Cogito Preview: IDA as a path to general superintelligence
a year ago
- #Superintelligence
- #LLM
- #AI
- Cogito releases open-license LLMs (3B, 8B, 14B, 32B, 70B) outperforming competitors like LLaMA and Qwen on benchmarks.
- Models use Iterated Distillation and Amplification (IDA) for scalable alignment and self-improvement toward superintelligence.
- Each model offers standard and self-reflective (reasoning) response modes.
- Upcoming releases include larger models (109B, 400B, 671B) and improved checkpoints.
- IDA combines advanced reasoning with iterative self-improvement, surpassing overseer limitations.
- Amplification uses computation to enhance intelligence; distillation internalizes improvements into model parameters.
- Cogito’s 70B model outperforms larger models like Llama 4 109B MoE.
- Models optimized for coding, function calling, and agentic tasks, with shorter reasoning chains.
- Benchmarks show IDA’s effectiveness, though real-world performance may vary.
- Deep Cogito aims for general superintelligence via scientific breakthroughs and top-tier research.