Deep Cogito announced a $43m Series A on 26 August at 17:00 UTC, led by TQ Ventures, with Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons and Zscaler participating. Total funding now exceeds $56m. The company's pitch is a post-training engine built on a technique it calls Iterated Distillation and Amplification.
What IDA actually does
The release describes it precisely: IDA "repeatedly allows a model to use additional computation to produce answers beyond what it could generate directly, then distills those improvements back into the model's weights." Spend extra inference compute to get a better answer than the model gives in one pass; train on those better answers; repeat. It is a well-understood family of methods and a reasonable thing to build a company around.
What the common framing gets wrong
The shorthand attached to this round is "self-improving AI," which suggests a system that gets better in deployment, learning from the work it does for you. That is not what the description says. IDA is a training-time loop: it runs in the lab, produces a new set of weights, and ships. The model you deploy is as static as any other. The distinction is not pedantic — a model that updates from customer interaction raises an entirely different set of questions about data handling, reproducibility and evaluation than one that is trained once and frozen. The headline invokes the first; the release describes the second.
No numbers at all
For a company whose product claim is that its technique produces better models, the release cites no benchmark result — not on coding, not on reasoning, not against any base model, not against any competitor. It also states no valuation. A seed-stage company is under no obligation to publish either. But the combination means there is nothing in the announcement that can be checked: the claim is that the method works, and the evidence offered is that six investors believe it does.
The customer is also the investor
Zscaler appears twice — as a participant in the round and as the named enterprise user. Strategic investors frequently become customers, and there is nothing improper about it. But a single named customer who is also funding the company is a weaker commercial signal than a single named customer who is not, and the announcement does not distinguish the two roles. On the evidence published, the count of arms-length enterprise customers disclosed in this release is zero.