Everything Intelligent Adapts
Why Mozilla Ventures backed Adaption’s bet against brute-force AI
For the past decade, AI progress has followed a simple formula: bigger models, more data, more compute. It worked. But the returns are diminishing and the costs keep compounding. Power, capital, and control have concentrated in a handful of labs that can afford the next doubling.
Meanwhile, the rest of us have become elevated prompt engineers. We contort our requests to fit the model’s limitations. We rephrase often, and we memorize recipes – “think step by step,” “you are an expert in…” – hoping the system will finally do what we actually need.
This relationship is backwards. Tools should adapt to people. Not the other way around.
Adaption is building AI that works the way it should: systems that evolve with the world, learn from context, and meet people where they are. Mozilla Ventures backed Adaption because this approach reflects a shared belief in open, community-driven technology. Learn more about Adaption on their website here.
Introducing Adaption
Adaption was founded on a simple premise: intelligence is not static, and AI should not be either. Instead of one-size-fits-all architectures optimized for averages, Adaption is building flexible systems that can be tuned to specific environments and requirements. Today’s AI remains expensive to customize and slow to change. Adaption’s work centers on lowering those barriers through efficiency-first design, modular composition, and data-efficient alignment.
The mission is to build systems that evolve across languages, cultures, and local realities.
Why this moment
Sara Hooker has been thinking about the limits of scale for years. Her recent essay, “On The Slow Death of Scaling,” lays out the case: scaling laws – the idea that performance predictably improves with compute – have become the industry’s central article of faith. But the evidence is shakier than the consensus suggests.
Scaling laws hold reasonably well for predicting pre-training loss. But when you measure actual downstream performance – the tasks people care about – the relationship gets murky. Some capabilities scale predictably. Others plateau, or emerge unpredictably, or don’t arrive at all. As Sara puts it: “The acceptance that there are emergent properties which appear out of nowhere is another way of saying our scaling laws don’t actually equip us to know what is coming.”
The frontier labs that bet everything on compute are likely under-investing in other directions. And those other directions are where the interesting work is now: gradient-free optimization, test-time compute, synthetic data that lets you reshape the training distribution on the fly, interface-algorithm co-design. These techniques can deliver 5-20x performance gains with a fraction of the compute.
This is the moment Adaption is built for. Not the end of progress – but a shift in where progress comes from.
Why this team
Mozilla Ventures doesn’t back teams for generic “technical credibility.” We back founders whose work and worldview align with ours.
Sara Hooker grew up across Africa – Mozambique, Eswatini, Kenya, Liberia. She went to a Portuguese-language middle school. For her, the gaps in AI language coverage aren’t abstract.
At Cohere, she led the Aya project: 3,000 researchers across 119 countries, building AI that works in 101 languages – more than 50 of which had never been served by generative AI. Time Magazine named her one of the 100 most influential people in AI.
Sara has been publishing on efficient adaptation, data pruning, and the limits of scale for years – before it was fashionable to question the scaling consensus. Her research on model merging, multilingual optimization, and synthetic data generation shipped in production systems like Aya Expanse.
Sudip Roy spent seven years at Google building ML infrastructure – Pathways, TensorFlow Extended – that powers thousands of production models, including Gemini. At Cohere he led inference: making AI work reliably and efficiently at scale. This team knows what to build and how to ship it.
AI ownership should not be a privilege
True AI ownership today is out of reach for nearly everyone. Developers build on distant monolithic models controlled by a privileged few. Customization requires steep expertise, expensive compute, and massive datasets. Builders who care about privacy, sovereignty, or local needs face a constrained choice between opaque APIs or limited local alternatives.
This centralization is economic as much as technical. It locks innovation behind unpredictable pricing, shifting vendor rules, and platform dependency.
Adaption is designed as both a technical and economic response built across three layers: continuously refining data to meet user intent, building intelligence that adapts across industries and specializations, and redesigning the human-AI interaction layer. Together, these make advanced AI capabilities more reliable and deployable in practice.
A bet on builders, not gatekeepers
Adaption is bringing together talent obsessed with new learning paradigms. It invites builders who want to reimagine the building blocks of AI progress.
Mozilla’s developer network, open-source heritage, and community-building experience are strategic assets that accelerate this flywheel.
In practical terms, this means a future in which a developer at a regional hospital, a startup in an emerging market, or a nonprofit working in local languages can build and own specialized AI models without relying on centralized APIs.
Everything intelligent adapts
The last decade was defined by scale. The next will be defined by adaptability.
Mozilla Ventures backed Adaption because we believe the future of AI won’t be built by whoever has the most compute. It will be built by teams that figure out how to make systems learn efficiently, evolve continuously, and work for the people the monolithic models leave behind.
This is early,but a future where AI adapts to more people – to their languages, their businesses, their needs – is worth betting on.
That’s why Mozilla Ventures is proud to support Adaption.
