00

Introduction

dfdx labs is an independent R&D company led by Matthias Plappert. We take on hard technical problems at the frontier, figure out how to solve them, and build the first working prototype.

New technologies constantly create new possibilities. But there's a wide gap between "this is theoretically possible" and "here's something that actually works." Closing that gap requires a way of working that's closer to R&D than to traditional engineering: you can't just look up the answer, you have to figure it out.

That's what we do. Sometimes for ourselves, sometimes for our clients.

01

Past work

Matthias spent years doing exactly this kind of work:

  • Power market forecasting (dfdx labs, 2024 – 2026)

    We built advanced price forecasting models for the European power market for a client. Those models are used by our client for proprietary trading and battery storage optimization and have traded several terawatt-hours of power in production.

  • Large language models (OpenAI, 2020 – 2021)

    Matthias was part of the team that trained GPT-3.5 (which later powered the first version of ChatGPT) and built OpenAI Codex, which became the foundation for GitHub Copilot—one of the first successful products built on top of LLMs.

  • Dexterous robot manipulation (OpenAI, 2017 – 2020)

    Matthias was part of the team that built Dactyl, a robot hand that could solve a Rubik's cube. This was a problem widely considered impossible at the time. It required combining reinforcement learning, simulation, and physical robotics into a system that actually worked in the real world.

  • Hardware/software integration (Square, 2012 – 2013)

    Matthias helped build the foundation for how Square's iOS point-of-sale system interfaces with hardware, including integration of the first Square Stand prototype.

02

Philosophy

Everything about how we work follows from a few principles.

  1. 02.1

    We focus on genuinely hard problems.

    We prioritize work where off-the-shelf approaches don't work and genuine research is needed to make progress. We have a strong preference for problems that cross disciplines—we bring approaches and intuitions from one field into another that domain specialists typically wouldn't consider.

  2. 02.2

    We build things that work.

    The goal is always a working prototype—something useful, not a slide deck or a research paper. We're not done until it runs in the real world.

  3. 02.3

    We work independently or with clients.

    We split our time between independent projects and client work. Independent projects are prototypes of things we believe should exist but don't yet. When a client has a genuinely hard problem, we partner with them to build the first working version.

03

Working with us

We work best with people who know what they want but not yet how to get there. There are two ways to work with us: research & prototyping, or advisory.

  1. 03.1

    Research & Prototyping

    • Hands-on work where we take on a problem and build a working prototype.
    • The ideal project has a hard technical core, a clear definition of what success looks like, and is independent from the client's day-to-day operations.
    • Projects are structured around milestones with clear deliverables. Either side can walk away at a natural stopping point if the work isn't showing promise. R&D has a real failure rate—the goal is to fail fast and cheaply, not to keep going out of obligation.
  2. 03.2

    Advisory

    • Execution stays with the client; we play a supporting role.
    • We work with technical teams navigating unfamiliar territory, leadership teams making strategic decisions about new technology, and investors evaluating technically complex companies.
    • Advisory complements prototyping naturally: many engagements start as prototyping and transition to advisory for the handoff and productionization.
  3. 03.3

    How we usually work together

    1. We meet to discuss the challenge you're facing, and figure out if we're a good fit.

    6. We collaborate to clarify the problem, define success criteria, and set a first milestone.

    26. We test the riskiest assumptions quickly to decide whether the path is viable.

    46. A first end-to-end prototype proves the core concept in a realistic setup.

    66. We harden weak points, improve reliability, and align with your constraints.

    101. We deliver outcomes, document key learnings, and plan next decisions together.

    This timeline might not fit all engagements, but it's a good starting point for understanding how we usually work with clients.

04

Get in touch

If you're interested in working with us, please get in touch.

The more detail you provide, the faster we can figure out if there's a fit.