He left the search box that answers a billion questions a day to build something stranger: models that learn to trust their own gut.
There is a switch inside the models Drishan Arora builds. Flip it one way and the AI answers immediately, the way you blurt out a capital city. Flip it the other and the same model stops, thinks in the open, and reasons its way to a conclusion. Two minds in one set of weights. That toggle is the small, concrete thing that explains the large, abstract thing Arora is actually after.
He is the co-founder and CEO of Deep Cogito, a San Francisco lab he started in June 2024 with Dhruv Malhotra, another Google alum. The company's stated job is almost comically large: build general superintelligence. The way they talk about it is the interesting part. Not a moonshot. Not a prophecy. A problem. "Building superintelligence," Arora says, "is fundamentally a tractable machine learning problem." Tractable is an engineer's word. It means: solvable, with the tools on the bench.
Before any of this, Arora had a comfortable, enviable seat. He was a senior software engineer at Google, a tech lead on the question answering team behind Google Search and Google Assistant, and he later led large language model work for Google's generative search. When you typed a question and Google handed back a clean, confident sentence instead of ten blue links, that was the machinery he tended. Few engineers get closer to the place where the entire world asks its questions.
He left it to argue with the field's central habit.
The reigning recipe for smarter AI has been blunt: let the model search more. Generate more candidate thoughts, branch further, burn more compute at the moment of answering, and the scores go up. It works. Arora's quarrel is that it papers over something. The model gets better at looking for the answer without getting better at sensing which direction the answer lives in. The intuition stays flat while the flailing gets more expensive.
His counter-move is a method with a mouthful of a name: iterated distillation and amplification, or IDA. First, let the model reason hard and search for a good solution - the amplification. Then take that hard-won reasoning and fold it back into the model's own parameters, so next time the model simply knows - the distillation. Do it again. And again. Each loop, the model's instinct sharpens. "The model should be able to directly guess the results of running reasoning," Arora puts it, "without actually doing so." A chess master who no longer calculates every line because the good move already looks good.
The proof he points to is efficiency, not just accuracy. Deep Cogito's flagship reaches results comparable to heavyweight rivals while spending roughly 60% shorter reasoning chains. Less thrashing, better aim.
There is a second choice baked into all of this, and it is louder than it looks. When Deep Cogito ships a model, it ships the weights. The 3-billion-parameter version and the 671-billion-parameter version both go out the door for anyone to download, run, fine-tune, or pull apart. In a field where the strongest systems increasingly sit behind paid interfaces and guarded APIs, that is a deliberate stance. Arora's company emerged from stealth not with a demo and a waitlist, but with models that landed near the top of open leaderboards on day one and a research note explaining exactly how they were made.
The wager is that openness compounds. Hand the method to the world and the world stress-tests it, extends it, and finds the failure modes faster than any single lab could. For a company built on a self-improvement loop, betting on a second loop - the one that runs through the open-source community - is a consistent kind of faith.
The timing helped. Deep Cogito built its first family on open foundations like Meta's Llama and Alibaba's Qwen, then layered IDA on top, which is part of how a two-person-led team got to frontier-competitive results in about 75 days. Standing on open shoulders, then handing the new height back. That is the rhythm.
We are building general superintelligence. Achieving this requires scientific breakthroughs - such as advanced reasoning and iterative self-improvement.
Deep Cogito, mission statementThe model spends compute thinking out loud, hunting for a stronger solution.
That reasoning path is folded into the model's own weights as a sharper prior.
Next time, the model guesses the good answer faster, with less searching.
Deep Cogito's headline claim is not only that the model is good, but that it gets there using markedly shorter reasoning chains than comparable open models - a stand-in for the intuition it has internalized. Illustrative comparison of relative reasoning-chain length.
~60% shorter chains while matching comparable results. Source: Deep Cogito v2 research notes.
A B.Tech in electrical engineering before the pivot into machine learning and language.
A master's in computer science - and a teaching assistant for five separate courses across two years.
Senior engineer and tech lead on question answering for Google Search and Assistant; later led LLM work for generative search.
Co-founds the company in San Francisco with ex-DeepMind product lead Dhruv Malhotra.
Cogito v1 hybrid reasoning models (3B-70B) ship and top open-model charts - trained in roughly 75 days.
Cogito v2 preview: four open models up to 671B, built on IDA and self-improving intuition.
A $13M seed round led by Benchmark; South Park Commons among the early backers.
Cogito v2.1 671B arrives, billed as the best open-weight LLM by a US company.
Two people. Seventy-five days. A stack of open models that landed near the top of the charts on arrival. The Deep Cogito origin reads less like a typical AI startup and more like a sprint run by people who already knew exactly where the bottleneck was.
The name is a tell. Cogito - Latin, "I think." Descartes' one load-bearing certainty, borrowed by a company whose entire product is the act of thinking, and the act of getting better at it.
Arora is not loud about himself. He lets the weights argue. Where many frontier labs treat their best models as crown jewels to be locked behind an API, Deep Cogito keeps shipping its strongest work as open weights for anyone to download, poke, and build on. That is a position as much as a product decision - a wager that openness, not secrecy, is the faster road to the thing they want.
His co-founder Malhotra came from Google DeepMind, where he worked on generative search - the same neighborhood of the map Arora knew from the Google side. Two people who had seen, up close, how the world's questions get answered, deciding the answering could be done differently.
The investors followed the thesis, not the hype. Benchmark, the firm with a long history of early, concentrated bets, led the $13M seed. South Park Commons, the community-and-fund built for founders at exactly this stage - the blank-page moment before a company has a shape - was there among the early backers. Neither is known for chasing crowds.
Trace the path and one habit repeats. Electrical engineering, then a hard turn into language and learning. A seat at the center of Google's question machine, then a hard turn into building a rival approach. The default recipe of more search, then a hard turn into intuition. Arora keeps walking up to the comfortable consensus and asking whether the obvious next step is actually the right one. Deep Cogito is what that question looks like when you give it a building and a few hundred million parameters to argue with.
A mixture-of-experts model billed as the strongest open-weight LLM from a US company, competitive with frontier closed and open systems.
The first Cogito models, frontier-competitive for their size, trained in roughly 75 days with a small team.
Turned iterated distillation and amplification from idea into a working pipeline for model self-improvement.
Hybrid models that answer instantly or stop to reason - instinct and deliberation in one set of weights.
Led question answering for Google Search and Assistant, then LLM work for Google's generative search.
A $13M seed led by Benchmark, with South Park Commons among early supporters.
Building superintelligence is fundamentally a tractable machine learning problem.
The model should be able to directly guess the results of running reasoning - without actually doing so.
We are building general superintelligence. Achieving this requires scientific breakthroughs - advanced reasoning and iterative self-improvement.
He trained as an electrical engineer at IIT Delhi before going all-in on machine learning and natural language.
At Columbia he was a teaching assistant for five different courses in two years - a lot of whiteboards.
The company name, Cogito, is the back half of Descartes' "I think, therefore I am."
Deep Cogito gives away its strongest models as open weights, while many rivals lock theirs behind APIs.
The team's analogy for self-improvement is AlphaGo - a system that got better by learning from its own play.
The goal is not a cleverer chatbot. It is a model that develops genuine intuition, improves itself loop by loop, and arrives at general superintelligence - released in the open, for anyone to build on, rather than locked away.