Generalist AI

Generalist AI - GEN-0 and Robotics Scaling Laws

Generalist AI

Home > Companies > Generalist AI


Overview

Generalist AI is a startup founded by a team from OpenAI, Google DeepMind, and Boston Dynamics (per company announcement). They claim to have discovered the first scaling laws in robotics using 270,000 hours of real-world physical interaction data.

ItemDetails
AnnouncementNovember 4, 2025
Core ModelGEN-0
Key ClaimDiscovery of robotics scaling laws
Data270,000+ hours of real data
Bloggeneralistai.com/blog

Founding Team

Background

OriginContribution
OpenAIChatGPT, GPT-4 scaling
Google DeepMindPaLM-E, RT-2 development
Boston DynamicsAtlas, Spot, Stretch
OtherAutonomous driving technologies

GEN-0 Model

Key Features

ItemDetails
Training Data270,000+ hours
Data SourceReal physical interactions
EnvironmentsHomes, bakeries, laundromats, warehouses, factories
TasksFrom peeling potatoes to tightening bolts

Architecture: Harmonic Reasoning

A new approach that doesn’t rely on System 1-System 2:

┌────────────────────────────────────────┐
│           Harmonic Reasoning           │
│                                        │
│  [Sensing Stream] ←→ [Acting Stream]   │
│       (continuous, asynchronous)       │
│                                        │
│  "Harmonious interaction"              │
└────────────────────────────────────────┘
FeatureDescription
CoreHarmonizing continuous streams of sensing and action
AdvantageScalable to very large model sizes
Supported Robots6, 7, 16+ degrees of freedom

Robotics Scaling Laws

Key Findings (Company Claims)

Predictable scaling in robotics, similar to LLMs:

L(D) ∝ D^(-0.5)

L = Downstream task error
D = Pre-training data volume

→ 2x data → ~30% error reduction

Observations

FactorEffect
Pre-training data ↑Performance ↑
Compute ↑Performance ↑
PredictabilityConsistent and predictable improvements

Intelligence Threshold (Phase Transition)

7B Parameter Threshold

Model SizePhenomenon
< 1BFails to absorb complex data, “ossification”
7B+Internalizes data, continuous improvement

Implications

< 7B: Limited improvement even with more data
≥ 7B: More data → continuous improvement
      + Adapts to new tasks with minimal fine-tuning

→ 7B could be robotics’ “GPT-3 moment”


Data Collection

Scale

ItemValue
Total Data270,000+ hours
Collection Environments1,000+ locations
Weekly Growth10,000+ hours
TypeReal physical interactions only

Philosophy: Real Data Superiority

Simulation: Physical accuracy limitations
Human Video: Difficult action extraction
Teleop: Slow and expensive

Generalist Perspective: "Only real physical interaction is authentic"
→ Proven with 270K hours of real data

Cross-Embodiment

Multi-Robot Support

Designed from the ground up to support various robot forms:

DoFTested
6 DoF
7 DoF
16+ DoF

Significance

If GEN-0’s claims are true:

ImpactDescription
ScalingProves robotics can scale like LLMs
Economic JustificationPredictable ROI on large investments
Data DebateStrengthens case for real data superiority
Industry Reaction”Robotics’ ChatGPT moment”

References


See Also