ROBOTS GO MENTAL

Brain-inspired AI models

AI that learns like a brain. Not by burning GPUs.

Robots Go Mental builds brain-inspired AI models and training methods. They are designed to need far less compute, memory and data than conventional models, run on the customer’s own infrastructure, and shift the paradigm so customers own their Ethernity models.

Our products Contact

The problems

Two problems we will solve

Problem 1

Bigger models, bigger GPUs, bigger bills.

Most progress in AI comes from scale. Each gain in capability tends to mean a larger model, more GPUs and a higher bill to train and run it. Long contexts add to the cost: in standard transformers, the cost of attention grows with the length of the context. And retraining a model on new data can degrade what it learned before.

Problem 2

Your data, their model, their IP.

Cloud AI services mean sending sensitive data to an outside provider and relying on a model the company can’t inspect, change or take with it. The know-how that goes into tuning it, often the company’s own IP, stays on someone else’s platform. Hosting open models in-house keeps the data inside, but needs a dedicated team to run them.

Products

Three AI/ML products

Foundation model · customer-deployed

Ethernity E9

A new foundational model with a brain-inspired architecture, trained from scratch: a sparse mixture of experts, linear recurrent net and a persistent latent memory. Built to run in the customer’s own environment or using hyperscalers.

Status: pre- and post-training · internal availability

How Ethernity works

Optimization layer · as a service

Mental

Guided Transfer Learning (GTL): a patented method that helps existing models learn a new task from very few examples, without overwriting what they already know.

Used by NASA GeneLab · PCT/EP2023/085102

How Mental works

Open-source engine · customer-owned AI

Myelin

The open-source engine: control plane, orchestrator and pipeline to build and run your own AI model. Open source for self-managed teams, with enterprise support, or as a managed service at a hyperscaler.

Developer · Enterprise · Enterprise Service

How Myelin works

Ethernity E9 · how it works

Powered by a brain-inspired cybernetic architecture that unites O(1) linear recurrent state tracking (Linear Recurrent Net), non-parametric episodic Ethernity Latent Memory vaults (ELM), and an asymmetric dense-sparse Mixture-of-Experts (MoE), Ethernity E9 delivers cognitive depth comparable to 14B–32B dense frontier models while activating only 4.173B parameters per token (25.04 GFLOPs/token). Unlike traditional Transformers, which suffer from quadratic memory exhaustion (O(S²)) and catastrophic forgetting, E9 deploys as a self-contained, air-gapped sovereign software stack running entirely on cost-efficient on-premises GPU hardware or private VPCs. E9 absorbs corporate memory continuously without degrading its weights, reduces compute requirements by 1.68× to 7.79× (30% to 87% fewer FLOPs), compresses memory by 20,500×, cuts MLOps fine-tuning costs by 80%, and slashes 3-year enterprise AI Total Cost of Ownership (TCO) by 70% to over 90% compared to legacy cloud rentals.

Ethernity E9 is a sparse mixture-of-experts model with 9.1 billion physical parameters and 4.14 billion active parameters per token. It is a new model, designed and trained from scratch, not a fine-tuned version of an existing open-weights model. Standard transformers attend over every previous token and keep a key-value cache that grows with the context. Ethernity replaces most of that with a fixed-size recurrent state and a separate latent memory, and spreads the work across specialized experts so that only part of the model is active for each token.

Architecture

Dense-sparse mixture of experts
Sparse expert layers route each token to its top 4 experts. Only a fraction of the parameters does work for any given token.
Shared language expert
One large expert is always active and handles grammar and sentence structure. The routed experts specialize in content.
Linear recurrence
Most layers keep a fixed-size recurrent state instead of attending over the full history, so memory per token stays constant, O(1), as the context grows.
Latent memory
Past context is compressed into persistent latent concepts that the model reads back when needed. The memory is tiered: a facts vault, fast local storage and, if needed, remote storage.

Training

Self-regulating learning rate
The model monitors its training signal for sudden topic shifts and lowers its learning speed when one occurs, to protect what it already knows.
Language kept apart from facts
Language rules and factual knowledge sit in separate parts of the model, so learning a new fact doesn’t disturb how it forms sentences.
Mixed-domain data
General text, code and mathematics are trained together. Dense mathematical content gets extra compute.
Checkpoint dampening
When older checkpoints or data are loaded, updates to the memory are dampened so the model stays stable.

Inference

Position tracking
Token positions and document order are tracked explicitly, so facts stay attached to where they appear in long documents.
Confidence filtering
During generation the model drops its lowest-confidence 5% of internal paths before producing an answer.
Read-only speculative decoding
To speed up generation, the model drafts future tokens. Drafts can read the latent memory but never write to it, so wrong guesses can’t pollute long-term memory.

