====== Poetiq ======
**Poetiq** is an AI startup founded by former DeepMind researchers that specializes in building recursive, self-improving [[meta|meta]]-systems designed to enhance the capabilities of existing large language models and AI systems. Rather than pursuing traditional model training or scaling approaches, Poetiq develops sophisticated orchestration layers and harness architectures that amplify performance through intelligent system composition.

===== Overview =====
Poetiq's core innovation lies in its approach to AI performance optimization. Instead of training new foundation models from scratch—a computationally expensive and resource-intensive process—the company creates [[meta|meta]]-systems that coordinate, refine, and iteratively improve outputs from existing models. This philosophy emphasizes engineering efficiency and cost-effectiveness while pursuing advanced reasoning and problem-solving capabilities.

The company's founding team drew expertise from DeepMind, one of the world's leading AI research institutions, bringing deep knowledge of machine learning systems, neural architecture design, and advanced reasoning techniques.

===== ARC-AGI-2 Breakthrough =====
Poetiq achieved significant recognition when their system attained **54% accuracy on the ARC-AGI-2 benchmark**, a major milestone in artificial general reasoning(([[https://alphasignalai.substack.com/p/anthropics-512k-line-code-leak-reveals|Alpha Signal AI - Anthropic's 512K Line Code Leak Reveals (2024]])).

This result is notable for several reasons:

* The performance was achieved through an orchestration layer rather than training a new frontier model
* The approach demonstrated lower computational costs compared to developing raw frontier capabilities
* The result highlighted the potential of [[meta|meta]]-system design for complex reasoning tasks

The ARC-AGI-2 benchmark, which focuses on abstract reasoning and few-shot learning capabilities, is considered one of the most challenging evaluation frameworks for assessing progress toward artificial general intelligence.

==== Cost-Efficiency Comparison ====
Poetiq's orchestration approach demonstrates substantial cost advantages when compared to frontier models. On the ARC-AGI-2 benchmark, Poetiq achieved its 54% accuracy at a cost of **$30.57 per problem**, significantly outperforming Google's Gemini 3 Deep Think, which attained 45% accuracy at **$77.16 per problem**(([[https://alphasignalai.substack.com/p/anthropics-512k-line-code-leak-reveals|Alpha Signal AI - Anthropic's 512K Line Code Leak Reveals (2024]])).

This comparison illustrates how an effective orchestration layer can surpass the capabilities of more expensive, standalone frontier models while requiring substantially lower computational investment per solution.

===== Approach and Philosophy =====
Poetiq's methodology represents a paradigm shift in AI performance optimization:

* **Recursive [[meta|meta]]-systems**: Self-improving systems that refine their own outputs through iterative processes
* **Orchestration over training**: Leveraging existing models through intelligent coordination rather than expensive retraining
* **Cost efficiency**: Delivering higher performance at reduced computational and financial expense
* **Compositional design**: Building complex capabilities by combining simpler, existing components

This approach suggests that significant performance gains may be achievable not solely through larger models or more training data, but through superior engineering and system design.

===== Significance =====
Poetiq's work has implications for the AI industry's future trajectory. The company demonstrates that the path to advanced AI capabilities need not rely exclusively on ever-larger foundation models trained on massive datasets. Instead, sophisticated harnesses and [[meta|meta]]-systems can unlock substantial improvements from existing infrastructure, potentially offering a more sustainable and cost-effective pathway to performance gains. The demonstrated cost advantage over frontier models suggests that orchestration-based approaches may represent a fundamentally more efficient strategy for achieving competitive reasoning capabilities.

===== See Also =====
  * [[google_deepmind|Google DeepMind]]
  * [[mistral_ai|Mistral AI]]
  * [[together_ai|Together AI]]
  * [[deepseek|DeepSeek]]
  * [[devin|Devin: Autonomous AI Coding Agent]]

===== References =====