Self-improvement AI attracts $650 million: productless startup is worth $4.65 billion

Artificial intelligence capable of self-improvement has won the biggest check ever given to an unprescribed startup. The Recursive Superintelligence (RSI), founded by Yuandong Tian — former Meta scientist — raised $650 million in round led by Alphabet's venture capital fund, reaching a value of $4.65 billion. The proposal is radical: to develop models that build other models, creating a cycle of continuous improvement without human intervention.

How does self-improvable AI work in startup?

The central concept of RSI is a system that, in addition to performing tasks, is also capable of modifying its own architecture and algorithms. Instead of relying on engineers to adjust parameters or re-train with new data, the model analyzes its performance and rewrites parts of its code. This promises to exponentially accelerate the advancement of AI, something Tian already studied at Meta AI Research, where he led learning projects by reinforcement and self-supervision.

Why does Alphabet bet on AI self-improvement without product?

Alphabet, through GV (formerly Google Ventures), has already invested in companies that challenged common sense — such as Uber and Slack — before having commercial traction. In the case of RSI, the thesis is that self-improvement capacity can solve the current bottleneck of artificial intelligence: dependence on labeled data and manual engineering for each new domain. If it works, technology could reduce development costs and expand the AI application to areas where today is unfeasible.

What does self-improvement AI mean for the market?

The movement signals that large players are willing to finance fundamental research, even without a deadline for financial return. Tian argues that RSI is not a product company, but an applied research laboratory — something that resembles DeepMind's initial approach before Google acquired it. Brazilian companies that develop AI should observe closely: if the model proves feasible, it can completely change the innovation cycle, making many of the current fine-tuning practices obsolete.

IA self-improvement and security: what is the relationship with current debates?

The possibility of an AI that improves itself rekindles discussions about alignment and control. While some celebrate the potential to overcome human boundaries, others warn of existential risks. Tian, in interviews, argues that RSI will be built with layers of security incorporated since design — something that in practice is still theoretical, since there is no product to test. The debate connects with the ongoing regulatory proposals, as advocated by Senator Bernie Sanders, which suggests confiscation of large laboratory actions to finance public research.

Frequently asked questions about self-improvement AI

What makes Recursive Superintelligence different from other AI startups?

It does not sell services or APIs; its sole goal is to develop a model that can improve itself, without human intervention, creating a recursive cycle of improvement.

How much money has the startup raised?

$650 million in a single round, led by GV (Alphabet), with a value of $4.65 billion — even without having released any commercial products.

What's the founder's background?

Yuandong Tian was a senior scientist at Meta AI Research, where he led advances in reinforcement learning, self-supervised neural networks and hyperparameter optimization.

---

Source: TechTimes. Read also: Bernie Sanders proposes to confiscate 50% of OpenAI and Anthropic shares for public funds