Deep Cogito, a San Francisco-based AI startup, formally launched its suite of hybrid AI models to a global audience on June 1, 2024. Company co-founders Drishan Arora and Dhruv Malhotra started the company to transform artificial intelligence. Specifically, they’re interested in building AI systems whose reasoning and non-reasoning capabilities can work fluidly together. This ground-breaking approach includes a unique fine-tuning method that trains the AI to carry out tasks with greater cognitive function.
The Cogito 1 family of models provides a robust arsenal of sizes. They began with 3 billion parameters and scale up to a remarkable 70 billion parameters. In the next few weeks and months, Deep Cogito will release longer versions of these models. These upgraded versions will allegedly be their 671 billion parameter models. The firm utilizes Meta’s open LLama model as well as Alibaba’s Qwen model in its technology. This strong financial undercurrent powers their aggressive moves into AI.
With Deep Cogito’s models, they can really stand apart in the market. They are consistently the best open models of similar size, even models produced by industry titans such as Meta and DeepSeek. According to the company, one of the most impressive features across each model is its ability to customize conversational response modes based on the task required.
“Each model can answer directly [… ] or self-reflect before answering (like reasoning models),” – Deep Cogito
This ability to do much more than generate simple answers is a huge boon in an age where generative AI applications require so much more. We’ve made both models freely available for download. They are freely available for adaptation through APIs on cloud platforms such as Fireworks AI and Together AI.
Malhotra provides the company with deep experience in generative search technology. He was formerly a product manager at Google’s AI lab, DeepMind. He imagines Deep Cogito to be a big player in the search for “general superintelligence.” This radical idea has captivated the tech community and inspired endless speculation.
South Park Commons is an early investor in the company. This partnership brings both credibility and resources to bear on Deep Cogito’s ambitious projects. As they keep scaling up their work, they definitely have an eye toward refining their models further, and so forth into other methodologies.
“Currently, we’re still in the early stages of [our] scaling curve, having used only a fraction of compute typically reserved for traditional large language model post/continued training,” – Deep Cogito
What’s most exciting here, perhaps, is the new research into complementary post-training approaches for self-improvement. This new initiative is another demonstration of Deep Cogito’s commitment to pushing the envelope with AI. It is instrumental in keeping their models at the forefront of what’s technologically possible.
