Deep Cogito Raises $43 Million To Build AI Models That Improve Themselves
Right now, most AI progress comes from “pre-training.” This involves feeding models massive amounts of internet data. At least, that’s how ChatGPT and Gemini became operative. However, it’s expensive and The post Deep Cogito Raises $43 Million To Build AI Models That Improve Themselves appeared first on Ventureburn .
Deep Cogito, an AI lab founded in 2024 by former Google AI Search leaders Drishan Arora and Dhruv Malrana, has secured $43 million in Series A funding led by TQ Ventures. The round was joined by Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler, bringing total funding to over $56 million. The company's mission is to build AI models that can improve themselves, a process they call "post-training."
Deep Cogito's approach is different from the conventional method of building larger base models from scratch. Instead, they focus on developing the engine that enhances existing models' capabilities. Their two core techniques are Iterated Distillation and Amplification (IDA) and Process Supervision + Reinforcement Learning. IDA involves giving a model a prompt, using more computing power to think harder, and then distilling the improvement back into the model's weights. This process is repeated, allowing the model to learn and get better without new human data.
The other technique, Process Supervision + Reinforcement Learning, grades every step a model takes to reach a final answer. This helps the model avoid wasted or incorrect steps, leading to better reasoning and lower hardware costs. Deep Cogito has already released results, with their latest model, Cogito v2.1 671B, debuting in November 2025. They claim it outperformed other US open models of the time while using fewer tokens, making it cheaper to run.
The new funding will be used to scale Deep Cogito's post-training engine, focusing on two areas: the Cogito family of open-weight frontier models that any company can run and modify, and enterprise models trained on a company's proprietary data. Deep Cogito's long-term goal is recursive self-improvement, where models progressively get smarter on their own and surpass the limits of human training data.
This could significantly impact the AI landscape, giving companies control over their AI and reducing costs for businesses like banks, healthcare, and manufacturers. For the US AI ecosystem, Deep Cogito could help bridge the gap with China, which currently leads in open models.
Written by urgent.news from Ventureburn's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
