AI progress races ahead while humans pedal to catch up

The pace of development in large language models has accelerated dramatically in early 2025, with breakthroughs arriving almost weekly that are reshaping our expectations of artificial intelligence. Simon Willison's recent talk, whimsically illustrated with pelicans on bicycles, captures this technological vertigo perfectly. As business leaders struggle to keep pace with these developments, Willison offers a clear-eyed assessment of where we are and where we're headed.

In his characteristically accessible style, Willison walks us through the current state of LLMs in 2025, highlighting how dramatically the landscape has shifted in a short time. He focuses on the extraordinary capabilities of models like Claude 3.5 Sonnet and GPT-5, using concrete examples that demonstrate their reasoning abilities and surprising emergent properties. The pelican metaphor runs throughout – suggesting we're all somewhat awkward creatures trying to balance on unfamiliar technology – yet the underlying message is deadly serious: AI capabilities are advancing at a pace that demands our attention.

Key developments shaping the AI landscape in 2025:

• Tool use has become sophisticated and seamless – Modern LLMs can now identify when to use tools, select the appropriate ones, and execute complex workflows without explicit instruction, representing a massive shift from earlier models that required careful prompting.

• Multi-modal integration has reached new heights – The latest models can reason across text, images, code, and other modalities in ways that feel increasingly natural, breaking down the artificial boundaries between different types of information.

• Self-improvement capabilities have emerged – Models can now critique their own outputs, identify weaknesses, and iteratively improve solutions, demonstrating a form of metacognition that was barely hinted at in previous generations.

• Reasoning abilities have deepened significantly – 2025's LLMs show dramatically improved abilities to work through complex problems step by step, maintain context, and avoid the hallucinations that plagued earlier generations.

Perhaps the most profound insight from Willison's talk is how the relationship between humans and AI is fundamentally changing. We're moving beyond the era of prompt engineering where humans carefully crafted inputs to extract useful outputs. Today's models actively collaborate with users, suggesting approaches, identifying gaps in reasoning, and bringing relevant knowledge to