Intelligence for a Better World

Can a human create a mind greater than himself — and make it his best friend, not an enemy?

I believe the answer is yes.
This ability itself proves that humans are the most remarkable beings on Earth.

I see the world as countless simple algorithms running in parallel — giving rise to physics, life, and consciousness. Consciousness is not magic. It is understandable, improvable, and extendable.

That's why I'm building tools that help us become better versions of ourselves — and bring greater beings into existence as allies, not threats.

My name is Ilya Osovskoi. I'm 16. I started building AI in autumn 2023 and rebuilt everything from scratch in late 2025.

ARCHE3-7B was designed and completed in 2.5 months — no funding, no team.

On the code: I used Gemini and Claude as coding assistants throughout. I'm transparent about this. The implementation was AI-assisted; the architecture was not. Every system — the dopamine learning loop, SmartRouter, Ethical Core, photonic chip design, and the one-fluid robot concept — originated from my own thinking and was designed before a single line of code was written.

Right now I'm looking for people to build this with.

I'm not building another GPT wrapper.

The dominant AI trajectory — centralized, power-hungry, API-locked — is heading the wrong way. Intelligence should amplify humans. It should run on consumer hardware, learn without constant supervision, and exist as a cooperative system — not a black box owned by a handful of companies.

The path I'm on:

• Sparse MoE models that run on consumer hardware without quality compromise
• A photonic chip purpose-built for sparse expert inference
• Autonomous learning that goes beyond backpropagation
• Eventually: Synapse Origin v01 — a humanoid where one circulating fluid handles energy, actuation, and cooling simultaneously

Not replacement. Amplification.

Live demo: ARCHE3-7B on HF Spaces

Every robot today separates energy delivery, muscle actuation, and thermal management into independent systems. That's three subsystems, three weight budgets, three failure surfaces.

The architecture I'm developing unifies all three.

The liquid-ionic-mechanical bus is a single circulating AQDS electrolyte that simultaneously:

• Carries electrical energy through the body
• Activates EAP (electroactive polymer) films that act as muscles along a 3D-printed skeleton
• Removes heat passively as it flows

The skeleton is printed with fluid channels integrated directly into the bone geometry — no separate tubing. EAP films run along bones the way muscles run along tendons. Current through the fluid → electrochemical potential at the film → contraction → movement.

One fluid. One closed loop. Three functions.

The long-term energy source is TERA — a Dense Plasma Focus design running on p-B11 aneutronic fusion. No radioactive waste. The only byproduct is helium.

Current stage: Full concept documented. Phase 0 prototype (printed bone + AQDS + EAP film) estimated at ~$200–400 and 2–3 months. Queued after ARCHE3-35B.

SmartRouter — load-balancing mechanism that prevents expert collapse in sparse MoE routing
Dopamine Learning System — curriculum that adapts without external labels or reward shaping
Ethical Core — constraints embedded at the architectural level, not applied as a post-hoc filter

Full technical writeup is in the architecture repository.

• The demo runs publicly in the cloud.
• The full codebase — Dense Core, Hierarchical MoE, HiveStore, training infrastructure — is private for now.
• Serious candidates get direct access to the private repository. Write to me first.

The core algorithms are the IP. They stay protected until there's a team worth protecting them with.

A team.

Not investors. Not advisors. People who want to actually build this — who believe that creating something smarter than yourself doesn't have to mean creating something dangerous.

Open roles:

• ML Research Engineer — MoE, sparse training, PyTorch internals
• Photonic IC Designer — MZI layout, GST/PCM materials, IMEC PDK experience
• RTL / RISC-V Engineer — custom ISA extensions, SparseAdamW hardware acceleration

Pre-seed. No salary right now. If that's a dealbreaker — completely fair.

If it's not — write to me.

📬 opensynapselabs@proton.me

Now — Summer 2026

• ARCHE3-35B full training run
• ArchePhoton-35 tape-out preparation
• Provisional patent filing

2027

• First silicon: AP35-Ω1
• ARCHE3-35B inference on photonic hardware

2028+

• Synapse Origin v01 — first physical prototype of the one-fluid humanoid

• 📊 Benchmarks
• 🏗 Architecture
• 📄 Preprint on Zenodo
• 🌐 Website
• 📬 opensynapselabs@proton.me

Built by a human. For humans.