Why this founder is building ‘alternative AI’

Eve Bodnia came to AI by way of a quest to understand real intelligence. The quantum physicist by training said she’s long been fascinated by how the human brain functions as a hyper-efficient computer. During her time as a PhD student at UC Santa Barbara, she said her work bridged the mathematics, physics, and neuroscience departments.

Perhaps that unconventional path helps explain why she’s taking a different approach to AI than much of the rest of Silicon Valley. Instead of chasing ever-larger LLMs, her company, Logical Intelligence, is betting on an alternative structure called energy-based models (EBMs), which it claims can better perform certain heavy reasoning tasks with much less compute.

Rather than predicting the next words in a sequence, energy-based models score different possible answers to a problem based on how well they fit the data and underlying constraints. Logical Intelligence shows off its own model, Kona 1.0, in a demo where it solves a Sudoku puzzle faster than leading LLMs.

“We’re not just exclusively saying, like, ‘LLM is bullshit.’ Logical Intelligence is not saying the LLM is dead. But we’re saying that the LLM is not enough,” Bodnia said. “It’s not enough for winning the enterprise market, especially when it comes to the mission-critical market.”

It’s critical

By “mission-critical,” she means industries like finance, power grids, and medical robotics that demand precision and involved reasoning capabilities. Those are the types of use cases where Logical Intelligence thinks its models can outperform reasoning LLMs in terms of precision. When she first met with Morning Brew in Manhattan in March, Bodnia was in town pitching financial firms on Logical Intelligence as well as looking to raise money from investors.

Right now, Logical Intelligence is looking to establish a set of key open-ended partnerships in each of the industries it’s targeting. Bodnia wants Logical Intelligence to develop custom models in collaboration with engineers at each of the client firms to solve a problem of their choosing.

So Logical Intelligence might work with a grid operator to build a model around real-time analysis of data on power distribution. Bodnia noted the company is working with a partner on material science for LED screen applications.

“We’re going to pick exclusive partners from each vertical,” Bodnia said. “We’re going to give them an advantage against their competitors, but in exchange, they’re going to help us build the product for this specific vertical.”

And Logical Intelligence’s coding verification agent, Aleph, offers machine-proofed secure code aimed at “operators of critical infrastructure and safety-sensitive systems,” per the company’s website.

“I feel like we actually nailed two of the most challenging fields in history together in one company,” Bodnia said. “So one is the EBM concept. Another one is the formal verification concept. So formal verification is just a mathematical technique which allows you to guarantee the correctness of your systems. And this guarantee comes as a machine-verifiable proof.”

Alt-AI

Bodnia’s line of thinking on LLMs echoes the views of prominent LLM critic and AI pioneer Yann LeCun, who recently joined Logical Intelligence’s research board. LeCun, who previously served as Meta’s chief AI scientist, has long argued that LLMs alone are not the path to so-called artificial general intelligence (AGI). Since leaving Meta earlier this year, he’s become more vocal in his criticism of what he characterizes as the herd mentality in the “LLM-pilled” tech industry, as he put it in a recent New York Times interview.

“He was advocating for alternative architectures for a very long time, both in academic and industry spaces,” Bodnia said, adding that Logical Intelligence is more of a complement to than competitor of LeCun’s own startup, Advanced Machine Intelligence Labs, which is focused on world models. “We are part of this new alternative AI ecosystem, but all of us also can work with LLMs.”

“Researcher at heart”

Bodnia said she didn’t go to grad school intending to join the startup world, at least not before a longer career in academia. But she said the trade-offs with research freedom and worries that her published research would get lost in a flood of other AI papers made her decide to take the leap early.

She started with the idea of energy-based models, and the B2B structure followed naturally from there.

“I’m a researcher at heart, so I just generate ideas for the sake of ideas. And sometimes they’re useful, sometimes they’re not,” Bodnia said. “And when you start a business, you have to think about, who’s gonna buy it? Where is the pressure on the market? And it seems like [there is] this problem of LLMs still not doing trading, LLMs not doing your taxes. You can’t rely [on an LLM] to drive a car; robotics tasks are not automated by LLM. And I’m like, this architecture seems to be exactly for those kinds of tasks.”