Self-improving AI startup raises $650M as researchers pursue self-evolving models

Yuandong Tian's $650 million raise is a large bet on an alternative to the scaling paradigm: that AI models can design and improve other models, breaking the dependence on ever-larger compute. Tian, a former Meta scientist, argues that brute-force scaling will hit physical limits — energy, memory, fab capacity — and that algorithmic and efficiency breakthroughs are the only sustainable path forward. The funding size signals serious investor appetite for anti-scaling theses at a moment when compute costs dominate the industry conversation.

Separately, MIT researchers published a preprint on self-evolving 'AI scientists' aimed at autonomous scientific discovery. The work demonstrated results in specific materials-science domains but remains unproven as a general method — a caveat the researchers themselves flag.

The theme ties into the week's macro mood: a recurring worry that energy and compute, not better models, will be the real limiting factor, and a corresponding hunt for efficiency. It also resonates with the broader 'self-improving AI' narrative that fuels both excitement and AGI-timeline debate.

The skeptical read is that 'models building models' and 'self-evolving AI scientists' are concepts with thin track records — easy to fund, hard to deliver, and prone to overhyped framing. The $650M is venture conviction, not proof. Readers should watch for concrete benchmarks showing a self-improvement loop that beats human-designed architectures, and whether the MIT approach generalizes beyond materials science.