Faris Sbahi is the Co-Founder and CEO of Normal Computing, a Celesta portfolio company developing enterprise AI systems that can reason and make decisions alongside humans, even when there is uncertainty and risk — empowering computers and humans to accomplish complex tasks together that require weighing tradeoffs and dealing with real stakes.
Faris Sbahi is the Co-Founder and CEO of Normal Computing, a Celesta portfolio company developing enterprise AI systems that can reason and make decisions alongside humans, even when there is uncertainty and risk — empowering computers and humans to accomplish complex tasks together that require weighing tradeoffs and dealing with real stakes.
While there is a lot of excitement around enterprise AI adoption right now, we still have huge issues related to things like hallucinations, lacking the ability to understand AI’s underlying logic and reasoning. Addressing this is incredibly important when it comes to complex, high-stakes workflows and we need to solve this if we expect AI to truly scale towards fundamental possibility and efficiency limits. This is where Normal has set its sights.
Our current focus is in the world of semiconductors. Right now, there's a gap from both a product and technology perspective in terms of solutions that can improve how we test, service, and design complex hardware like semiconductors and industrial machinery. From a user and product perspective, there are mission-critical requirements around reliability and experience.
Because of reliability issues, AI has struggled in many ways to penetrate high-stakes, critical, resource intensive projects. Our AI systems now understand the advanced formal mathematics and physics behind hardware, and help unlock complex AI automation and increase the velocity of this work, without the costly mistakes. There are parallels with DeepMind’s AlphaGeometry, but for hardware instead of pure mathematics, and led by former LLM leads from Meta, Palantir, and Google Brain – alongside silicon experts from Graphcore, Los Alamos, and more.
We believe the hardware sector faces some of the most important problems in the world. Geopolitical and environmental challenges are compounding and there's a lot of disruption and unreliability happening within our supply chains and critical manufacturing. Increasing capability for the semiconductor and manufacturing sector can have a positive impact on a whole range of other sectors like healthcare or sustainability. The industry struggles to tackle wide-ranging topics like the “AI energy crisis” because of what I call the “silicon complexity crisis”. Even with the simplest kinds of physical architectures, like memory, complexity is now at the PhD level. We need AI to help design and manufacture our chips.
One of the fundamental issues we can immediately address is the talent shortage issue. Most companies we partner with are 10-20x short of qualified test engineers that can read specs, write tests, service equipment, and so on. It's an order-of-magnitude gap in terms of our capacity to test and service these complex hardware products. We believe the only way to make up for that gap is a technology-driven approach. We are trying to add a 10-100x multiple on top of the existing headcount of these organizations – humans aided by digital Robots.
Quite exciting as well – we were recently selected for the UK’s Advanced Research and Invention Agency (ARIA) program. We will use our technology to de-risk unconventional physics-based architectures which are thousands of times more energy efficient for critical AI primitives. This involves our trademark “thermodynamic computing” approach which harnesses noise as a resource rather than fighting against it, a pioneering design space to tackle in partnership with ARIA.
Many companies that have come up in the wake of the ChatGPT storm are already gone a year or so later. We were founded and VC-backed before ChatGPT was released, and built some of the early pioneer Physics + ML tools over the last 5 years – like with Antonio and Tensorflow Quantum. So, we have that DNA in the company and deeply understand the landscape and how it’s evolving, which informs our approach to differentiation.
One element is our domain focus. We don’t believe there will be one generalist LLM to rule them all, at least for some time. As AI is gaining more adoption, everyone is finding it necessary to have a purpose-built view. Every industry has its own considerations and requires different kinds of products and technology, going all the way down to the models.
This is why we’re initially laser-focused on semiconductors, electronics, and industrial equipment with a strategy of becoming critical infrastructure for these kinds of companies by delivering a complementary AI workforce that is 10x multiple on the critical engineering headcount that they need. With our current partners, we regularly outperform on accuracy, quality, and calibration versus the generalist models because we are purpose-built and focused on specific domains.
Another element is our vision to expand into a full-stack AI company – as with the aforementioned ARIA UK program. The probabilistic AI we are building translates into transparent AI reasoning and greater user confidence – we believe this is the future of AI. Advanced models like these run too inefficiently on today’s hardware and are much less energy efficient than is possible. Developing our own hardware infrastructure technology to support Normal’s proprietary software will enable us to scale our AI to a broader market more quickly, strengthening our cost advantage while enhancing the quality we already deliver.
AI has struggled in many ways to penetrate high-stakes, critical, resource intensive projects, because of reliability issues. Our AI systems now understand the advanced formal mathematics and physics behind hardware, and help unlock complex AI automation and increase the velocity of this work, without the costly mistakes.
Founders are looking for someone who can be a trusted advisor, but every company’s needs are different. At our current stage, all the various introductions have been invaluable. These connections help us navigate the underserved market we operate in and connect us with key decision-makers to form essential partnerships.
At the same time, the best investors know when to trust the founding team, as it can be challenging to invest without the same level of insight you’d have in an operator role. It's a delicate balance. I think it’s incumbent on the startups to guide this – and the great ones do it really well – which is to figure out how best to utilize your investor base and ask for help when it will be most beneficial. This starts with bringing the right investors into the fold. We are tremendously grateful for the support from Steve Fu, Nic Brathwaite, and the entire Celesta Team.
You have to make it fun, even the challenges. There's a funny quote from Nvidia CEO Jensen Huang, where he wished great “pain and suffering” upon every Stanford graduate, in the sense that you learn so much and build character by pushing through challenges. I believe what has made this journey so rewarding for us is having a tight-knit founding team that truly enjoys tackling the challenges that come our way.
In doing so, it's crucial to foster a team culture where everyone feels comfortable with the unglamorous aspects. Instead, you're relishing bringing those issues center stage and figuring them out together. This is the whole point of a startup, you're trying to build something that lasts by working to solve deep, fundamental problems that no one has before. Don't try to battle the hard parts on your own, make it a team sport. We tend to work with a lot of ex-founders, in many cases recently exited, for this reason, and – I like to think – future founders as well.
In many ways, tackling challenging, highly ambitious problems is actually easier than working on something more conventional and routine. At Normal, this is how we’ve been able to attract an incredibly brilliant and passionate team.
Haha, yes! There are two meanings.
One is in reference to the normal distribution, which is a probability theory representing the most common distribution of naturally occurring results. As we're using probabilistic machine learning models, we wanted our name to be an ode to the technology we're developing.
The second aspect was more about the culture we’re aspiring to create. We want to build a real company, one that's no-nonsense. With all the buzz in AI, and tech startups in general, we wanted to disconnect ourselves from that and develop meaningful technology, build high-impact products, and create genuine value. Normal Computing is a nod to this approach as sort of “back to basics” or the “normal” way we think all new startups should be thinking, all the way down to the physics and fundamental limits of our Universe.