Frequently Asked Questions About Emit Jane Luma Foundation Lab Method
21 answers covering everything from basics to advanced usage.
// Basics
What's the difference between a foundation lab and a traditional AI research lab?
A traditional AI research lab publishes papers and hands off models to product teams who build separately. A foundation lab eliminates this handoff entirely—product and research are one unified system. Research produces the product directly, and every product deployment generates training signals for the next model. This joint optimization creates compounding returns impossible in siloed organizations where research and product have separate roadmaps, separate incentives, and separate timelines.
What does end-to-end optimization mean in the Foundation Lab Method?
End-to-end optimization means jointly optimizing the entire stack—from base model pre-training through product deployment and back to data collection—as one system. Never optimize a narrow sub-problem in isolation for months. If the model can't do something today, that's a data collection job for the next training run, not an engineering workaround project. Every product decision must feed back into the model, and every model improvement must make the product better.
Why does the Foundation Lab Method say to think in professions instead of verticals?
Verticals like 'entertainment' or 'healthcare' are abstractions that obscure the real end-to-end workflows humans need solved. Professions are concrete: filmmakers have a specific workflow from concept to final output, architects go from brief to permit-ready package. Targeting professions forces you to design for complete workflows with specific failure modes and magic moments, rather than building generic tools that solve fragments no one asked for in isolation.
What is a Forward Deployed Creative and how is it different from a sales engineer?
A Forward Deployed Creative (FDC) is a Luma-invented role analogous to Palantir's forward deployed engineers but for creative and visual domains. Unlike sales engineers who focus on closing deals and technical support, FDCs serve two simultaneous functions: helping enterprise customers deploy AI systems into complex organizational workflows and piping intelligence from real customer usage directly back to model research and training. Every enterprise deployment becomes an optimization loop feeding the data flywheel, not just a customer support interaction.
// How To
How do I identify the scarce data problem for my AI modality?
Ask two questions: Is there a YouTube of this modality? Is there a Wikipedia of it? If the answer to either is no, you have a scarce data problem. Your first product must generate that data at scale—build something people love to use for free that produces training data as a byproduct. Don't wait to know exact scale requirements; scaling laws for new modalities are unknown early. Expect uncertainty about whether you need 1 million or 1 trillion examples.
How do I build a thin product stack on top of my AI model?
Map every feature in your product that compensates for something the model can't do natively. Each workaround, orchestration layer, or engineering harness represents 'fatness' in your stack. Resist the urge to build complex systems to paper over model gaps—these are six-to-eight-month dead ends. Instead, build the minimum product that delivers real value today, flag each gap as a data collection job for the next training run, and expect each model iteration to reduce the product layer.
How do I capture process data from my deployed AI product?
Design your product so that the full path to every artifact—not just the final output—is logged and usable for training. Record every user action, iteration, decision, undo, and refinement. Deploy agents to real customers and observe how the best practitioners use them. Forward Deployed Creatives serve this dual function: helping customers succeed while piping usage intelligence directly back to model training pipelines. Every enterprise deployment is an optimization loop, not a support ticket.
How do I apply the 10x logarithmic scaling test?
Before committing to a major training run, ask: if the next model were 10x larger in compute and parameters, would it be a categorically different thing—not just incrementally better? If the answer isn't an obvious yes, diagnose the real constraint. Is it insufficient modality coverage (missing audio or language tower)? Data quality or process data gaps? Architectural limitations? Fix the actual bottleneck before scaling. This prevents burning compute budgets on problems that scale alone cannot solve.
// Troubleshooting
What if my AI product team and research team are already separate organizations?
You must restructure. The Foundation Lab Method is explicit: treating product and research as separate teams with separate roadmaps is the first pitfall listed. This isn't a nice-to-have organizational preference—it's the core thesis. Start by creating shared objectives where every research output must improve the product and every product deployment must generate training data. Eliminate separate roadmaps. Move toward a single unified team with one set of metrics measuring the compound loop.
What do I do when my model has capability gaps but I need to ship a product now?
Ship the thinnest possible product that delivers real value despite the gaps—but critically, do not build elaborate engineering harnesses to compensate. Each gap is a data collection job for the next training run, achievable in two to three weeks, not a six-to-eight-month engineering project. The product you ship today has two jobs: serve customers and generate training signals. Accept temporary limitations rather than building complex workaround systems that the next model iteration will make irrelevant.
Why does my consumer AI app have high initial engagement but terrible retention?
You're likely below the intelligence threshold. The Foundation Lab Method predicts this exact pattern: consumer generative products create novelty spikes followed by retention collapse when the model doesn't understand context, humor, or the user's local state. Generated content is not interesting because it's generated—it's interesting because of what's happening in it. Until your model passes this threshold, focus on enterprise and professional deployments where end-to-end workflow value doesn't depend on contextual entertainment intelligence.
