Six-month-old US start-up launches new ‘energy based’ reasoning model as it targets $1bn-plus valuation
Yann LeCun, Meta’s former chief AI scientist, will be chair of Logical Intelligence’s technical research board © Fabrice Coffrini/AFP/Getty Images
Logical Intelligence, a start-up that claims to have made a breakthrough in developing a more advanced AI model, has appointed Yann LeCun to its board as it seeks funding at a $1bn-plus valuation.
The six-month-old Silicon Valley-based firm on Wednesday unveiled a new “energy based” reasoning model called Kona, which it says can solve a variety of problems with greater accuracy and less power than popular large language models such as OpenAI’s GPT-5 and Google’s Gemini.
Eve Bodnia, Logical Intelligence’s founder and a quantum physicist, told the FT that the high accuracy of its mathematically-grounded system made it suitable for industries where errors are critical, such as advanced manufacturing, robotics and energy infrastructure.
The launch comes as Logical Intelligence prepares to kick off a funding round in the coming weeks, targeting a valuation of between $1bn and $2bn, according to two people familiar with the matter.
It has also appointed LeCun, Meta’s former chief AI scientist who left late last year to set up his own AI start-up , as chair of its technical research board.
The push comes as some of the world’s top AI groups including Google DeepMind and Nvidia have stepped up their focus on alternative systems such as world models to achieve machine “superintelligence” amid concerns that LLMs are reaching a ceiling in their progress.
Logical Intelligence is the first start-up to attempt to commercialise an energy based model. In a demo on Wednesday, the company showed Kona beating rival LLMs from OpenAI, Google and Anthropic to solve a sudoku, and is planning future demonstrations of the games chess and go.
The company claims that Kona’s ‘reasoning’ is based on the model’s ability to recognise and correct its errors on one task and extrapolate that across other tasks.
“If general intelligence means the ability to reason across domains, learn from error, and improve without being retrained for each task, then we are seeing in Kona the first credible signs of AGI,” said Bodnia. “It is not the end state, but it is a clear break from narrow AI.”
In contrast to LLMs, which respond to queries by predicting the likely next word in a sequence, energy based models are trained on a set of fixed parameters, such as the rules of sudoku, the capacity of an energy grid or the physical constraints on a warehouse robot.
The more EBMs seek to deviate from the rules, the more energy they use, preventing digressions — or hallucinations — that can occur with LLMs. EBMs grade their complete answer against the parameters they have been set, with the lowest energy answers scoring highest.
Bodnia describes LLMs as “a guessing game”. “We think that because it sometimes gives you the right answer, it must be intelligent. If you take a cat and you teach it to bark, that doesn’t mean it’s a dog,” she said. “I’m creating the dog.”
LeCun, who this month launched Advanced Machine Intelligence Labs, a start-up that will focus on developing world models, has long been an outspoken critic of the idea that LLMs alone could achieve AGI. World models aim to navigate the physical world by learning from videos and robotic data rather than just language.
“Logical Intelligence is the first company to move EBM-based reasoning from a research concept to products, enabling a new breed of more reliable AI systems,” said LeCun.
He and Bodnia both believe that true human-level AI will come from combining models.
“[Humans] can dance, we can sing, we can do a lot of things that are not just language based. So you need all of it to navigate the world: you need to recognise the danger of the world, you need to maximise plausible outcomes,” said Bodnia.
本文转自FINANCIAL TIMES（https://www.ft.com），仅作知识交流使用，不涉及商业用途。若存在侵权问题，可随时联系我们删除。
知更小筑