Sooth Labs

Sooth Labs (also called Sooth) is an American artificial intelligence company based in Pittsburgh, Pennsylvania, that builds models to forecast the probability of geopolitical, economic and market events.^{[1][2][4][14]} It was founded in January 2026.^[3] Its chief executive is Yaser Sheikh, a former vice president at Meta and consulting professor at Carnegie Mellon University, and its chief scientific officer is Ruslan Salakhutdinov, a Carnegie Mellon machine learning professor.^[2][4] Sooth describes its system as a "world model for global foresight" that returns calibrated probabilities for questions about the future.^[4][5] Felicis, its lead investor, has said the underlying model is a time series model rather than a large language model.^[3] In April 2026, Bloomberg reported that the company was raising about $50 million in a round Felicis was set to lead, at a valuation of roughly $335 million including the money raised, with backing from Yann LeCun and Jeff Dean.^[1]

Overview

History

Background and founding

Sheikh was on the faculty of Carnegie Mellon's Robotics Institute from 2006 to 2019, where his group developed the OpenPose human pose estimation system and the Panoptic Studio.^[6] He was a vice president at Meta from 2015 to 2025. There he founded the Meta Reality Lab in Pittsburgh and led the development of Codec Avatars, which he describes as real-time, photorealistic telepresence.^[6] His Carnegie Mellon homepage describes his new venture as focused on "foundational advances in long-horizon foresight."^[6]

According to Felicis, Sheikh and Salakhutdinov founded Sooth in January 2026.^[3] Felicis partners Aydin Senkut and Feyza Haskaraman wrote that the founders came to their first meeting with a prototype that "came close to matching the accuracy of prediction markets," and that the firm "committed on the spot" to leading the seed round.^[3]

Sheikh named a larger founding group when he announced the company on LinkedIn on April 22, 2026. He wrote that after "a wonderful decade at Meta" he had teamed up with Salakhutdinov, Chuck Hoover, David LaRose and Shih-En Wei to found Sooth Labs, with Senkut and Haskaraman of Felicis as lead investors.^[5] The company's website lists the same five people under "Founders."^[4]

Public launch (April 2026)

Bloomberg News reported the company's funding on April 22, 2026, in an article by Rebecca Torrence. It described Sooth as "a new artificial intelligence lab founded by former Meta Platforms Inc. employees" that was raising about $50 million to build models that help businesses forecast "the likelihood of specific geopolitical and market events taking place."^[1] Sooth's co-founders told Bloomberg that the round, which Felicis was set to lead, would value the company at roughly $335 million including the money raised. Bloomberg also reported that LeCun, described as a former Meta executive, and Google chief scientist Jeff Dean had backed the company, and that Meta chief technology officer Andrew Bosworth was advising it.^[1] Felicis published its investment announcement the same day.^[3]

In a demonstration described by Bloomberg, Sooth's software let a user query the probability of specific events. It put the chance that the World Health Organization declares another pandemic by 2028 at 16%, and the chance that Anthropic goes public in 2026 at 33%.^[2] Sheikh told Bloomberg that the company wanted to help businesses make better decisions in areas such as capital allocation and risk management, and the company said it was in discussions with potential customers in finance, defense, insurance and real estate.^[2]

Later activity

As of October 2026, the company's LinkedIn page listed its headquarters as Pittsburgh, its founding year as 2026 and its size as 11-50 employees.^[14] In a Lab to Startup podcast episode dated September 15, 2026, Salakhutdinov, identified as Sooth's chief scientific officer, talked mainly about his Carnegie Mellon lab and former students; the episode notes list building Sooth Labs among the topics covered.^[13]

Founding team

Sheikh and Wei co-authored work on human pose estimation at Carnegie Mellon. Convolutional Pose Machines, by Wei, Varun Ramakrishna, Takeo Kanade and Sheikh, built a sequence of convolutional networks that refine estimates of body-part locations stage by stage, operating on the belief maps produced by earlier stages.^[11] The follow-up paper by Zhe Cao, Tomas Simon, Wei and Sheikh introduced Part Affinity Fields for real-time multi-person pose estimation and placed first in the inaugural COCO 2016 keypoints challenge.^[12] Both men are listed among the authors of the open-source OpenPose library, which is based on the Part Affinity Fields method.^[18]

Salakhutdinov joined Carnegie Mellon's Machine Learning Department in February 2016 after teaching at the University of Toronto.^[9] MIT Technology Review reported in October 2016 that Apple had hired him as its first director of AI research while he continued to work part time at Carnegie Mellon.^[10] His homepage lists his research interests as deep learning, probabilistic graphical models and large-scale optimization.^[8] Bloomberg described him as moving from Apple to conduct AI research at Meta, and Felicis gave his Meta title as vice president of generative AI research.^[2][3]

Technology

Approach

Felicis described the product as "a world model for long-horizon, cross-domain forecasting." A user poses a time-bounded, verifiable proposition, and the system returns a calibrated probability along with a structured, causal timeline of the events driving that estimate. The user can then change an assumption in that causal chain and see how the downstream probabilities shift. Felicis used the examples "Will U.S. CPI exceed 4% year-over-year at any point in 2026?" and "assume Iran signs a ceasefire."^[3]

