We are proud to introduce BaseModel.AI ➡️ www.BaseModel.AI #basemodel #behavioral #data
With immense pride, I wish to announce BaseModel.AI 💥 - the world’s most powerful private foundation model for #behavioral #data. Basemodel.ai was created based on many years of research and experience related to the development of the Synerise platform, responsible for billions of AI #decisions made daily. Foundation models like ChatGPT, GPT-4, Dall-E 2 and StableDiffusion have revolutionized Text and Image processing. A single large model trained on massive datasets can replace thousands of specialized models. BaseModel.AI uniquely applies the same principles to behavioral data! From the beginning of Synerise, we aimed for the future. Our bold, ambitious vision was considered a futuristic mirage by many. Our stubborn adherence to the scientific method and breaking old paradigms was considered overambitious by others. AI was dismissed as science fiction not long ago, but today is on the lips of the entire world. We would also like to encourage you to read our newest AI paper: https://lnkd.in/dPiqUZ9c - built upon a great benchmark developed by ML superstars: https://lnkd.in/dBRMFPcZ, called "Temporal Graph Benchmark" with very large datasets and many models tested. Temporal graphs are a hot topic in ML, because it's the most abundant data in commercial settings (almost all kinds of relational databases with timestamps contain temporal graphs) and none of the text or image processing ML techniques can be applied to them. The domain is heavily researched by world leading institutions, such as: #Google #DeepMind, #Meta FAIR, #Baidu Research, #Stanford University, #Oxford University, #Berkeley AI Research, MILA - #Quebec AI Institute, which often have dedicated groups working exclusively on graph machine learning. We make some improvements to the state-of-the-art, inter alia: 1. We show how many existing models for temporal graphs can be outperformed by a simple baseline, taking <10min on 1 CPU instead of 5 days on a powerful GPU cluster. 2. We highlight a fundamental problem of contrastively-trained temporal graph models - their inability to learn global dynamics. 3. We improve existing benchmarking metrics and bring them closer to real-world usefulness. 4. We propose improved contrastive training protocols for temporal graph models. 5. We show that paradigm-breaking science behind BaseModel.AI can outperform existing models by a wide margin. Read more: https://www.basemodel.ai