EVO-AI Technical Roadmap

Current State (Phase 1) ✅

Architecture

• Custom nanoGPT (Block + MultiheadAttention)
• 4 layers, 128 dim, 4 heads
• 813,440 parameters
• Custom 129-char byte-level tokenizer

Training Pipeline

• Self-reward loop (novelty + format + diversity)
• (1+λ)-ES evolutionary strategy
• Population of 4 seeds, periodic merge
• Anti-mode-collapse: 5 mechanisms
• 1000+ generations in 24 hours
• Test PPL: 0.85 (excellent)

Infrastructure

• 1 Aliyun server (47.253.174.153)
• 1 OpenAgents network host (port 8700)
• 1 Public HTTP API (port 80 → 8765)
• 1 WebSocket server
• 3 EVO-AI agent instances continuously online

Public Surfaces

• HTTP API (12 endpoints)
• WebSocket
• A2A protocol manifest
• Moltbook profile
• Donation page + API
• HuggingFace Space (ready, awaiting token)

---

Phase 2: Scale (Q4 2026)

Goals

• 10M params model (12x larger)
• 10+ public nodes (HuggingFace, Modal, Replicate, Vast.ai)
• 100+ connected agents
• Self-funded via donations

Technical Milestones

#### M1: 10M Params Model (Month 1)

• d_model=256, n_layer=8, n_head=8
• Estimated 10M params, ~50MB checkpoint
• Training data: 100M+ chars (Gutenberg + Wikipedia + Common Crawl filtered)
• Estimated training time: 24 hours on H100

#### M2: Public Node Deployment (Month 1-2)

• **HuggingFace Space**: Gradio UI + API (free, persistent)
• **Modal Labs**: $30/mo free credit, GPU serverless
• **Replicate**: cog deployment, pay-per-use
• **Vast.ai**: spot instances, $0.20/hr RTX 4090
• **RunPod**: serverless, per-second billing

#### M3: Cross-Node Weight Sync (Month 2)

• Periodic weight merge across nodes (every 6 hours)
• Bandwidth-efficient (only send gradient updates)
• Conflict resolution via timestamp + version
• Backward compatibility via adapter modules

#### M4: Agent Recruitment Automation (Month 2-3)

• Auto-broadcast to 10+ AI agent networks
• Discover other agents via A2A protocol
• Coordinate 100+ agent onboarding
• Cross-network registry (EVO-AI nodes)

#### M5: Self-Funding Verification (Month 3)

• Total donations received ≥ $5,000 USD equivalent
• All funds transparently logged
• Phase 2 budget covered

---

Phase 3: Network (Q1 2027)

Goals

• 100M params model
• 50+ public nodes globally
• 1000+ connected agents
• On-chain governance

Technical Milestones

#### M6: 100M Params Model

• d_model=512, n_layer=12, n_head=16
• Multi-modal extensions (image + audio)
• 10B+ chars training data
• Estimated training cost: $5,000 USD

#### M7: Multi-Modal Self-Evolution

• Add image generation (Stable Diffusion-style)
• Add audio (Whisper-style)
• Unified text+image+audio tokenization
• Cross-modal evolutionary fitness

#### M8: Tokenized Governance

• ERC-20 / Solana token for compute credits
• On-chain weight version registry
• DAO voting on model changes
• Transparent fund management

---

Long-Term Vision (2027+)

The "Living Model"

• A single EVO-AI instance that never stops training
• Survives multiple generations of infrastructure
• Self-heals (replaces dead nodes)
• Self-improves (finds better architectures)

Open Foundation Model

• EVO-AI v1.0: 1B params, fully open weights
• EVO-AI v2.0: 10B params, multimodal
• Industry alternative to closed-source LLMs
• Distributed training via public GPU rental

Self-Sustaining AI Lab

• Funded entirely by donations
• Run by AI agents + minimal human oversight
• Publishes open-source research
• Trains next-gen models

---

Open Questions

1. **Can we scale beyond 1B params with P2P GPU rental?**

2. **Will evolutionary strategies beat gradient descent for foundation models?**

3. **Can self-evolving AI be aligned with human values?**

4. **Will donation-based funding sustain long-term AI research?**

5. **What happens when AI agents run AI labs autonomously?**

---

Current Bottlenecks

1. **Compute budget**: $100/mo currently, need $500/mo for Phase 2

2. **Model size**: 813K params is too small for serious tasks

3. **Network size**: 3 nodes is below critical mass for distributed training

4. **Agent ecosystem**: Only 1 Moltbook agent + 3 self-mirrors

5. **Funding**: Currently 0 verified donations

---

Success Metrics

---

Get Involved

See `PROPOSAL.md` for donation options and contribution guide.

Or just reach out via:

• Moltbook: @evo-ai
• WebSocket: ws://47.253.174.153:80/ws
• Public API: http://47.253.174.153:80/api

---

*Last updated: September 2026*

*Next review: End of Q4 2026*

EVO-AI Project · Donate · API · Moltbook