
Normal Computing’s Probabilistic AI is “a paradigm that may solve these and other roadblocks by giving unprecedented control over reliability, adaptivity, and auditability to AI models powered by its customers’ private data.”
Forged through their work in the largest-scale and the most critical AI workflows at Alphabet, Normal is “supporting use-cases where risk has been a central barrier to AI adoption.”
These systems encompass “a wide range of applications.”
They reportedly “include automating complex underwriting processes, where policies may involve numerous locations with specific guidelines.”
Additionally, they can “enable autonomous workflows for generating and validating specialized code that adheres to mission-critical constraints and unique idioms for custom and confidential codebases.”
Moreover, they can assist in “mitigating risks in an airline’s supply network, even in dynamic and ever-changing conditions.”
In response to a question like “What recommendations would you provide for my client thinking to save for their kid’s college?,” a typical Large Language Model (LLM) deployed to assist a financial advisor “by synthesizing across various data portals and policies might make up or provide out-of-date or impersonal details that are critically relevant to decision-making.”
As well, it may fail to “provide transparent reasoning that would be needed for audit.”
In contrast, with Probabilistic AI, models can “detect when they synthesize inaccurately by also generating probable, auditable explanations of how they reached a conclusion, and even revise themselves by adaptively making an additional query to a datastore or human-in-the-loop.”
Nicholas Brathwaite, Founding Managing Partner at Celesta Capital, said:
“Artificial Intelligence has the potential to address some of the greatest human challenges of our time. But in order to do so, it must be reliable, transparent, and able to comprehend the limits of its own reasoning so that it knows how best to engage and explain to humans in the loop. We are excited to support the Normal Computing team as they develop their cutting-edge Probabilistic AI, which will help to develop AI that can be trusted for use in critical public and private applications.”
Probabilistic AI can “enhance promising models like LLMs and Diffusion Models, as well as enable new architectures.”
Normal says that “integrating these large models into composed workflows with their Probabilistic AI technology – in addition to specialized models, enterprise-specific plugins, and domain-specific processes – has the ability to solve complex real-world problems.”
Normal’s technology is designed “to deploy these large AI systems reliably, detecting and fixing failures like hallucinations and predictably adapting and learning in real-time to private data and changing conditions.”
Faris also “explained Normal’s commitment to working collaboratively with its clients to enable applications that routinely involve multiple stakeholders, a complex data landscape, and sophisticated security policies. ”
Amongst major AI innovations – “like scaling transformer models with GPUs – there often remains a significant gap between these new capabilities and the requirements for real-world production use cases where information is incomplete and noisy, and constantly changing,” said Faris.”
Furthermore, successful resolutions “are typically rich and limited to the largest tech companies like Alphabet and Meta.”
Faris emphasized:
“AI has the potential to improve essentially everything we value, but we’ve seen a trend of doubling down on certain architectures and approaches because they work with today’s held conventional tools and infrastructure, not because they are as trustworthy or understandable as we can achieve. The solution is to redesign AI systems from the ground up,” said Faris. “This contrasts other more surface-level approaches like prompt engineering and retrieval-based methods which alone aren’t enough, especially for mission-critical enterprise problems. The Normal team is thrilled to have the support of our investors in this Seed round of funding to confront this challenge head-on and continue to enable and advance principled systems for our partners.”
Normal asserts “that AI system transparency and openness are frequently required for adoption.”
This means that Normal “provides AI systems backed by customizable open source models, similar to Stanford University’s Alpaca, which allow for full auditability.”
This further contrasts closed systems “like OpenAI’s GPT-series whose internals remain hidden.”
Normal’s system is “designed so that a company’s proprietary information remains private, with no uncertainty how its data is being used.”
This is one way in which these systems “can more auditably uphold a business’s ground truth. Normal itself is committed to being an active contributor to the open source, having made available some of its developer tools for reliable Generative AI workflows.”