Jack Clark says TasteVal points to self-improving AI by 2027
AI might soon start making itself smarter, without much help from humans.
Jack Clark, co-founder of Anthropic, thinks we could see self-improving AI as early as 2027.
His prediction comes partly from a new benchmark called TasteVal and also from recent AI maths breakthroughs that look creative.
Basically, AI's "experimental research taste" has been doubling roughly every three months since December 2025, and Anthropic's Opus 5.5 beat the human baseline on the benchmark at about 1/30 of the experts' average cost per attempt.
Opus 5.5 outperforms experts cost-effectively
TasteVal tested AIs on real-world challenges like prepping training data and fine-tuning models, and Anthropic's Opus 5.5 model even outperformed human experts from places like OpenAI and Google DeepMind, all while being way more cost-effective.
If AIs keep getting better this quickly, they might soon upgrade themselves with little human input.