Uploaded October 2026 | Updated November 2026, 17 hours ago
Thanks to Neon for sponsoring this video. Check out Neon for your backend needs for your apps and agents: get.neon.com/4dTo3CI
Google Research's TabFM is a zero-shot foundation model for tabular data: you pass it the rows you already know, and it predicts the rest in a single forward pass, with no training. In this video I explain how TabFM learns from synthetic tables, why it can't read column names, and how TabFM-Auto adds an LLM agent around it. Then I build that agent on Neon Postgres to predict customer churn, and show how a leaky column fooled the agent and how a branch of the database's history fixed it.
Neon: get.neon.com/4dTo3CI
TabFM blog post: https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/
TabFM paper: arxiv.org/abs/2609.37959
TabFM-Auto paper: arxiv.org/abs/2609.37989
TabFM on GitHub: github.com/google-research/tabfm
TabFM on Hugging Face: huggingface.co/google/tabfm-1.1.0-pytorch
Timestamp:
00:00 - The Problem with LLMs on Tabular Data
01:52 - What is Google TabFM?
02:26 - How TabFM Works (Synthetic Causal DAGs)
03:17 - The Architecture: Database Backends & Neon
03:57 - Live Example: Churn Prediction Without Training
04:58 - TabFM’s Blind Spot: Semantic Understanding
06:00 - TabFM Auto & Gemini Orchestration
06:58 - Database Requirements (Branching & Leakage Prevention)
08:52 - Production Setup & Scheduled Pipelines
09:49 - Key Takeaways & Architectural Blueprint
Thanks to Neon for sponsoring this video. Check out Neon for your backend needs for your apps and agents: get.neon.com/4dTo3CI
Google Research's TabFM is a zero-shot foundation model for tabular data: you pass it the rows you already know, and it predicts the rest in a single forward pass, with no training. In this video I explain how TabFM learns from synthetic tables, why it can't read column names, and how TabFM-Auto adds an LLM agent around it. Then I build that agent on Neon Postgres to predict customer churn, and show how a leaky column fooled the agent and how a branch of the database's history fixed it.
Neon: get.neon.com/4dTo3CI
TabFM blog post: https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/
TabFM paper: arxiv.org/abs/2609.37959
TabFM-Auto paper: arxiv.org/abs/2609.37989
TabFM on GitHub: github.com/google-research/tabfm
TabFM on Hugging Face: huggingface.co/google/tabfm-1.1.0-pytorch
Timestamp:
00:00 - The Problem with LLMs on Tabular Data
01:52 - What is Google TabFM?
02:26 - How TabFM Works (Synthetic Causal DAGs)
03:17 - The Architecture: Database Backends & Neon
03:57 - Live Example: Churn Prediction Without Training
04:58 - TabFM’s Blind Spot: Semantic Understanding
06:00 - TabFM Auto & Gemini Orchestration
06:58 - Database Requirements (Branching & Leakage Prevention)
08:52 - Production Setup & Scheduled Pipelines
09:49 - Key Takeaways & Architectural Blueprint








![System 1 vs System 2 AI models: whats actually different?
A System 1 model like Jev takes one look: a single pass, no scratchpad. A System 2 model runs the same brain in a loop, writing notes and building on them step by step. I tested both on chains of questions that get longer each time. Jev falls off a cliff; Qwen 3.8 with thinking barely moves.
The rule: count the steps. If the answer is visible in one look, use System 1. If it needs results you dont have yet, use System 2.
Full breakdown and the experiment you can run yourself: [link to full video]
#AI #LLM #Jev #SystemOne #SystemTwo #AIModels #Shorts System 1 vs System 2 AI models: whats actually different?](https://i.ytimg.com/vi/8daNaFE-qOg/mqdefault.jpg)