What is the biggest mistake AI startups make according to the Foundation Lab Method?
Building complex engineering harnesses to paper over model capability gaps. This is described as a six-to-eight-month dead end. Teams spend months building orchestration systems, multi-agent pipelines, and workaround harnesses—only to have the next model iteration make all of it irrelevant. The correct response is to treat every capability gap as a two-to-three-week data collection job for the next training run. The product should be thin; the model should do the work.
How do I convince my investors or board that product and research shouldn't be separate?
Frame it as a compounding vs. linear growth argument. Separate product and research teams produce linear improvements—each side makes progress independently. A unified foundation lab produces compounding returns—every product deployment generates training data that makes the next model better, which makes the product thinner and more capable, which generates more users and data. Show the six-to-eight-month dead-end cost of engineering harnesses that get invalidated by the next model. The math favors the compound loop at every time horizon beyond six months.
// Comparisons
How does the Foundation Lab Method compare to the typical Silicon Valley approach of shipping fast and iterating?
The Foundation Lab Method doesn't oppose speed—it opposes building the wrong thing fast. Traditional 'ship fast' culture often produces thick product harnesses to compensate for model gaps, creating technical debt that each new model iteration invalidates. The Foundation Lab approach ships thin products fast while simultaneously treating every gap as a training data problem. Speed is directed at the model-product compound loop rather than at engineering workarounds. The iteration unit is the training run, not the sprint.
How does the Foundation Lab Method differ from how OpenAI or Google structure their AI product strategy?
Most large AI labs operate with research teams and product teams as distinct organizations with separate roadmaps and incentives. The Foundation Lab Method argues this separation is fundamentally wrong—it prevents the compounding loop where product generates training data and research directly produces product improvements. Additionally, large labs often build separate modality towers (one for language, one for vision, one for audio), while the Foundation Lab Method insists on a unified single tower architecture as the only path to true world models.
// Advanced
Can I apply the Foundation Lab Method if I'm building on top of someone else's model like GPT or Claude?
Only partially. The full method requires control over the base model training loop because the core thesis is joint optimization between product and research. If you don't control the model, you can't feed product data back into training. You can apply the profession-targeting, thin-stack, and end-to-end workflow principles. But the data flywheel and compound loop—the most powerful elements—require either training your own models or having deep partnership with a model provider who will incorporate your process data into their training runs.
How do I know when my model is ready for consumer deployment vs. staying enterprise-focused?
Apply the intelligence threshold test: does the model understand why specific content would be interesting to a specific person in a specific context? Can it grasp context, humor, and the user's local state? If not, consumer deployment will produce novelty spikes followed by retention collapse. Enterprise is correct when the model can solve clear end-to-end professional workflow problems. Businesses are responsible for 99% of pixels on screens daily and have concrete workflow problems solvable without the contextual entertainment intelligence consumers require.
What's the relationship between unified models and world models in this framework?
A unified model is the architectural path to a world model. A unified model fuses language, image, video, and audio tokens into one single backbone, jointly trained. A world model goes further—it has understanding of the physical world and can simulate it, comprehending laws of physics, causality, time, and human language as one signal stream. The unified model architecture (single tower) is necessary because separate modality towers cannot jointly optimize and thus cannot develop the integrated physical understanding that defines a world model.
Why shouldn't I build separate AI models for different modalities?
Separate modality towers cannot jointly optimize, preventing the model from developing true physical world understanding. A unified single tower—one backbone processing language, audio, video, images, and physical context as one signal stream—enables things categorically impossible with separate towers. For example, understanding a character's identity across a long film production, or reasoning about visual states referenced in code. The human brain doesn't have separate systems per modality; your model shouldn't either.
How do I prioritize which modalities to fuse first in a unified model?
Language + video + audio covers approximately 90% of the path to a world model. Start with the highest-leverage fusion—typically language + image or language + video—and expand from there. At each fusion step, measure whether the combination enables things that were categorically impossible before with separate models. If a fusion only produces incremental improvement, the integration approach needs rethinking. Physical context (3D, robotics, embodiment) comes after the core language-vision-audio fusion is working.
What does it mean that foundation labs can launch products at 1% of the balance sheet?
Because foundation lab economics are driven by compute and research rather than individual software products, the marginal cost of launching a new product is approximately 1% of the company's total resources. The base model and infrastructure already exist; a new product is a thin layer on top. This is fundamentally different from traditional software companies where each new product requires its own engineering team, infrastructure, and significant capital. It's the economic advantage of the unified product-research architecture.