Ruslan Salakhutdinov has founded a company called Sooth Labs after leaving Meta.
Salakhutdinov is no ordinary "former Meta employee." He served as Meta's chief AI scientist, previously a professor at Carnegie Mellon University, and is a heavyweight in deep learning. Co-founders Chuck Hoover and Yaser Sheikh are also former Meta research leaders.
The company announced on April 22 that it had completed a $50 million funding round at a $335 million valuation, led by Felicis Ventures.
Two investors' names stand out: Yann LeCun (the "godfather of deep learning" and Turing Award winner) and Jeff Dean (founder of Google Brain and current Google Chief Scientist). Former Meta CTO Andrew Bosworth serves as an advisor.
Having these three names on a single company's cap table is rare in the AI industry. What are they betting on?
What Sooth Labs Does
In a nutshell: It helps enterprises predict the probability of specific geopolitical and market events occurring.
This is not a conversational assistant, a coding tool, or a RAG application. Sooth Labs aims to be a "prediction engine" — providing AI-driven probability assessments that tell businesses how likely an event is to happen.
Specific scenarios could include: the probability of a policy shift in a certain country, regulatory risks in a particular market, or geopolitical conflict risks in a supply chain. These judgments previously relied on the experience and intuition of human analysts; now the goal is to use AI for quantitative estimation.
This direction has its historical context.
Why Now
Prediction markets have grown rapidly in recent years with the help of AI. Platforms like Polymarket and Metaculus have already demonstrated that aggregating the beliefs of market participants can generate more accurate probability estimates than traditional forecasting.
Sooth Labs wants to go a step further: not just aggregating human beliefs, but training AI models to learn the patterns of events directly from massive datasets.
This problem presents several technical challenges:
- Limited sample size for geopolitical events (large-scale wars are rare, historical data is scarce)
- Poor comparability between similar events (each crisis has a different context)
- Difficulty separating signal from noise (media coverage does not represent true probability)
Whether they can succeed remains an open question. However, the three founders did extensive work on structured data and human feedback training at Meta, experience that is transferable to predictive modeling.
What LeCun and Jeff Dean's Bet Means
LeCun recently founded AMI Labs, focused on "world models" — enabling AI to truly understand the three-dimensional physical world, not just process text.
His investment in Sooth Labs suggests he believes that AI's ability to understand "events" is also a crucial component of world models. Geopolitical events are among the most complex patterns in the human world; accurately predicting them implies a deep understanding of human societal dynamics.
Jeff Dean has led extensive work on prediction and time-series ML at Google. His bet on Sooth Labs reads more like an endorsement that "this direction is technically feasible."
Both men, from different research perspectives, point to the same conclusion: the path of using AI to predict complex events is worth pursuing.
Who Will Pay
A $335 million valuation and $50 million in funding — not huge by today's AI fundraising environment.
But the potential customers for this sector are very clear: risk management departments at multinational corporations, hedge funds, insurance actuarial teams, and government agencies. These clients are willing to pay and have the budget.
Compared to consumer-facing AI for everyone, the B2B vertical prediction AI sector has much less competition, though its ceiling is also relatively clear.
The core proposition Sooth Labs needs to validate is: Are AI-generated probability estimates more accurate, faster, and cheaper than teams of human analysts?
If they can prove this in a few specific areas, the business can take off. If not, $50 million is enough for a few years of experimentation.
The endorsements from LeCun and Jeff Dean help open doors for data procurement and institutional client negotiations — which may be more valuable than the money itself.
Whether this sector can take off also depends on competition with prediction markets and policy think tanks. But starting with top-tier AI researchers' backgrounds gives them at least a high starting point.
Sources: CocoLoop, AI Pioneers Back Startup Building Models to Predict Events (Bloomberg, April 22, 2026)