What Parametric does

Parametric PBC is a San Francisco-based AI research company founded in 2025 that applies mechanistic interpretability techniques to Vision-Language-Action (VLA) models used in robotics. The company trains transcoders on third-party foundation models — notably π₀-FAST, an open-source VLA developed by Physical Intelligence — to extract human-interpretable features from internal neural representations. Its published research identifies two distinct feature classes: "scene" features that fire on objectives and environmental landmarks in external camera input, and "ego" features that encode robot-centric information such as end-effector state. The company's stated objective is to "white-box" black-box robotics models so that general-purpose robots can be deployed safely and reliably, with demonstrated potential to steer model behavior by encouraging or suppressing identified features.

The company serves AI/ML researchers and robotics engineers, and operates a public benefit corporation structure with a Y Combinator F25 affiliation. Its technology stack leverages the TopK activation function with L2 penalty (following Paulo et al.) and sweeps expansion factors in [16, 32, 64, 128] and k in [32, 64, 128]; training used 1 billion activations with 250 million reserved for validation, drawn from the DROID dataset captured on a Franka Panda arm. The company publishes findings openly on its research blog rather than through gated commercial channels, and has no disclosed pricing, SDK, or commercial API.

Commercially, Parametric is pre-revenue. Its stated revenue model is "licensing royalties," which appears prospective rather than active. The company has raised $500,000 from Y Combinator (December 2025), employs 1–10 people, and has not announced customers, partnerships beyond its research engagement with Physical Intelligence, or follow-on funding rounds. The leadership team consists of co-founders Cody Swain (CEO) and John Newsom (CTO), both co-authors of the company's research. Strategic direction as articulated on the company website points toward "physical businesses that run themselves," signaling eventual ambition to move from interpretability research into enabling autonomous physical operations.

Name

Parametric

Legal name

Parametric PBC

Website

https://www.parametric.company

Company type

Private

Founded year

2025

Operating status

Operating

Headcount range

1–10 employees

Short description

Parametric PBC is an AI research company that applies mechanistic interpretability to robotics VLA models, training transcoders on systems like π₀-FAST to extract human-interpretable features and enable safer, steerable deployment of general-purpose robots for researchers and engineers.

Ownership category

akta.pro rank

Product category

AI Robotics Interpretability

NAICS

Custom Computer Programming Services (541511)

SIC

Services-Computer Integrated Systems Design (7373)

akta.pro primary industry

Model Transparency, Explainability & Interpretability (XAI) (HDAAAMAB)

akta.pro secondary industry

Safety, Alignment & Content Moderation (Guardrails, Red Teaming) (HDAAACAI)

Keywords

• Mechanistic interpretability
• Robotics AI safety
• Vision-language-action models
• AI transparency tools
• Robotics interpretability

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Operations

Revenue model

1. Research and Development Services: As an early-stage Y Combinator company focused on fundamental AI safety research, Parametric appears to be in a pre-revenue research phase, publishing findings to build academic credibility and potential licensing opportunities.

Go-to-market motion1 record

Distribution channels1 record

Marketing channels2 records

Core offering

Parametric conducts research and develops mechanistic interpretability techniques for Vision-Language-Action (VLA) robotics models. The company trains transcoders on VLA models such as π₀-FAST to uncover human-interpretable internal features (categorized into "scene" and "ego" feature classes), with the goal of enabling safer, more controllable deployment of general-purpose robotic systems through behavioral steering and validation of AI decision-making.

Product overview

Parametric is an AI research company (Y Combinator F25) focused on developing interpretability tools for vision-language-action (VLA) models in robotics. The company conducts research on making black-box robotics models explainable through techniques like training transcoders on VLAs (such as π₀-FAST from Physical Intelligence). Their work centers on identifying 'scene' features that highlight objectives and 'ego' features focused on the robot itself, with the goal of enabling safer, more controllable robotic systems.

Differentiator

Problem solved

Functional benefit

Quantifiable outcome

• Over 98% of features fall into 'image' regions; identified features with high multimodality characteristics
• +2 more outcomes

Segments1 record

Ideal customer profiles1 record

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Feature4 records

Partnerships

One partnership is on record.

• Physical IntelligencecoreTechnology or IntegrationPhysical Intelligence developed π0-FAST, the open-source Vision-Language-Action model that Parametric analyzed for their interpretability research. Parametric trained transcoders on π0-FAST to uncover interpretable features in the model's internal representations. Physical Intelligence's model serves as the foundation for Parametric's research on VLA model interpretability techniques.

Scale indicators1 record

Recent moves5 records

Emerging players

• Apollo Research: Apollo Research focuses on evaluating and interpreting advanced AI systems, including alignment-focused interpretability research. It overlaps with Parametric in mission and emerging-stage positioning within AI safety/interpretability.
• METR (Model Evaluation and Threat Research): METR is a non-profit research organization that evaluates frontier AI models for dangerous capabilities and reliability. It addresses safe deployment of advanced AI systems, a mission adjacent to Parametric's VLA interpretability work and a potential collaborator or counterparty for shared safety tooling.
• Conjecture: Conjecture is an AI safety startup focused on aligning and controlling advanced AI systems. It is an adjacent peer addressing safe deployment of frontier AI, sharing customer overlap with Parametric's mission-driven interpretability work.
• Transluce: Transluce develops AI observability and interpretability tools for foundation model behavior. It is a comparable early-stage peer operating in mechanistic interpretability with overlapping researcher talent and similar target buyers.

Direct peers

• Goodfire: Goodfire builds an AI interpretability platform focused on mechanistic understanding of foundation models. It is the closest direct competitor to Parametric in bringing interpretability tooling to model developers, including enterprise teams deploying AI systems.

Others

• Physical Intelligence: Physical Intelligence develops the π₀ family of VLA foundation models that Parametric's research is built on. While not a direct competitor, it is the strategic ecosystem partner whose model roadmap most directly shapes Parametric's opportunity and risks.

Broad incumbents

• Anthropic: Anthropic operates a prominent mechanistic interpretability research group that originated many of the sparse-autoencoder and transcoder techniques Parametric applies. As a frontier model lab with vastly more capital, it sets the research agenda and talent benchmark in the interpretability space.

Market position

Strengths5 records

Weaknesses4 records

Competitive moat3 records

Key risks5 records

Key highlights5 records

Customer concentration

• Xx.com/parametricpbc
• LinkedInlinkedin.com/company/parametric-pbc

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Funding overview

Funding rounds1 record

Investors2 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Parametric