Sauraj Gambhir
Co-Founder at Prior Labs
29 Jul 2026Anthropic Responds | SSI’s $5B | Prior Labs Close €1B SAP Acquisition | Human Fridges
Prior Labs' Siraj on the €1B SAP Acquisition
Following a Hex ad read, the hosts recap the deal — €700M upfront plus €300M in stock, with SAP committing €1B to scale the lab — before Siraj explains the structure was inspired by DeepMind's acquisition by Google, preserving independence, brand, and speed. He details how SAP will distribute TabPFN across its product suite to a third of its customers with its ~25,000-person sales team, the new moonshots research arm, and rising attention on tabular data including Google's new model and Nvidia buying Kumo. He also covers Yann LeCun's role on the scientific advisory board, the move to a co-CEO structure, and open go-to-market roles.
22 Nov 2025Intercom, BackedVC $100M, Beside $30M, TabPFN-2.5, Alex Karp, Molly O’Shea, Bryan Johnson
Saraj: Prior Labs' TabPFN 2.5 & Tabular Foundation Models
Saraj, co-founder of Prior Labs, recounts his path from finance (RBC tech M&A and venture) into startups, and how Prior Labs formed around co-founders Frank Hutter and Noah and their research on tabular foundation models. He explains why LLMs struggle with tabular data — scarce quality public datasets and sequential architectures — and how Prior Labs instead trains transformers on synthetic tabular data. TabPFN 2.5 now tops the Tab Arena benchmark, handling up to 100,000 data points and beating methods trained for 4+ hours in seconds, with 2.5M downloads and an open-source push to replace hundreds of per-use-case models with one; the segment ends as he describes expanding tooling for data scientists' workflows.
TabPFN Founder on Tab Arena Benchmarks & Missing Data
The guest explains how their tabular foundation model is benchmarked: a public, academic-maintained leaderboard called Tab Arena, where they are now the leading model, plus an internal benchmark on curated datasets. The tasks are classification and regression across many datasets. He also highlights the model's ability to impute missing values with high accuracy, something traditional ML methods can't do without a data scientist doing manual work.
TabPFN Go-to-Market: Licenses, Banks, Hitachi & Healthcare
The guest describes the ~20-person team, an inbound-only go-to-market, and commercialization via an API plus commercial model licenses, following the foundation-model approach with Robin Rombach as an angel. Customers include banks, a healthcare diagnostics company building earlier lung-disease detection, and Hitachi for railway track anomaly detection. He defends staying horizontal, citing daily user feedback calls revealing use cases from materials discovery to energy trading.