Emit Jane Luma Foundation Lab Method

// When should I use the Luma Foundation Lab Method?

Use this skill when designing or evaluating an AI company's research-product strategy, deciding whether to build a narrow vertical product vs. a generalist system, or planning how to couple model training with real customer usage data.

// What inputs do I need to apply the Foundation Lab Method?

• Company or project descriptionrequired
  What the company or project is building — the core modality, problem domain, or product area (e.g., visual AI, coding agents, robotics).
• Current stagerequired
  Where the team is today: pre-product, early product, scaling, or enterprise deployment.
• Target professions or customer segmentsrequired
  Which professions or types of people the product serves — framed as professions, not verticals (e.g., filmmakers, marketers, game artists, not 'entertainment industry').
• Known model capability gaps
  What the current model cannot do that the product needs — the delta between model capability and product promise.
• Data availability
  What proprietary or scarce data exists or could be collected that the internet cannot supply (e.g., process data, how artifacts were made, not just the artifacts themselves).

// What are the core principles of the Foundation Lab Method?

// How do you apply the Foundation Lab Method step by step?

1. 1

   Define the North Star as Multimodal AGI and Joint End-to-End Optimization

   Before any product or research decision, state the prime directive explicitly: (1) multimodal AGI as the destination and (2) joint end-to-end optimization as the method. Every subsequent decision is evaluated against these two poles. If a product decision does not feed back into the model and if a model improvement does not make a product better, it is misaligned.

2. 2

   Identify the scarce data problem for your modality

   Ask: is there a YouTube of this modality? A Wikipedia of it? If no, the first product must generate that data. Do not wait to know the exact scale needed — scaling laws for new modalities are unknown early. Release something people love to use for free that produces data at scale. Expect to not know if you need 1 million or 1 trillion examples.

3. 3

   Map current base model capability honestly against the end-to-end product promise

   List what the model can and cannot do today. Flag every gap where the product currently requires an engineering harness or workaround. Each gap is not an engineering project — it is a data collection and training job. Categorize each gap: does it require a fine-tuning run, a new training run, or a full pre-training investment in compute?

4. 4

   Build the thinnest possible product stack on top of current model capability

   Resist the urge to build complex orchestration systems to paper over model gaps. Build the thinnest product that delivers real value to real professions today. Fatness in the product stack is technical debt that the next model iteration must pay down. The product's job is also to generate the training signal for the next model.

5. 5

   Target professions, not verticals

   Reframe every market conversation from 'which vertical' to 'which profession.' Professions have specific workflows, specific failure modes, and specific magic moments. Ask: what does end-to-end look like for this profession? A filmmaker's end-to-end is concept → shoot → edit → set changes → final output — not 'make a clip.' A marketer's end-to-end is understanding the environment → resonant message → localized assets at scale.

6. 6

   Deploy Forward Deployed Creatives (FDCs) to enterprise customers

   FDCs (Forward Deployed Creatives) are not sales engineers — they serve two jobs simultaneously: (1) help customers actually deploy powerful systems into their complex organizational workflows, and (2) pipe the resulting intelligence — what works, what breaks, what data is needed — directly back to research and model training. Treat every enterprise deployment as an optimization loop, not a support ticket.

7. 7

   Capture process data, not just artifact data

   The internet supplies artifacts (movies, images, code). It does not supply how those artifacts were made — the actions, iterations, and decisions. End-to-end agents require process data. Every interaction in your deployed product is a training signal. Build the product so that the path to the artifact, not just the artifact, is logged and usable for training.

8. 8

   Apply the 10x logarithmic scaling test at each model iteration

   Before each major training run, ask: if this model were 10x larger in compute and parameters, would it be categorically different — or just incrementally better? If the answer is not obvious yes, the bottleneck is not scale. Diagnose whether the constraint is: (a) insufficient modality coverage (e.g., missing audio, missing language tower), (b) data quality/process data gaps, or (c) architectural limitations. Fix the real constraint before scaling.

9. 9

   Unify modalities into a single tower progressively

   The shape of a world model is a single tower that jointly models language, audio, video, images, and physical context as one single signal stream. Do not build separate towers per modality — fuse them. Prioritize: language + video + audio covers approximately 90% of the path to a world model. Start with the highest-leverage fusion (language + image or language + video) and expand. Measure whether each fusion enables things that were categorically impossible before.

