We invested in Sooth Labs' seed round alongside Felicis. Yann LeCun (Meta's Chief AI Scientist) and Jeff Dean (Chief Scientist at Google DeepMind) backed the round as angels. Andrew Bosworth, Meta's CTO, is advising.
The Team
World-class teams are the foundation of world-class companies, and one of the primary reasons we invested in Sooth. Rarely do we encounter a founding team where each person brings world-class credentials in their own right, spanning both the frontier of AI research and the hard work of shipping systems at scale.
Yaser Sheikh founded and scaled Meta's Pittsburgh AI lab from pure R&D into shipped products over the span of a decade. Codec Avatars is the flagship output of this work. He is a CMU Robotics Institute faculty member and the author of OpenPose and Panoptic Studio, which have become foundational infrastructure for AI's understanding of human motion.
I've known Russ Salakhutdinov for years. Russ completed his PhD under Geoffrey Hinton, who is commonly known as the “Godfather of AI”. He has amassed over 256,000 citations (among the most-cited active ML researchers in the world), was Apple's first Director of AI Research, and most recently served as VP of Research in Generative AI at Meta. His work is foundational and lifelong, not derivative.
Chuck Hoover directed billion-dollar innovation programs at Meta. He has navigated the full arc of taking a large-scale technical bet from internal conviction to organizational commitment to shipped product inside one of the most resource-intensive, politically complex companies in the world. The gap between research results and enterprise products is where most companies fail, and he has closed that gap before.
The Problem
Every institution makes high-stakes commitments under fundamental uncertainty every day. The tools they rely on are gut calls, proprietary internal models, and occasional expert opinion. Specialist forecasting platforms exist, but each is siloed to a single domain, and almost none are measured against what actually happens.
The biggest nonlinearities in any domain come from outside that domain. A geopolitical shift reprices insurance exposure. A rate environment collapses real estate assumptions. A technological development reshapes defense procurement before the procurement cycle can catch up. Single-domain forecasting, by design, misses exactly the signals that matter most.
Foresight should be a system, not a skill.
Why LLMs Don't Fix This
LLMs are trained to produce plausible text. They are not trained to produce calibrated probabilities graded against real outcomes. Next-token prediction is not a proper probabilistic scoring. Plugging an LLM into a forecasting workflow doesn't change what it optimizes for.
Simulation-based approaches (Aaru, Simile) take a different path: build a bottom-up model of individual behavior and aggregate up to predictions. The problem is tail risk. Tail events are precisely where bottom-up granularity fails, because the required accuracy at the individual level is unachievable at the exact moments when it matters most. And tail events are where the value is. A forecasting system that works in normal conditions but fails in crisis conditions isn't a forecasting system.
What Sooth Does
Sooth's core is a continuously trained world model for long-horizon, cross-domain forecasting. The architecture has three components: a multimodal encoder that ingests unstructured text and structured time series simultaneously; a forecaster that optimizes in latent world-state space rather than direct output accuracy; and foresight engines that surface the model through a chat interface for queries, counterfactuals, and scenario planning.
This is what The Moat Just Moved looks like in forecasting. The advantage isn't domain expertise bolted onto an analyst interface. It's the architecture underneath: a world model trained across domains, optimized for calibration, and continuously updated as events unfold.
Why Now
The team spent decades building the foundational infrastructure that makes this possible. Panoptic Studio, OpenPose, Codec Avatars, Apple AI Research, Meta's generative AI stack at scale. The compute and data availability that make a continuously trained world model viable didn't exist five years ago. The demand from institutions that are running out of road with existing tools is real and growing.
We invested because Russ, Yaser, Chuck, and their team are the right people to build this. And we think the time is now.
Follow Sooth Labs at soothlabs.com.
