Research & Development
Exploring What AI Can Do Next
Building smarter, safer AI
At Yalla2x, we explore how AI can learn, improve, and solve real-world problems. Our focus is on building reliable, secure, and practical AI solutions that help businesses work smarter while keeping humans in control.
Research direction
Built to use recursive self-improvement systems
Our lab investigates how AI systems can iteratively improve their own reasoning, coding, and research capabilities while remaining measurable, safe, and human-governed. We focus on real world evaluation loops rather than purely theoretical takeoff scenarios.
-
Methods and evaluation
Loop
Closed feedback cycles where models propose, test, and refine better versions of themselves
Evaluation
Stable benchmarks and external verifiers to measure whether each change is a real improvement
-
Safety and governance
Safety
Guardrails for reward hacking, model collapse, and misaligned self-modification behaviors
Oversight
Human review stages for high-impact changes, critical decisions, and deployment boundaries
Publications & Notes
Recent work from the lab
- Sealed
- an app-level seal: built and passed its own checks; not audited by a third party
- In progress
- research under way, no final result yet
-
RSIS
RSIS is our approach to building practical, trustworthy AI that runs under the owner's control.
-
ARC-AGI-3: Measuring AI Learning
We test how AI learns and adapts to unfamiliar challenges across 25 environments, measuring progress in reasoning and problem-solving.
Research highlights: Level 6 reached in one environment · 149 recorded winning runs
-
The Riemann Hypothesis
Exploring one of mathematics' greatest unsolved problems through AI-assisted reasoning and formal verification.
Research highlights: 7 of 14 project targets reported as met · The hypothesis remains unsolved
-
AI - Sahih Verification Layer
Sahih Verify checks citations in AI-assisted legal documents, identifying unsupported claims and producing signed verification records.
Research highlights: 12 application checks passed · Currently tested with sample data; accuracy not yet established
-
Proof of Human: Verifying Identity in the Age of AI
Proof of Human is designed to record evidence of human presence during content creation through liveness challenges and signed records.
Research highlights: 12 application checks passed · Liveness accuracy not yet measured
-
Artificial Superintelligence: Exploring Self-Improving AI
We are exploring how AI systems can improve through testing, comparing new models with existing ones before adopting changes. This is not yet artificial superintelligence.
Research highlights: Previous evaluation score improved from 90 to 97 · Evaluation loop currently under investigation