The Powerful Alternative To Fine-Tuning

Poetic's Core Technology

• Builds recursively self-improving AI reasoning harnesses that sit on top of foundation models to outperform them.
• The "Poetic Meta-System" automatically generates systems to solve hard problems, optimizing prompts, data, and reasoning strategies without human intervention.
• Unlike traditional fine-tuning or training models from scratch (which costs hundreds of millions of dollars), Poetic's approach is significantly cheaper and faster.
• The harness remains compatible with new frontier model releases, allowing startups to gain performance bumps without re-training or rebuilding.
• This approach effectively acts as "stilts" for startups, making them "taller" than the underlying base models.

Performance Benchmarks and Economics

• Arc AGI v2: Achieved results 9 percentage points higher than Gemini 3 DeepThink (54% vs. 45%) while costing less than half the price ($32 vs. ~$70 per problem) by utilizing Gemini 3 Pro.
• Humanity's Last Exam: Reached 55% accuracy on 2,500 expert-level questions, surpassing the previous state-of-the-art (Claude Opus 4.6 at 53.1%).
• Cost Efficiency: The Humanity's Last Exam run was optimized for under $100,000, whereas training foundation models to similar capabilities costs hundreds of millions.
• Team Structure: The optimization is performed by a lean team of seven research scientists and engineers.
• Historical Context: Previous manual optimization on a hard task with Gemini 1.5 Flash yielded only 5% performance, which jumped to 95% after adding reasoning strategies.

Strategic Philosophy and Market Position

• Treats frontier models as a foundational layer to stand on, not as competitors.
• Positions the company against the "bitter lesson" where fine-tuned models become obsolete immediately upon the release of new, stronger base models.
• Views the technology as a new paradigm distinct from Reinforcement Learning (RL), with its own S-curve that shifts upward as both the Meta-System and underlying models improve.
• Currently operates in stealth and has not publicly released the product; early access is available via a signup button on poetic.ai.
• Actively seeking startups with reliable, robust agents that require further optimization for specific verticals.

Founder Background and Origins

• Ian Fisher, co-founder and co-CEO, previously founded a mobile dev tools company acquired by Google, where he spent a decade researching AI and robotics at Google Research and DeepMind.
• Initially joined Google to work on robotics but pivoted to machine learning research upon realizing hardware constraints were too limiting.
• Recently demonstrated the rapid capability of AI by building an iPhone app over a weekend using GPT-5, a task he hadn't attempted in a decade.

Call to Action and Future Outlook

• Encourages engineers and founders to "try things" daily with AI to push boundaries, citing personal experience as proof of how quickly capabilities are advancing.
• The goal is for the Poetic Meta-System to hit the performance ceiling of AGI or superintelligence before the underlying models do.
• Y Combinator's next batch is accepting applications at ycombinator.com/apply.