10. 10

    Evaluate consumer vs. enterprise deployment using the intelligence threshold test

    Consumers consume; creators create. A generative product aimed at consumers is premature until the models are intelligent enough to understand context, humor, and the local state of the user. Apply the test: does the model understand why this content would be interesting to this specific person in this specific context? If no, enterprise deployment is the correct focus — businesses are responsible for 99% of pixels on screens every day and have clear end-to-end problems the model can solve now.

// What does the Foundation Lab Method look like in practice?

// What mistakes should I avoid when using the Foundation Lab Method?

• Treating product and research as separate teams with separate roadmaps — in a foundation lab they are one unified system and must be jointly optimized.
• Building complex engineering harnesses to paper over model capability gaps — this is a 6-to-8-month dead end; the correct response is to treat the gap as a 2-to-3-week data collection job for the next training run.
• Thinking in verticals instead of professions — verticals are abstractions that obscure the actual end-to-end workflow a human needs solved.
• Chasing consumer deployment before the models are intelligent enough to understand context and local user state — this produces novelty spikes followed by retention collapse, as the content is not interesting because it is generated.
• Collecting only artifact data (finished outputs) and not process data (how the artifact was made) — agents that do end-to-end work require process data that the internet cannot supply.
• Assuming scale alone (10x parameters, 10x compute) will fix a categorical capability gap — if the answer to 'would 10x scale make this categorically different' is not an obvious yes, the constraint is architectural or data quality, not scale.
• Building separate modality towers instead of a unified single tower — separate towers cannot jointly optimize and prevent the model from developing true physical world understanding.
• Solving spot work problems and calling it AI transformation — the promise of AI is the full end-to-end solution (the book, the campaign, the film), not a faster fragment of the workflow.

// What are the key terms and concepts in the Foundation Lab Method?

Foundation Lab

A company architecture in which product and research are not separate functions but one unified system. Research produces the product; the product works in research. Foundation labs are described as 'the blueprint of companies of the future' because their economics are driven by compute and research, not by individual software products, enabling new products to be launched at approximately 1% of the balance sheet.

End-to-End Optimization

The prime directive and guiding methodology: joint top-to-bottom optimization across the full stack — from base model training through product deployment and back. The only way AI systems will do meaningful things in the world. The opposite of optimizing a narrow sub-problem in isolation.

World Model

A model that has understanding of the physical world and is able to simulate it. Not defined by real-time speed or autoregressive architecture. Defined by understanding laws of physics, causality, time, and human language — all as one single signal stream. The shape of a world model is a single tower jointly modeling language, audio, video, images, and physical context.

Unified Model

A single-backbone model with a language tower and one or more modality towers (image, video, audio) fused into one single thing, jointly trained on both language tokens and continuous signal tokens. The unified model is the architectural path to a world model. Unified models enable things categorically impossible with separate modality towers.

Single Tower

The architectural ideal for a world model: one model that processes language, audio, video, images, and physical context as one single signal stream without separate towers per modality. Analogous to the human brain operating across all modalities without separate systems.

FDC (Forward Deployed Creative)

A Luma-invented role analogous to Palantir's forward deployed engineers, but for creative and visual domains. FDCs serve two simultaneous functions: (1) help enterprise customers deploy powerful AI systems into their complex organizational workflows, and (2) pipe intelligence from real customer usage directly back to model research and training pipelines.

Process Data

Training data that captures how an artifact was made — the actions, iterations, and decisions in the path to a final output — as opposed to artifact data (the finished output itself). Process data is what the internet cannot supply and what end-to-end agents require to learn to do end-to-end work.

Promise of AI Is Not Spot Work

A core principle stating that AI's value is not in doing fragments of workflows faster ('I can make copy,' 'I can make a little image') but in delivering full end-to-end solutions — the book, the campaign, the film production. Spot work solutions will be commoditized; end-to-end solutions are the durable value.

Think in Logarithms

A scaling evaluation heuristic: when assessing the next major model investment, ask whether a 10x increase in compute or parameters would produce a categorically different model — not just an incremental improvement. If the answer is not obvious yes, the constraint is not scale but architecture, data quality, or missing modality coverage.

Thin Stack

The product architecture principle of building the thinnest possible product layer on top of base model capability. Fatness in the stack represents problems the model cannot yet solve natively. The next model's job is to reduce that fatness. Thick stacks built around model gaps become irrelevant with each new training run.

Intelligence Threshold Test

The test for whether a consumer generative product is viable: does the model understand context, humor, and the local state of the specific user well enough that the output would be genuinely interesting to that person? Below this threshold, consumer generative networks produce novelty spikes followed by retention collapse.

// FREQUENTLY ASKED QUESTIONS