Archive — US — turtlesai
Cogito v1: The Open Source Self-Improving AI
Deep Cogito has launched Cogito v1, a set of open-source large language models (LLMs) ranging from 3B to 70B parameters, trained using the Iterated Distillation and Amplification (IDA) technique, which enhances self-improvement. These models outperform existing models like LLaMA and are available on platforms such as Huggingface. Future models up to 671B parameters are planned, continuing the push towards general superintelligence.
Key points
- Deep Cogito released Cogito v1, consisting of open-source LLMs with parameters from 3B to 70B.
- The training utilized the Iterated Distillation and Amplification (IDA) technique, promoting iterative self-improvement.
- Cogito v1 models excel in standard benchmarks when compared to other models like LLaMA, DeepSeek, and Qwen.
- Future plans include larger models up to 671B parameters and improved versions of existing models.
- The models are usable on platforms like Huggingface, Ollama, Fireworks AI, and Together AI.
Why it matters
- The release of Cogito v1 aligns with trends in AI research focusing on self-improvement and robustness in models.
- IDA techniques are gaining attention for their potential to enhance model capabilities beyond human limitations.
- Previous AI advancements have paved the way for more effective training methods that facilitate model alignment with user objectives.
The launch of Cogito v1 signifies a notable advancement in AI technology, promoting self-improvement through innovative training methods, contributing to open-source efforts, and setting the stage for even more powerful models in the pursuit of general superintelligence.