Poetic's core technology is a 'metasystem' that uses recursive self-improvement to automatically generate optimized 'harnesses' for LLMs. These harnesses, composed of code, prompts, and data, enhance the reasoning and knowledge extraction capabilities of foundation models for specific, difficult tasks.
The rapid release of new, more powerful foundation models often invalidates costly fine-tuning efforts. Poetic's model-agnostic harnesses provide a durable performance layer, allowing companies to leverage the latest models immediately without re-engineering their systems.
Poetic consistently achieves state-of-the-art results on challenging AI benchmarks while using cheaper underlying models. They surpassed Google's top score on Arc AGI V2 at half the cost and achieved a new high score on the Humanities Last Exam with an optimization cost under $100,000.
The discussion highlights a shift from manual prompt and context engineering to automated system optimization. Poetic's metasystem outsources the task of understanding a dataset and designing complex reasoning strategies to the AI itself, producing non-obvious solutions.
Keep pulling the thread on Ian Fisher.