A lean team of just seven at Poetiq is challenging the conventional, resource-intensive approach to AI development, consistently setting new AI benchmark records. This disruptive strategy, highlighted in a recent YC podcast, demonstrates how agility and innovative meta-AI systems can outperform the industry's largest players.
Poetiq's core innovation lies in its 'recursively self-improving systems,' often referred to as AI harnesses or meta-systems. These systems operate atop existing large language models (LLMs), dynamically generating optimized code, prompts, and data to enhance performance for specific tasks. Unlike the traditional method of training new LLMs from scratch, which can cost hundreds of millions of dollars and take months, Poetiq's approach avoids this massive expenditure and time commitment.
The efficacy of this strategy is evident in their benchmark achievements. Poetiq recently topped the ARC AGI V2 leaderboard, outperforming Google's Gemini 3 DeepThink at half the cost by leveraging the more economical Gemini 3 Pro. They also surpassed Anthropic's Claude Opus 4.6 on Humanity's Last Exam, scoring 55% against 53.1% and demonstrating a 9 percentage point improvement on official verification.
This 'stilt' approach offers a critical advantage for startups and developers. By continuously optimizing how existing LLMs reason and extract knowledge, Poetiq enables clients to achieve superior results that evolve with the latest foundational models. This ensures sustained high performance without the need to re-fine-tune or retrain, effectively 'vaccinating' against the rapid obsolescence of costly custom models.