Add without starting over
Connect a new specialist, tool, sense or memory without rebuilding everything that already works.
ERAIS / Efficient Routed Adaptive Inference System
Not a language model. A cognitive architecture.
ERAIS is a native-byte, modular adaptive AI architecture. It is designed to add language, vision, audio, reasoning and memory as composable expertise, then activate only the parts each task needs.
Use the right tools and the right amount of effort for the task. Leave the rest idle.
More capability. Less active compute.
See what has runWhy it matters
Make capable AI practical in more places.
If each task can use only the parts it needs, increasingly capable AI does not have to remain tied to the largest datacentres.
Bring AI to the work
Move capable systems from datacentres toward workstations, field equipment and edge devices.
Build better machines for it
Eventually design hardware around the work this architecture actually performs—not brute force alone.
Change where intelligence can exist. That is the destination. The experiments decide how far ERAIS earns the right to go.
What makes it different
Learn something new. Keep what already works.
ERAIS can add an audio, vision or specialist skill as a separate candidate. Use it only when needed. Check the accepted language system directly.
ERAIS learned from real audio and labelled video while the frozen language core—and the capability not being trained—stayed exact in the declared test. Learn something new. Keep the accepted core stable. Retain the complete result.
See how this was testedBrute-force changeERAIS approach
Retrain the shared model
Attach one new skillTest everything later
Check the accepted core nowHunt through scattered logs
Keep the change, result and failure connectedRepeat the whole experiment
Fix or restore the named partHas anything real run?
Yes. Here is why we are excited.
A native-byte model has trained at 87.3M parameters, Fracture has reproduced a learned speech projection exactly at full expert coverage, and 1080p vision has traversed the shared world path.
3.40×descriptive same-host throughput ratio
68.45% lessCPU-package energy per reported valid target
An earlier covered-bank component configuration, compared with its previous implementation rather than a dense-model baseline. It does not establish complete-bank quality, end-to-end generation speed, held-out quality, or current serving savings.
Fracture component: less execution overhead
Selected Gemma MLP execution measured 1.982× the speed of its previous implementation on 48 retained rows, with unchanged output for those rows.
Cloud CPU choice: a faster exact continuation
Two cloud CPU configurations resumed the same retained 3M checkpoint and completed the same 27-update interval. The faster configuration measured 1.390163× the training throughput.
Local CPU: a clear small-scale systems signal
At TinyStories paper-text scale on the same CPU host, the corrected ERAIS full-stack window measured 3.40× the dense reference throughput and 68.45% lower CPU-package energy per reported valid target.
Cloud CPU: sustained full-stack continuation
A cloud CPU continuation completed 16.79M valid targets with exact terminal state, sustained throughput, zero swap events, and high use of its declared thread budget.
Cloud GPU: full-stack execution at higher throughput
The repaired TinyStories full stack completed 205 updates on a T4 with bounded quality movement, active cognitive paths, compact sparse execution, and an exact-resume checkpoint.
Cloud GPU: materially more memory headroom
A governed T4 training intervention reduced peak accelerator memory by 45.13% while retaining more than 100,000 valid targets/s and terminal quality inside the registered window.
Stored scale: feasibility, headroom, and the next bottleneck
At 1.61B stored-parameter scale, a larger-batch A10G arm completed ten finite updates, exercised the full governed stack, retained exact restartability, and exposed GPU occupancy as the next optimisation target.
Native + Fracture + worldTrain from raw bytes. Reorganise learned capability into experts. Bring language, speech and vision into one adaptive system.
The three execution paths are now real. The next sprint joins and scales them.
See the architecture move, not just the thesis.
The public demo connects measured Fracture, native-byte language, sparse vision and symbolic reasoning results without exposing the proprietary engine.
Explore the ERAIS demosCurrent native programme
Now the hard question: what happens as ERAIS grows?
The current adaptive raw-byte architecture has trained cleanly at 87.3M parameters with sparse K1–2 routing active. The programme is now driving longer matched runs to discover where quality and active-compute advantages emerge.
87.3M parametersnative adaptive byte architecture
2,048 updatesclean nontrivial-scale training run
3.4964 BPBterminal source measurement
K1–2 activesparse routing stayed operational
Does the learning hold?
Can the richer system improve on a fixed test—not only the checks that change during training?
Can it compress safely?
Can routine input become less work without losing meaning, order or useful surprises?
Can it know more without doing more?
Can stored capability grow while the work performed for one task stays under control?
Does it survive a fair comparison?
Does any advantage remain against a conventional model under the same declared rules?
The next curve changes the scale of the opportunity.
If quality keeps rising while active work stays selective, ERAIS can carry more capability into workstations, products and edge systems that cannot afford brute-force execution.
One architecture / many ways to work
Real work is read, heard, seen and done in motion.
ERAIS is being built so language, voice, vision, memory and action share one evolving model of the world instead of living in disconnected products.
Private work on a workstation
Read procedures, compare evidence and prepare a decision without sending every task to a giant datacentre.
Hands-busy field support
Hear a spoken note or an alarm, then bring in only the knowledge needed for that moment.
Visual inspection
See an asset or defect, compare the relevant evidence and leave unrelated parts of the system idle.
Observatory / how the work stays honest
Find the change. Fix the part. Keep the rest.
The ERAIS Observatory keeps each result connected to what changed, what ran, what failed and what happened next.
That means less time reconstructing an experiment from scattered logs. A result can be explained, repeated, repaired or rejected from the record that produced it.
See how NeuroForge tests its own workResult traceConnected
- Capability
- Real audio + labelled video
- Core check
- Language core + other capability exact
- Evidence
- Metrics · artifacts · failures
- Decision
- Keep · repair · restore
01Isolate
Keep the accepted system still.
02Observe
Record what actually ran.
03Explain
Find the part that changed.
04Decide
Keep, repair or restore.
Fracture / the migration bridge
Keep what today's models have already learned.
Fracture converts learned model components into modular ERAIS expert structures. It builds on the knowledge already captured in existing models.
Executed speech and language conversions demonstrate the path. The destination is the full adaptive native architecture: composable experts, shared cognition and continued learning. Reconstruction measurements establish a baseline; recovery and adaptive execution develop the resulting capability.
Donorstart with capability already learned
Fracturereorganise it into modular expert structure
Recover + adaptrestore quality and continue native ERAIS learning
Scale nextcarry the result across complete models and modalities
Build the next cognitive architecture
Help NeuroForge prove how large this can become.
We are looking for aligned investors, research collaborators and design partners to fund the experiments that matter and turn a founder-scale programme into an amplified team.