Google DeepMind chief Demis Hassabis warned that today’s top AI systems exhibit “jagged intelligence,” excelling at elite math tasks while stumbling on basic problems, underscoring gaps in reasoning and planning. Speaking on a Google podcast, he argued that more data and compute alone won’t close the gap to AGI and called for tougher, more diagnostic benchmarks. The remarks echo concerns voiced by OpenAI’s Sam Altman about the lack of continuous learning and rising risks from hallucinations and misinformation. The message: the industry must fix brittleness now to avoid repeating social media’s early missteps at scale.
Related articles:
BIG-bench: Beyond the Imitation Game benchmark
Measuring Massive Multitask Language Understanding (MMLU)
Tree of Thoughts: Deliberate Problem Solving with Large Language Models
Meet Gemini: Google’s multimodal AI models
Training Compute-Optimal Large Language Models (Chinchilla)
