Turing Award winner Richard Sutton argues that conventional generative AI systems cannot perform genuine scientific discovery because they lack built-in evaluation mechanisms to assess their own results.
Sutton identifies a fundamental limitation in current generative AI: without the ability to evaluate outputs, these systems cannot sustain scientific breakthroughs. Novel discoveries emerge only momentarily before disappearing, he explains.
According to Sutton, AI systems like AlphaGo and AlphaProof demonstrate what's required for genuine creativity in science—integrated feedback loops that allow systems to test and validate their own findings. This evaluation capacity separates truly creative AI from systems that merely generate plausible-sounding but unverified outputs.
The distinction matters for fields where verification and reproducibility are essential. Generative models excel at pattern recognition and synthesis but lack the iterative testing mechanisms necessary for scientific validation. Sutton's assessment suggests that achieving AI-driven scientific discovery will require architectural changes beyond current generative approaches, incorporating built-in evaluation and refinement capabilities.
Singaporeans show strong willingness to use AI shopping agents, but demand strict safeguards to govern how the technology operates, according to new research.
Rabbit has released OS3, a cloud-based agentic operating system that runs on Windows, Mac, and Linux devices without requiring the company's R1 hardware. Users can connect up to five devices to a single account and select their preferred AI models.
Anthropic has launched Claude Opus 5.5, the latest iteration of its flagship AI model. The release comes as competition intensifies in the large language model market.
OpenAI released two cheaper versions of GPT-6 on Tuesday, with pricing roughly half that of GPT-5.6. GPT-6 Luna offers the steepest discounts at $0.10 per million input tokens and $0.50 per million output tokens.