GPT-7 Will Be Smarter Than We Are

The last time we talked to Bert Van Hoof, president and chief operating officer of Willow, Bert told us in December that WillowTwin was linking with Microsoft’s Sustainability Cloud to develop an AI-driven intelligent rules engine designed to integrate building systems and replicate the knowledge of aging operators.

Since our last talk, Microsoft has invested $10B in OpenAI, which unleashed GPT-4 on the world and soon will unveil a more powerful—by a factor of at least seven—GPT-5, a chatbot which can synthesize voices as well as images and language.

Bert couldn’t tell us everything he knew in December, and he knew a lot—the Willow COO spent nearly a decade working in Microsoft’s AI skunkworks, which has now grown to more than 8,000 AI specialists.

Now, he can tell us that the rapidly evolving ChatGPT language model has become the foundation of an exponentially growing number of “co-pilot” applications layers, including WillowTwin.

“Microsoft has a process called grounding. All of that is based on underlying graph technology—grounding proves the accuracy of a prong. The co-pilot improves the prong and that goes into the large data modeling for ChatGPT,” Van Hoof explained.

As an example of how quickly GPT technology is being adapted, Van Hoof cited a co-pilot Microsoft launched 12 months ago that focused on understanding and generating code, aimed at app developers.

“Now, 40% of all the code written on the planet is done by co-pilot. That’s a stunning number, but it takes away a lot of the endless copy-and-paste repetition that humans were doing,” Van Hoof said. “AI just does it better and creates productivity for developers.”

“The big debate right now is which kinds of activities do these domains not necessarily replace the humans, but in a field like X-ray analysis, AI can detect more things than the human could,” he said.

How quickly do the machines get rid of the lawyers? “GPT-3 was in the 10^th  percentile of passing the bar exam, GPT-4 is in the 90^th percentile,” Van Hoof noted. “It’s an area a lot more conducive to AI.”

An AI co-pilot is a natural fit for WillowTwin, Van Hoof told GlobeSt.

“If you think of WillowTwin, we have the graph for the physical world,” he said. “We can define all this data we have, the spatial, the static, the live data; we can overlay that with skills and code it from experts—then we can overlay that with the large language models.”

“Day-to-day operators can get more concise instructions on what to fix, in which order and why. The human can decide whether this response is appropriate or not,” Van Hoof said. “The AI intelligence can help inform the human to better understand the science of a big impact score and why any action is warranted.”

“Ultimately, generative AI [becomes] the feedback loop that looks at the results of those actions and has the measured outcomes that improve the human encoded skills based on the response of the system and the result of it,” he said.

Willow is focused on developing an AI-driven platform that will serve an entire building life cycle. Earlier this year, the company announced a partnership with Johnson Controls that integrates WillowTwin with JCI’s OpenBlue platform.

“They have a huge installed base and they’re a highly relevant player in the market from the equipment side. We’re tapping into all of their network platforms from acquisitions all the way through their services arm, including financing,” Van Hoof said.

Will we see a fully automated building operations platform soon?

“I don’t think so. Every time we introduce an opportunity for autonomous command and control, we give the operator the opportunity to approve the action first,” he said. “You’re going to see very repetitive and safe things where the operator goes ‘I’ve seen this in action, I now trust it, I don’t need to prove this every time and we’ll let the system run it.’”

The rapid advance of the chatbots surprised even insiders like Bert Van Hoof. To understand how rapidly the brain power of AI “language models” is growing, we’ll use as a benchmark the human brain—which has an estimated 86B neurons firing off 60 trillion synapses to run the cognitive universe inside our noggins.

Substitute the word “parameters” for synapses—this is what the AI technicians call the correlations that language models build as they are trained—and here is the timeline for AI:

A Microsoft translator program in 2018 (that was named “Bert” in honor of you know who) had 340M parameters. According to Van Hoof, GPT-3 had 75B parameters, while GPT-4 logged in at 540B parameters.

This means GPT-5—a bot that will be able to generate deep-fake videos with perfect audio mimicry, just in time for next year’s election—will likely have a “brain” that is firing nearly 4 trillion synapses.

Even people who have trouble multiplying using a calculator can figure out this means GPT-7—and a multitude of competing bots—will be smarter than we are, coming a lot sooner than anyone expected—which is why Elon Musk has publicly asked for an immediate six-month pause in AI development.

Van Hoof echoed what OpenAI CEO Sam Altman said in a recent interview, referring to the generative machine learning that is driving AI: “The genie is out of the bottle.”

“The advance from GPT-3 to GPT-4 went very rapidly,” Van Hoof said.

Are the machines ready to train each other? “No, definitely not yet, but there is potential there that it could get there more quickly than people anticipated,” the Willow COO said.

“Even the AI experts—what we thought would take 20 years is now happening in five years. The last six months have really accelerated things,” he said.

“The genie is out of the bottle,” Van Hoof said. “It’s going to be an unstoppable train.”

Last week, President Biden indicated that the federal government is contemplating guardrails for AI development. The president said the verdict is out on whether rapid AI development will take a dangerous turn. “It could,” Biden said, on the same day Italy banned ChatGPT.

Whenever GPT-7 shows up, it’s going to have a big head start on understanding what makes the human race tick: its Beta version, out and about as GPT-4, already has been accessed by more than 200M of us in less than five months. All of the questions 200M people and counting can think of asking are now part of the GPT training module.

“Every time there are breakthroughs, there are positives and [there is also] somebody that gets affected in a negative way,” Van Hoof noted. “When people live longer, that’s bad for funeral homes.”