Recursive self-improvement as a startup alternative to fine-tuning

Executive overview

Fine-tuning frontier models costs millions and is obsolete the moment a new model ships. Poetic builds "reasoning harnesses" — agentic systems layered on top of existing LLMs — that outperform the base models without retraining.

The harness is compatible with any new model, so performance gains compound rather than reset. A seven-person team hit state-of-the-art on both ARC-AGI v2 and Humanity's Last Exam at a fraction of frontier training costs.

Automated prompt and reasoning-strategy optimisation is a more durable moat than fine-tuning.

What Poetic builds and why it beats fine-tuning

• Fine-tuning locks you to one model version; the next frontier release erases the investment
• A reasoning harness is code, prompts, and data sitting on top of one or more LLMs
• When a new base model ships, the same harness applies immediately — often with a larger performance jump
• Poetic's meta system generates these harnesses automatically, far faster and cheaper than hand-building them
• The meta system can also optimise an existing agent a startup has already built — targeting prompts, reasoning strategies, or the full pipeline

Benchmark results

• ARC-AGI v2: Poetic scored 54%, versus Gemini 3 Deep Think at 45% — achieved on Gemini 3 Pro (a cheaper model), at half the cost (~$32 vs ~$70 per problem)
• Humanity's Last Exam: 55% versus Anthropic's Claude Opus 4.6 at 53.1%, with optimisation costs under $100K
• Both results produced by a team of seven research scientists and engineers

How the meta system works

• The meta system observes a task, generates candidate reasoning strategies, and recursively refines them
• Outputs include non-obvious prompts — ones a human would not write — and deliberately simple (sometimes technically wrong) examples that still improve performance
• The system decides how much context stuffing, example generation, or re-ranking is needed; humans do not tune this manually
• Prompt optimisation alone yields modest gains; adding reasoning strategies encoded in code is where large jumps occur (one task: 5% → 95% accuracy)
• Each underlying model has its own S-curve; as models and the meta system both improve, the ceiling keeps rising

Who Poetic is for

• Startups that have exhausted context engineering and can't get reliable, robust results from base models
• Teams that have already built an agent but need a performance step-change
• Any builder who wants to stay at or above SOTA without re-running expensive training each model cycle
• Early access available at poetic.ai

Ian Fischer's path from mobile tools to DeepMind to Poetic

• First YC company (Aportable) was acquired by Google; used the transition to move into AI and robotics research
• Realised hardware was not the focus; pivoted fully into machine learning research at Google Brain and DeepMind for ~10 years
• Advice for engineers entering AI: experiment daily, push the boundaries of what models can do, build what you want to build