
A memory worth keeping
A sensor may pick up a sound that lasts only an instant. Dirac measures give us a way to describe an event at a precise moment. Which detail is worth keeping, and how much work will it take to recall it later?

Local & Hybrid Hyper AI
We want AI to learn on the devices we use every day, from laptops to prosthetic hands. When you need a larger model, you choose whether to bring it in.
Explore our research goalsSuppose a model takes ten hours to learn a task. Our goal for Christmas 2027 is to train ours in one hour on the same computer and have it do the job just as well. We’re committing ten months of research to that 10× speedup. We also want our model to use less memory, so a field team could run it on a laptop it already owns.
Now imagine teaching a field recorder to recognise a robin’s song. When you take it out tomorrow, you’d like it to recognise both the robin and the blackbird you taught it last spring. We’ll test whether our models can adapt in just that way. In the longer term, we want self-learning models that can pick up new skills from experience, right there on the device.
Faster training is part of our wider ambition to make general AI 100× more efficient. For that field recorder, our energy target would mean identifying the same birds just as reliably on one hundredth of the energy.
Why local models matter
A survey drone could count nesting birds even when it flies beyond mobile coverage. At home, a washing machine could listen for the rattle of a worn bearing and let you know when it needs a repair. It could give you that warning without sending sounds from your kitchen to a cloud service.
A model running on your device can answer even when you’re offline. There’s no wait for a server to reply and no cloud fee each time you use it. You can keep a version you trust. If its licence allows, you can also adapt it yourself and choose which content filters your app uses.
An appliance, a small aircraft and a prosthetic hand all have room for only so much memory. The processor has to be affordable, too. With MCUNet, researchers at MIT showed that an off-the-shelf microcontroller, the kind of small chip found in everyday devices, could recognise images. We want to find out how much more a general-purpose model could do within those limits.
Google’s on-device AI guide · MIT’s MCUNet study · Ai2’s open model
Research directions
Since the Transformer appeared in 2017, researchers have shown how much language models can gain from more data, larger networks and more computing power. We’d like to know how much a model could gain by remembering the right detail, or by learning from the precise moment a signal arrives. The Transformer paper · Scaling laws

A sensor may pick up a sound that lasts only an instant. Dirac measures give us a way to describe an event at a precise moment. Which detail is worth keeping, and how much work will it take to recall it later?

Spiking neural networks send short pulses. The gap between two pulses can carry information. We’re asking whether a network could use that timing to learn, and what a neuromorphic counterpart to backpropagation might look like.

Liquid neural networks can follow a signal as it changes over time. Neural operators learn how one whole pattern maps to another. We want to see whether either approach could help a small device make sense of what its sensors are picking up.
Why BionicMind Labs?
Imagine walking a trail and feeling your watch tap out the time. A different rhythm could tell you your heart rate, or a familiar double tap could let you know an email has arrived. You could keep your eyes on the path.
Picking up a cup with a prosthetic hand takes that idea a step further. A small model could learn to read the electrical signals from your muscles as you reach for it. As the fingers close, could the hand send a signal back that lets you feel how tightly you’re holding the cup?
Parts of this are already possible. Some watches tell time by touch. In a prosthetics study, people used their own muscle signals to move a hand and could feel the pressure of its grip. At BionicMind Labs, we want to find out whether local models that learn from the timing of electrical pulses could bring that two-way sense of touch to more devices. That exchange is what we mean by a “feeling” machine.
The shared playground

Back in the field, an ecologist records a bird she can’t quite identify. One of her AI assistants finds a note about a blackbird she saw there last spring. Another listens to the new recording and suggests a robin. When she comes back tomorrow, she’ll want to know what each assistant noticed and why they disagreed.
That gives us a way to test how our agents’ memory works. In our shared playground, we’ll give several agents observations that overlap and sometimes conflict. Then we’ll see what they remember, what they forget and whether they can trace each suggestion back to the recording or field note that prompted it.
If the ecologist wants a second opinion, she could choose a few seconds of birdsong and send just that clip to a specialist AI online. Its answer would go into the field notes on her laptop, alongside her own observations. The rest of the recording would stay with her.
That’s what we mean by hybrid AI. We want people, open-source agents and proprietary models to work together through speech, text and sensors. If you subscribe to an AI service from OpenAI, Anthropic or Google, we want your local agent to be able to work with it. We plan to use each provider’s API under its own access and billing terms. You decide when to ask for help and what to share.
Early supporters
We’re laying the groundwork for a startup. If you build devices, use one in the field or are thinking about backing this research, I’d love to hear from you. What would you teach it, and what would you want it to remember?
Email Lucy Nowacki