Deep Cogito

Deep Cogito is an AI research startup dedicated to developing open-source general superintelligence. Founded in June 2024 by Dhruv Malhotra, a former Google DeepMind product manager, and Drishan Arora, a former Google engineer, the San Francisco-based company aims to create AI that surpasses human capabilities. The company is backed by investors including South Park Commons and Benchmark, having secured $13 million in a seed funding round.

Deep Cogito's core product is its family of Cogito AI models, which are instruction-tuned generative models released under an open license for commercial use. A key feature of these models is their hybrid reasoning capability, allowing them to switch between a fast, standard mode for quick responses and a deep-reasoning mode for complex problem-solving. This approach addresses the common trade-off between the high computational cost of reasoning models and the efficiency of standard AI. The models are built upon established architectures like Meta's Llama and Alibaba's Qwen.

The company utilizes a training technique called Iterated Distillation and Amplification (IDA), where the model iteratively refines its understanding of a task, stores the improved reasoning within its own weights, and thereby builds a stronger intuition. This method of self-improvement reduces the reliance on human feedback and allows for faster, more cost-effective model development. The models have demonstrated strong performance on industry benchmarks and exhibit emergent abilities, such as reasoning about images without explicit training. Deep Cogito makes its models available for download and through cloud APIs, providing flexibility for developers.

Keywords: general superintelligence, open source AI, artificial intelligence, large language models, hybrid reasoning, machine learning, Iterated Distillation and Amplification, generative models, deep reasoning, AI research, Drishan Arora, Dhruv Malhotra, self-improving intelligence, AI models