Artificial intelligence (AI) is now a broad term. Venture capital, Hollywood (think Schwarzenegger’s Terminator), and constant media use have blurred its meaning. It can refer to a chatbot writing an email, an algorithm sorting a social feed, a recommendation engine influencing purchases, or software that helps machines make decisions. Most people do not separate these uses. Everything is AI. Ben Wolff, President and CEO of Palladyne AI, sees this as an issue.
“Not all AI is the same. Those two little letters can legitimately mean a thousand different things,” Wolff tells the California Business Journal.
Wolff has spent enough time around emerging technologies to know how quickly language can race ahead of reality. Before his current work in embodied artificial intelligence at Palladyne, an AI robotics firm, he was involved in telco. An earlier company of his built a wearable humanoid robot—an exoskeleton funded by the U.S. military. It was meant to let a human lift up to 200 pounds without strain. Then, President Barack Obama publicly referred to the effort as Iron Man.
“The President was talking about my company, which created science fiction-like expectations at the Department of Defence, which was bad for public dialogue on AI,” Wolff says.
Beyond the unreal expectations, though, the exoskeleton project led to a stronger idea. If a person can teach a machine real-world tasks, maybe the machine can later perform those tasks on its own. For Wolff, the key was not to make machines better at conversation, like chatbots, but to teach them to work within limits.
Language to action
This difference is central to how he sees AI. Most people think of large language models (LLMs), which use huge amounts of computing power, big data centers, and lots of energy. These models are good at finding patterns and working with language. But they are less useful when a machine is off-grid, running on batteries, and needs to make quick decisions.
“What happens when you have a machine that is disconnected from the electrical grid, and it’s using batteries? How do you get those machines to be able to think when they have very limited processing power, very limited electricity?”
He calls his approach, which is the foundation at Palladyne, constrained embodied AI. This means intelligence that runs on the machine and is designed for action in the real world, not just analysis in the digital world. The goal is not to copy a data center at the edge, but to focus only on what matters.
Wolff’s team took inspiration from biology. He points to a statistic from their recent white paper: The human body collects about a billion bits of sensory data every second, yet conscious thought can process only about ten bits per second. Everything else is filtered out. That, he says, is the model of Palladyne.
Machines in factories, drones, and other physical settings do not need all the information in the world. They need only what is relevant to their task. A robot in a car plant does not need to consider healthcare or geopolitics. It needs to focus on its job. A drone in a combat zone faces different challenges. The intelligence must be specific and immediate.
“We have to figure out how to distil all of that information down to only those bits of information that actually matter to be able to operate in the physical world,” Wolff says.
He uses the phrase “last mile of intelligence”. In telco, the last mile is the final link between the network and the user. For Wolff, the data center is like the brain, but the machine at the edge is where intelligence is actually used.
“It’s only that stuff that happens on the edge, the device that is actually interacting in the physical world,” he says.
Constraint, not creativity
Consumer AI is different. Most people see AI as systems that respond to prompts and seem fluent. These systems are built to generate and improvise, which leads to errors and drift. Wolff uses this technology, but he is clear about its limits.
“Every time I get into a lengthy dialogue with ChatGPT, I’m asking it to come up with something creative. And when I think what I’m seeing is that, the more times I double down on the ask for creativity, that’s where I start seeing drift and hallucination,” he says.
Hallucinations occur in LLMs because they must always complete the prompt—even if they produce irrelevant or inaccurate information. Their output sounds like language, but may distort substance.
His systems are different. They are not designed for creativity. Instead, they focus on finding the most likely correct answer for a specific task.
“We’re asking the machine and the algorithms to come up with the most probable outcome to complete a mission and to do it right,” Wolff says.
This design choice shapes his view on responsibility. Many AI companies focus on what their systems can do. Wolff believes it is more important to ask whether the system improves human outcomes and whether its actions are sufficiently limited to be accountable.
In manufacturing, that means making robotics usable by ordinary workers rather than reserving it for engineers and PhDs. “If we can take robotics out of the hands of PhDs, and put it into the hands of people that might only have a GED [General Educational Development; a high school equivalency credential, not a diploma] education, we’re opening the aperture on who can participate in the new AI robotics economy,” Wolff says. His aim is to democratize AI-powered robotics.
Bubble and aftermath
This goal is most important for small businesses. Large manufacturers can afford to program robots for a single high-volume task. Smaller companies need robots that can handle several jobs, not just repeat the same task all day. Until recently, that was difficult.
“An industrial robot could only do one task. I refer to it as being single-threaded,” Wolff says.
He argues that software and embodied AI can change the economics. A flexible robot that can switch tasks like a person makes sense for businesses with many products and small production runs. For these companies, robotics is about helping workers, not replacing them.
“When a line worker has spent 15 or 20 minutes programming the robot for a new job, you know that’s the job it’s going to do. There isn’t the opportunity for drift.” Wolff says this confidence comes from focus, not general ability. The machine is a specialist, not a generalist.
Wolff does not believe that only edge-based systems will matter. He expects centralized LLMs models and local machine intelligence to work together. The central systems will handle training and analysis. The machines will handle real-world tasks.
His metaphor is American football. Coaches call plays from the sideline after reviewing film and preparing for scenarios all week. But once the ball is snapped, players on the field must improvise in real time because the defence rarely behaves exactly as planned.
“That is kind of like the large language models in the big data centers where they’re calling in the play. But then, invariably, things happen on the field that the offence wasn’t expected,” Wolff says.
He thinks the same will happen in industry. Central systems will make the plans. Edge intelligence will deal with problems and real-world changes.
When prompted by this columnist about the ongoing debate over whether today’s AI boom compares to the nineties’ dot-com boom—and then bust—Wolff sees similar hype and high investor expectations in today’s AI market. On this, he is direct. “I think we are in a boom, funded by greedy investors,” he says.
He believes the bubble is real, but the technology is not empty. Many companies will fail before the strong ones last. He has seen this before. The internet survived because it solved real problems. He thinks AI will do the same.
“AI is never going to go away. There is AI fatigue today just like there was internet fatigue back in the late nineties, early twenty hundreds. But it didn’t go away.”
What matters is not who makes the biggest claims, but who builds systems that work in the real world. Wolff prefers to tackle hard problems rather than seek easy consensus. He likes to prove people wrong when they say something cannot be done.
This may be why he sees AI as infrastructure rather than spectacle. He focuses on context, bandwidth, power use, training, safety, and task switching. In a time of big promises, his view is narrower, but likely more lasting.
AI could change industries, jobs, and the line between digital and physical work. But first, it needs a clear definition. Wolff’s Palladyne bets that the most important systems will not be the ones that talk best, but the ones that know what matters, stay within limits, and work in the real world.
Copyright © 2026 California Business Journal. All Rights Reserved.
For California Business Journal Disclaimers, go to /terms-conditions/.
