New Physics-Based Decision Model Released via Open-Source Community
A new physics-based typed decision model, named Vega, has been released to the open-source community. Developed by Nandhakishor, the model utilizes a Test-Time Training (TTT) architecture implemented through an engine-and-adapter framework. According to documentation provided by the developer on Reddit and GitHub, the model is available in two sizes: an 800-million parameter version and a 4-billion parameter version.
The architecture employs a unique decision-making process modeled after physical movement. According to the developer’s technical blog, the model converts inputs—such as observations and questions—into vectors extracted from the hidden layers of a Large Language Model (LLM) or Vision-Language Model (VLM). These vectors are used to create a “landscape” of potential outcomes represented as valleys. The model then simulates a “ball” moving through this landscape, influenced by friction, which eventually settles into the valley representing the most likely correct answer.
Performance Benchmarks and Technical Specifications
The Vega model supports a context window of 73,000 tokens and includes image support. In unverified performance comparisons shared by the developer on Reddit, the 800-million parameter version demonstrated a latency of approximately 17ms on a GPU, while the 4-billion parameter version recorded approximately 36ms on a GPU.
In a comparative test involving 20 cases, the developer reported that the Vega 4B model achieved 17/20 correct answers, while the 0.8B model achieved 15/20. For comparison, the developer noted that the Plumb-4B model, running on eight CPU cores without a GPU, achieved 19/20 correct answers but with a significantly higher latency of approximately 0.9 seconds per question. The developer claims that Vega is capable of answering a calibrated “won’t answer” flag, a feature the Plumb-4B model reportedly lacks.
Community feedback on Reddit suggests that such models could serve as a cost-effective alternative to calling full-scale LLMs for tasks such as topic classification.
Broader Context of AI in Scientific Disciplines
The release of specialized decision models occurs as AI integration expands into complex scientific fields. While some debate continues regarding the necessity of true superintelligence for industrial “flywheels” to begin, current applications are already visible in biotechnology and physics.
In the medical sector, several organizations are utilizing AI frameworks for clinical and research purposes. Merck and Moderna are currently in Phase 3 of an AI-assisted cancer trial. Additionally, researchers at Stanford have developed an AI biotech company that designed a lung cancer drug which recently received FDA therapy designation. Other developments include Utah researchers using AI to analyze molecular features for cancer treatment predictions, and Dutch scientists utilizing the PARM model to predict how non-coding DNA mutations affect cancer. These developments suggest a growing trend of using AI as a verification and simulation engine for complex biological and physical systems.