Model can accept 1M tokens doesn't mean it can reason across those 1M tokens (twitter.com)

In a recent discussion, a Prime Intellect engineer challenged the prevailing narrative around large-context models, specifically those boasting capacities of up to a million tokens. He revealed that while models like GPT-5.5 perform admirably with 256k tokens—achieving an 80% score on retrieval—they falter significantly at a million tokens, dropping to just 36%. This phenomenon, referred to as "context rot," highlights a critical limitation: models can accept a larger context but struggle to effectively reason across it. The engineer emphasized that simply expanding context windows won't enhance AI capabilities; instead, he advocates for solutions like continual learning, training on personal data traces, and real-world environments to improve reasoning. In other noteworthy news, Andrew Ng has launched a free course on Claude Code, developed in collaboration with the Anthropic team. This course promises to demystify Claude Code's surprisingly simple architecture and its agentic capabilities, allowing users to apply it to any codebase. Ng asserts that this brief resource can effectively replace numerous paid coding agent courses, making advanced AI training more accessible to the community. Together, these developments underline the ongoing discourse in the AI/ML field about optimizing model performance and the educational resources available for practitioners.