Mental · how it works

Guided Transfer Learning adds one element to a standard training setup: a guidance matrix with one value for every parameter of the model. It sets how much each parameter is allowed to change when the model learns something new.

  1. Step 1

    Scout

    The model is trained briefly on many small tasks related to the target domain. For each parameter, Mental records how much it needed to change.

  2. Step 2

    Guide

    These measurements form the guidance matrix. Parameters that matter for this kind of task get high values; the rest stay low.

  3. Step 3

    Learn

    During the actual training, each parameter’s updates are scaled by its guidance value. Learning concentrates where it is needed, so the model can adapt from very few examples and leaves its existing knowledge largely intact.

Works with
Existing neural networks, including LLMs
Training data
Down to a single example per task
Delivery
As a service. Use Mental to trade compute for data, so you can post-train or fine-tune your Ethernity model with small datasets.
Applied
NASA GeneLab, RNA analysis for space biology. Read the paper (arXiv)
IP
Full international patent PCT/EP2023/085102

Myelin · how it works

Myelin covers the full path from raw company data to a running model, in the environment the customer chooses. Each stage ships as a ready-to-deploy container, so the pipeline can run end to end or slot into an ML pipeline a team already uses.

  1. Stage 1

    Data curation

    Collect, clean and prepare the customer’s data for training, inside the customer’s chosen environment.

  2. Stage 2

    Training

    Train and tune Ethernity on that data, with Mental where only a few examples exist.

  3. Stage 3

    Inference deployment

    Deploy the tuned model for production use, on the customer’s own infrastructure or managed at a hyperscaler.

Open source

Developer

The full pipeline for teams that run and manage it themselves.

Self-managed · supported

Enterprise

Runs on the customer’s own infrastructure, with support from Robots Go Mental.

Fully managed

Enterprise Service

The managed version, running at a hyperscaler, so the customer’s team doesn’t operate the pipeline.

Customer-owned AI

Your model, on your infrastructure.

Today, companies choose between renting AI from a cloud provider and running open models themselves. Customer-owned AI is a third option: Ethernity is installed in the customer’s own environment, adapted to their data there, and the result belongs to the customer. There is no dependency on a closed AI vendor, and the data stays under the customer’s control.

What the customer owns

  • Model
  • Codebase
  • Data
  • Parameters
  • Training
  • Tuning

How a deployment works

  1. Install. Ethernity runs with the Myelin pipeline on the customer’s own infrastructure (Developer or Enterprise edition) or managed at a hyperscaler (Enterprise Service). Myelin’s containers also fit into an existing ML pipeline.
  2. Adapt. Myelin curates the customer’s data and tunes the model in that environment. Mental can be used where only a few examples exist.
  3. Own. The tuned model, its parameters and code stay with the customer, licensed for enterprise use.

Where we are

Early stage, built step by step

  1. Now

    Internal availability

    Ethernity E9 is in pre- and post-training and tested on internal benchmarks.

  2. Next

    Dedicated GPU cluster

    Our first cluster to evaluate E9 on real enterprise data.

  3. Then

    Pilot deployments

    Customer-owned deployments with a small group of enterprise pilot customers.

  4. Later

    Enterprise API

    Ethernity and Mental available through an API.

We will publish performance results once they are validated on real enterprise data.

Team

Enterprise builders and a brain scientist

Portrait of Davor Andric

Davor Andric

CEO

Software and AI/ML platform engineer. Built and scaled DXC Robotic Drive into a business of more than $300M. At AWS, worked on Amazon SageMaker Ground Truth and the incubation of Amazon Q to 20,000 customers.

Portrait of Danko Nikolić

Prof. Dr. Danko Nikolić

Chief Scientist

Neuroscientist, formerly at the Max Planck Institute for Brain Research and the Frankfurt Institute for Advanced Studies. Author of the practopoiesis theory and of research on ideasthesia.

Portrait of Vjekoslav Nikolić

Vjekoslav Nikolić

VP Engineering

Software engineer with a background in DevOps. Turns research results into production-ready model code and infrastructure.

Portrait of Simon Reichert

Simon Reichert

GTM & Investor Relations

Built and scaled DXC Robotic Drive from 0 to 1 into a business of more than $300M. At AWS, led the global go-to-market for Accelerated Compute for Automotive and Manufacturing.

Contact

Talk to us

For pilot projects, research collaboration or investment questions, write to us directly.

Davor Andricdavor@robotsgomental.com
Simon Reichertsimon@robotsgomental.com