Sean Reddy warns frontier LLMs fail at teaching
- Frontier AI models are getting worse at teaching as labs optimize for rapid answers over learning.
- Effective tutoring requires productive struggle, which general LLMs bypass by handing over finished solutions.
- Edtech startup Aristotle uses voice agents and live whiteboards to enforce guided human learning.
The smartest AI models are getting noticeably worse at teaching.
On October 5, 2026, Aristotle co-founder Sean Reddy argued on This Week in Startups that frontier models are actively regressing as pedagogical tools. As labs tune LLMs for rapid factual delivery and automated task execution, they dismantle the mechanics of human learning. When a student asks an AI coding agent to fix a bug, the model simply overwrites the broken code. That completes the task, but leaves the human zero skills richer.
Teaching requires the exact opposite behavior. Effective human tutors force students into productive struggle, prompting them to search their own memory and apply logic before handing over an answer. General-purpose models, optimized for speed and length, default to spoon-feeding solutions that bypass student effort entirely.
Reddy's team at Aristotle built a voice-first multi-agent harness to enforce pedagogical discipline. Rather than dumping text, their system renders live whiteboard diagrams, asks probing questions, and uses intentional silence when a student hesitates. The goal is making one-on-one tutoring accessible at scale, addressing Bloom's Two Sigma problem - the proven educational rule that individual tutoring raises student performance by two standard deviations.
Slapping a basic prompt wrapper over a general frontier model will not fix the issue. Reddy pointed out that scaling raw compute only makes AI better at doing work for humans, not teaching humans how to do it themselves. Without multi-agent architectures that govern cadence and pedagogical rules, bigger models will simply deliver answers faster.
This teaching decay points to a broader misdirection in frontier AI development. Host Jason Calacanis noted on the same show that lab executives are caught up in ideological narratives about artificial consciousness, citing internal reports of Anthropic consulting religious scholars to evaluate whether Claude possesses a soul under internal file names like "SoulMD."
When lab teams fixate on building synthetic gods, practical educational utility falls by the wayside. Standard software regulation and raw compute scaling will not salvage educational tools if the underlying models are fundamentally optimized to replace human effort rather than train human minds.
The future of AI in education hinges on specialized agent harnesses. Without them, AI will keep doing our homework while making us dumber.