Self-Improving AI

Google researchers propose RRSI to curb test memorization in self-improving AI agents, improving unseen benchmark scores while reducing runtime token use.

Google DeepMind’s Dream-RSI lets AI agents reuse past search runs to tune exploration, reducing costly attempts without changing the underlying model.

Anthropic researcher Jacob Coxon resigned over fears of self-improving AI, urging frontier labs to slow development and negotiate safety agreements.

Deep Cogito has raised $43 million to advance self-improving AI models and expand enterprise model training, though funding details remain limited.

Mirendil says it signed a multiyear Google Cloud deal worth more than $100 million to expand compute for self-improving AI research.

Recursive Superintelligence signed a $410M AWS compute deal, underscoring how self-improving AI startups are shifting capital from headcount to infrastructure.

OpenAI says GPT-5.6 Sol post-trained the smaller Luna model from a sparse prompt, a notable step toward more automated AI research workflows.

Fortune covers Anthropic's warning that self-improving AI systems could create major societal risks.

Meta's new hyperagent framework allows AI systems to autonomously rewrite their own logic and scale across non-coding tasks without human intervention.

Silicon Valley is pouring money into the dream of recursive self-improvement, with startup Ricursive Intelligence, founded by ex-Google researchers, reaching a $4 billion valuation to build AI that can improve itself.