Ghola · neuromorphic R&D
AI should learn where it lives.
Today AI is trained once in a data centre and rented out. We are building Ghola, a neuromorphic core that learns by itself on the device where it runs, and does it cheaply. This is an R&D programme: here is the vision, the road, and what already works.
Vision
AI learned to speak before it learned to remember.
Today intelligence is rented. A few companies train very large models in a few data centres; everyone else sends data there and gets answers back. The models are frozen after training and remember nothing on their own.
Memory is not storage. A brain tags what surprised it, strengthens what turned out to matter, lets the rest fade and consolidates what survived during sleep. No central process decides this. Each synapse follows local rules.
That is what makes it portable. Memory that needs no controller can run on the device itself: a chip, a sensor, a robot. It learns where it works, cheaply, and belongs to whoever runs it.
Many small minds, not one large one. We think the concentration of AI is a design choice, not a law of nature. Learning that grows on small, low-power hardware is the counterweight.
We decide what would prove us wrong first. Every stage is pre-registered, judged against published animal behaviour, and failures are published next to passes.
Roadmap
From learning rules to a learning chip.
Each step has a gate: a result that must hold before the next one starts. Dates after the current quarter are targets.
| When | Stage | What it answers | Status |
|---|---|---|---|
| 2026 Q2–Q3 | Learning rules | Can local rules reproduce how animals learn, extinguish and consolidate memories? | done |
| 2026 Q3 | Hardware arithmetic | Do the rules survive the low-precision maths of neuromorphic chips? | done |
| 2026 Q3–Q4 | Closed-loop learning | Does the core learn behaviour online in a body and keep it through sleep? Tested on a virtual worm. | in progress |
| 2027 Q1 | Baselines | Is it better than simple alternatives at the same budget of memory and compute? | planned |
| 2027 Q2 | Open benchmark | A closed-loop continual-learning task that others can run their systems on. | planned |
| 2027 H2 | Core on a chip | The same learning on a neuromorphic chip, with latency and energy measured. | seeking partners |
| 2028 | First pilot | A sensor product that learns its own machine in the field. | target |
Where it leads
The technology at the end of the roadmap is a learning core that chip makers and device builders can embed. These are the uses we are designing it for.
Sensors that learn their machine
A vibration or wearable sensor learns what normal looks like for this motor or this person, adapts when the regime changes, and relearns fast when it changes back.
edge · industrial · healthRobots that adapt on board
A robot or drone learns behaviour that keeps it charged and cool, on its own chip, without sending data home.
embodied AIAutonomy without a link
Systems that must keep learning where there is no network, and keep what they learned after a restart.
space · remoteThe core
What we have built so far.
Ghola keeps learning after it is deployed, with no backpropagation, no dataset and no training run. Each synapse updates itself from local signals: it is tagged when something surprising happens, consolidated during a sleep phase if the tag survives, and fades otherwise. Every result below was pre-registered before the run.
Animal learning effects reproduced
Tagging, extinction, spontaneous recovery, savings and renewal emerge from the same local rules.
Risky external prediction passed
A prediction frozen before the run held against independent published animal data.
Bits are enough
The rules survive fixed-point chip arithmetic: gates at 5 bits, consolidation at 7. The tagging rule runs bit-exact in Intel's Lava framework.
Learns while it acts
In a virtual body, the core learns online and keeps what it learned when moved to new environments.
Studio
We develop and pilot AI-native solutions with real teams.
Alongside the lab, our studio builds AI-native tools for four areas of work and tests them with the people who use them. What we learn about memory in real products feeds back into the core.
Operations
AI automation of routine intellectual work: preparing and summarising meetings, drafting protocols and documents, checking them against requirements, tracking decisions and follow-ups. People keep the judgement; the assistant takes the rest.
CaseOS
A case management system for experts. Each client case gets its own workspace where the assistant reads the materials, follows the expert's methodology, drafts the artifacts and keeps the context of every case over time.
Communications
Practice grounds for the conversations that matter most, against AI counterparts that push back.
Communication trainers
Public-speaking and media training for executives with AI audiences and journalists, including a "digital troll" who plays the hardest person in the room.
Learning
Simulators that let professionals practise skills that are hard to train on real people.
Anna 0.8
A synthetic psychotherapy client for training therapeutic speech. After each session it gives a psycholinguistic analysis of the therapist's speech, ready for supervision and case review.
Education
AI-native programmes, labs and assistants for universities, designed with the pedagogy first and the model second.
Team
Two founders, one lab, Amsterdam.
Alexander Eliseenko
CEO · research and psychology12+ years in organisational consulting and 5+ years designing the logic of ML systems. Leads the Ghola research programme. Acting head of the neuroscience lab at Central University; lecturer at HSE University.
Andrey Kulikov
CTO · engineering and ML18+ years in IT, security and big data. Former CEO of SocialLinks, which he took international in web investigation and cybersecurity. Builds the core, its test environments and every product we ship.
Contact
Test it, fund it, or build with us.
Researchers
Critique the design, replicate a stage, or run your system on our benchmark once it opens.
Chip makers and funders
Hardware partners, neuromorphic labs, R&D grants and consortia.
Pilot partners
Teams in operations, communications, learning or education who want to pilot an AI-native solution.