Felicis also wrote that the model "isn't an LLM." It argued that language models optimize for next-token prediction rather than proper probabilistic scoring rules, and said Sooth instead "runs on a time series model."^[3] Bloomberg reported that Sooth trains its models on large cross-industry datasets, including data owned by its clients, and was working on models that take a mix of inputs, including video, audio and text.^[2] Sheikh described the system as one that integrates "discrete events and continuous signals into calibrated, auditable predictions."^[5]

The company's website argues that forecasting today is fragmented: prediction markets, time series models, financial markets and climate models each cover part of reality without capturing how the parts interact. It also states that large language models "don't maintain a persistent model of the world, continuously learn from outcomes, or produce calibrated probabilities that improve over time."^[4] Sooth uses the term world model for a system that represents how economic, geopolitical and operational events interact.^[4]

Product components

The website describes a "Global Foresight Engine" with three parts.^[4]

The company says its probabilities are calibrated: "when the system assigns a 70% probability to an event, approximately 70% of those events occur."^[4] As of October 2026, the website did not publish a forecast track record or benchmark results, and it asked interested organizations to request a briefing by email rather than offering self-serve access.^[4] Its homepage shows example questions labeled with the kinds of people who might ask them, such as an insurance chief risk officer asking about 2027 Atlantic hurricane losses and a defense planner asking about escalation in the Taiwan Strait.^[4]

Felicis said Sooth's initial market focus is financial services, defense and national resilience, because those are "the areas where the cost of being wrong is highest."^[3]

Funding

Andrew Bosworth is an advisor rather than a reported investor.^[1] The company has not published a final round size or closing date on its website.^[4]

Context

Bloomberg noted that Sooth's push coincided with surging interest in prediction markets such as Kalshi and Polymarket.^[2] Senkut told Bloomberg: "Because Kalshi and Polymarket have been so successful, the world has started thinking in bets."^[2]

Academic researchers have studied automated forecasting of world events with language models, an approach that Felicis and Sooth's website describe as poorly suited to the task.^[3][4] The Autocast benchmark (Zou et al., NeurIPS 2022) collected thousands of questions from forecasting tournaments with a date-organized news corpus. It found language models performed far below a human expert baseline, although performance improved with model size and with access to relevant news.^[16] Halawi et al. (2024) built a retrieval-augmented language model system that, on questions published after its models' knowledge cut-offs, came close to the crowd aggregate of competitive human forecasters on average and in some settings surpassed it.^[17]

Debarghya Das's "The Ultimate Neolab List" includes Sooth Labs as a frontier lab in a "Prediction" subcategory. In its "Neolab?" column the list answers "Mostly," with the note "Researcher-founded forecasting-model lab; small raise and unclear compute need."^[15]

See also

• List of neolabs
• Neolab
• World model
• Calibration (machine learning)
• Time series
• Kalshi
• OpenPose

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Rebecca Torrence, Bloomberg News. AI Pioneers Back Startup Building Models to Predict Events. Bloomberg Government, April 22, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Bloomberg, via Moneycontrol. AI pioneers back startup building models to predict events. Moneycontrol, April 22, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Aydin Senkut and Feyza Haskaraman. Foresight as Infrastructure: Our investment in Sooth. Felicis, April 22, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27Sooth Labs. Sooth Labs company website. Accessed October 27, 2026.
5. ^1 ^2 ^3 ^4Yaser Sheikh. LinkedIn post announcing Sooth Labs. LinkedIn, April 22, 2026.
6. ^1 ^2 ^3 ^4Yaser Sheikh. Yaser Sheikh (personal homepage). Carnegie Mellon University. Accessed October 27, 2026.
7. ^1 ^2Carnegie Mellon University Robotics Institute. Yaser Ajmal Sheikh. Accessed October 27, 2026.
8. ^1 ^2Ruslan Salakhutdinov. Ruslan Salakhutdinov (homepage). Carnegie Mellon University. Accessed October 27, 2026.
9. ^1 ^2Ruslan Salakhutdinov. Brief Biosketch. Carnegie Mellon University. Accessed October 27, 2026.
10. ^1 ^2Will Knight. Apple Gets Its First Director of AI. MIT Technology Review, October 17, 2016.
11. ^1 ^2Shih-En Wei, Varun Ramakrishna, Takeo Kanade and Yaser Sheikh. Convolutional Pose Machines. CVPR 2016. arXiv:1602.00134.
12. ^1 ^2 ^3Zhe Cao, Tomas Simon, Shih-En Wei and Yaser Sheikh. Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. CVPR 2017. arXiv:1611.08050.
13. ^Lab to Startup. The Professor Behind a Generation of AI Scientists and Founders. September 15, 2026.
14. ^1 ^2 ^3 ^4Sooth Labs. Sooth Labs company page. LinkedIn. Accessed October 27, 2026.
15. ^Debarghya Das. The Ultimate Neolab List. Accessed October 27, 2026.
16. ^Andy Zou, Tristan Xiao, Ryan Jia, Joe Kwon, Mantas Mazeika, Richard Li, Dawn Song, Jacob Steinhardt, Owain Evans and Dan Hendrycks. Forecasting Future World Events with Neural Networks. NeurIPS 2022. arXiv:2206.15474.
17. ^Danny Halawi, Fred Zhang, Chen Yueh-Han and Jacob Steinhardt. Approaching Human-Level Forecasting with Language Models. arXiv:2402.18563, February 2024.
18. ^1 ^2 ^3CMU Perceptual Computing Lab. OpenPose README. GitHub. Accessed October 27, 2026.