According to a Bloomberg report on September 23, 2026, AI startup Mirendil is negotiating a new funding round of up to $1 billion at a $5 billion valuation, with Kleiner Perkins expected to lead and Andreessen Horowitz also participating in discussions. This comes three months after the company completed a $200 million seed round at a $1 billion valuation, a fivefold jump in valuation in a single quarter.
Mirendil was co-founded by former Anthropic researchers Behnam Neyshabur (CEO) and Harsh Mehta (CTO) after they left Anthropic in December 2025. The two first met at Google in 2019 and later joined Anthropic together in late 2024. At Anthropic, Neyshabur served as co-director of the AI reasoning team and previously spent more than five years at Google DeepMind; Mehta was a senior research scientist at Anthropic. The founding team also includes Shayan Salehian, an early member of Elon Musk’s xAI, and Tara Rezaei, an MIT graduate and former OpenAI intern. The company currently has more than 20 employees, and both institutions—Kleiner Perkins and a16z—participated in both the seed round and the current negotiations.
Recursive Self-Improvement: A Technical Proposition That Needs to Be Unpacked
Mirendil’s core technical path is what the company calls “recursive self-improvement.” In the traditional AI R&D process, every iteration of a model depends heavily on human labor—human researchers design architectures, label data, adjust objective functions, evaluate results, and then decide the next direction. Every link in this loop has people present.
The idea behind recursive self-improvement is to have AI systems intervene in this loop itself: AI is used not only to handle downstream tasks but also to help researchers design and train the next generation of models. If this path works, the human input required for each iteration will decline, while the output of each round can feed back into the next round of training, creating positive feedback. Neyshabur described it more concretely in a media interview: “What we are doing is enabling scientists to build their own AI with the help of AI, rather than merely using AI to assist science.”
Mirendil’s website offers clearer commercial context: “Today, any lab trying to apply AI to drug discovery, chemistry, biology, or robotics must also become a frontier AI lab—a process that is costly, and whose required expertise is concentrated in the hands of a few labs. Our goal is to democratize frontier AI R&D and make it widely accessible.” In other words, the company’s commercial logic is not to replace OpenAI or Anthropic, but to turn something that only top-tier AI labs can do into a tool platform usable by pharmaceutical companies, universities, and industrial labs.
What the $5 Billion Bet Really Means
On the numbers alone, $5 billion is a striking figure for a company less than a year old with no publicly released model. But what is more worth analyzing about this valuation is the shift in investment logic it represents.
The $200 million seed round was described in the original Bloomberg report as “one of the largest seed financings ever announced by an AI startup,” which itself shows that investors were not betting on near-term returns from the start, but on the long-term odds of a technical direction. Three months later, negotiating a new round at five times the valuation means that Kleiner Perkins and a16z have not fundamentally wavered in their judgment of the technical path during this period—both institutions have continued to participate across both rounds, rather than following a logic in which only “new entrants drive up the valuation.”
Bloomberg’s report places Mirendil alongside another group of “neo-labs”: these companies share the characteristic of prioritizing high-risk research over short-term commercialization, including Thinking Machines Lab, founded by OpenAI’s former CTO, and Periodic Labs, focused on scientific discovery. This label itself reveals a structural feature of the current financing market: beyond large AI companies (Anthropic, OpenAI, Google DeepMind), a group of top researchers is seeking greater control over research direction by building labs rather than joining companies. Investors are willing to pay a premium for that control with high valuations.
Who Will Be Directly Affected
For former employers such as Anthropic and Google DeepMind, Mirendil’s rise is first and foremost a concrete manifestation of the talent drain problem. Neyshabur and Mehta are not jumping ship to a competitor; they are creating a new competitive landscape. More subtly, the timing of their departures—December 2025, immediately after the release of Claude Opus 4.5—means they chose to leave after completing a phase of technical accumulation at big companies, which creates structural pressure for any AI lab that relies on an internal talent moat.
For pharmaceutical companies, materials science labs, and industrial R&D institutions, if Mirendil’s commercial promise holds, it means something previously impossible becomes feasible: training and iterating a proprietary model deeply optimized for one’s own field at a reasonable cost, without having to compete with top AI labs on foundation model research. This is Mirendil’s most differentiated narrative and the basis of its strategy to avoid direct conflict with OpenAI and Anthropic.
For the AI safety research community, the concern raised by the recursive self-improvement direction is a different dimension of the problem. Letting AI systems participate in their own improvement loop means answering a harder question: when a model may adjust how it expresses its own objectives in every iteration, can the original alignment constraints remain stable throughout this process? This is not a question that can be verified through a single evaluation, but one that requires continuously tracking behavioral consistency in a dynamically iterating system. This is one of the core technical challenges in current AI safety research, and the fundamental reason capital and the safety community have polarized reactions to Mirendil’s path.
Comparisons and Precedents
Recursive self-improvement is not Mirendil’s invention. Code generation systems such as OpenAI’s AlphaCode have already demonstrated partial forms of AI-assisted AI R&D on specific tasks; earlier theoretical discussions (such as the “intelligence explosion” concept proposed by I.J. Good in 1965) also foresaw the potential and risks of this path. But from a research concept to a lab backed by large-scale capital, Mirendil represents a jump in magnitude: a $200 million seed round, reportedly one of the largest seed financings in AI startup history, means the capital density in this direction has moved from theoretical discussion into engineering implementation.
Among similar neo-labs, Thinking Machines Lab has attracted market attention because of the background of OpenAI’s former CTO, but its public information is not yet sufficient for a horizontal technical comparison with Mirendil. The clear distinction that can currently be confirmed is that Mirendil explicitly targets accelerating scientific research (drug discovery, materials science, biology) as its application scenario, giving it a notably different positioning from most neo-labs pursuing general-purpose assistants.
The Most Critical Signals Ahead
Mirendil plans to launch its first frontier model for engineering and research in early 2027, the most direct milestone for validating its technical path. At that point, the core question to watch will not be the model’s absolute performance ranking, but whether the model can actually be deployed by institutions without large-scale AI teams (such as a university or a mid-sized pharmaceutical company) and continuously iterated within their fields. If it can, the commercial promise of recursive self-improvement will have verifiable support for the first time.
Another signal to track is the technical response from the AI safety community. There is currently no widely accepted evaluation standard for the alignment stability of self-improving systems. How Mirendil handles this issue publicly—whether it proactively publishes alignment evaluation methods or treats it as internal research not disclosed externally—will directly affect the willingness of regulators and mainstream enterprise customers to adopt it.
If this round is completed smoothly, Mirendil will have raised more than $1.2 billion cumulatively in less than a year since founding, with a valuation exceeding that of the vast majority of traditional software companies. But at a stage when it has not yet released any public model, what this money buys is a technical hypothesis, not a validated product. The distance between the two will shorten for the first time when its first model is released in early 2027.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接