Anthropic co-founder Jack Clark says "I now think it's more likely we'll see full RSI become possible in 2027", earlier than his previous prediction of 2028

The capability to perform intuitive visual reasoning is now expected from advanced AI, potentially changing how we design and interact with AI systems.

AI startups Elorian and Scale AI have released a new benchmark called Humanity’s Sixth Sense (HSS) designed to evaluate AI models' intuitive visual reasoning capabilities, specifically their ability to infer information not explicitly shown in images or videos. The benchmark includes 522 questions across 288 images and 234 video clips, covering temporal, causal, physical, spatial, social, and abstract reasoning. Human performance on HSS is exceptionally high, with an aggregate pass@1 score of 93.1%. In contrast, the leading AI model, GPT 6 Astra, achieved a score of 53.6%, while Chinese open-weight models like GLM-5.3-flash and Kimi K3 scored significantly lower at 26.8% and 25.5%, respectively. Common AI errors involved missing crucial details or misidentifying elements within scenes.

The development of the Humanity’s Sixth Sense benchmark suggests that advanced LLMs like GPT and Claude are demonstrating capabilities previously thought to require distinct AI systems, such as dedicated 2D/3D world models. This challenges the notion that complex visual reasoning is strictly dependent on specialized architectures, indicating that current large multimodal models are already developing a form of emergent world understanding. The performance gap between frontier models and Chinese open-weight models highlights a divergence in capabilities, potentially influencing the competitive landscape for AI research and development.

Jack Clark, co-founder of Anthropic, has revised his prediction for when full recursive self-improvement (RSI) in AI might become possible, now anticipating it for 2027 instead of 2028. This shift is influenced by recent advancements in AI's ability to exhibit 'research taste,' which includes selecting problems, designing experiments, and interpreting results, as measured by the new TasteVal benchmark. The observation that AI systems are doubling their experimental research taste every 3.0 months, with top models like Opus 5.5 exceeding human expert baselines, suggests a faster-than-expected trajectory towards AI systems capable of self-improvement.