← Back to Home

💚 Support Independent AI Research

Help fund breakthrough research on shoestring budgets

Why Your Support Matters

Grey Liquid Lab and G4 Turbo development are conducted entirely independently—no corporate backing, no grants, no institutional funding. Just systematic testing, transparent documentation, and a mission to push AI boundaries on budgets measured in tens of dollars, not millions.

17,000+ Users

benefit from stable G4 Turbo/Nano models and transparent research findings. Your support helps me focus full-time on AI research instead of DoorDash.

❤️ GitHub Sponsors ☕ Ko-fi 💵 CashApp: $ssfdre38

Impact by the Numbers

21K+
Total Downloads (Ollama + HF)
100%
Organic Growth (Zero Ads)
4
Published Experiments
$100
Typical Research Budget

What You're Funding

  • 🔬 Grey Liquid Lab: Fundamental AI systems research exploring model compression, memory architectures, autonomy, and infrastructure. Currently investigating the 3-bit quantization barrier.
  • 🚀 G4 Turbo Family: Production-stable Gemma 4 models (Turbo, Nano, Mini) serving 17K+ users. Sub-1GB breakthrough models that run on phones, Raspberry Pis, and laptops.
  • 🤖 Ash & Autonomy Research: Democratized AI systems where everyone can run their own autonomous agent. Exploring AI agency, emotional processing, and self-aware architectures.
  • 🛠️ Open-Source Tools: forge-train, ash.cpp, OpenClaw CE, and other infrastructure projects that make advanced AI accessible to everyone.
  • 📖 Transparent Documentation: All findings published openly—successes and failures. If an experiment breaks, we document exactly how and why. No paywalls, no gatekeeping.

Research Philosophy

Compete with billion-dollar labs on $100 budgets. DeepMind and OpenAI have massive resources—we have creativity, transparency, and a willingness to test unconventional ideas. Grey Liquid experiments cost server rental for a weekend, not millions in compute.

Document everything. Every experiment gets a detailed report—whether it succeeds or fails. The community learns from both. Failure modes teach more than successes.

Democratize access. If advanced AI only runs on $10,000 GPUs, it's not really accessible. G4 Turbo/Nano prove that powerful models can run on phones and single-board computers. Everyone should be able to participate in the AI revolution.

"DeepMind might want to look out—I might make the next breakthrough with $100 worth of renting servers and make the next generation of AI systems."

Current Research Status

Grey Liquid Lab - Q1 2026

Investigating the 3-bit quantization barrier that prevents models from being compressed below ~3.4 bits per weight. Four experiments completed:

  • Experiment #1: Q2_K quantization (2.96 bpw) - compiler alignment failure
  • Experiment #2: Importance matrix analysis - incomplete tensor coverage discovered
  • Experiment #3: GPTQ comparison - confirmed universal 3-bit floor across all methods
  • Experiment #4: Mixed-precision bypass - precision distribution matters more than average

Next: Test non-PLE architectures (Llama 3.3, Phi-4) to isolate whether Gemma 4's PLE pathway is the specific bottleneck, or if this is a universal mathematical floor.

G4 Turbo Production Status

Three stable models in production serving 17K+ users:

Ready to Support?

Choose your preferred platform:

❤️ GitHub Sponsors ☕ Ko-fi 💵 CashApp: $ssfdre38

Questions? Reach out on GitHub or Ollama.