The Powerful Alternative To Fine-Tuning

TL;DR

Poetic's recursively self-improving AI harness outperforms fine-tuned models at a fraction of the cost, and stays current as new frontier models release.

Key Points

• 1.Fine-tuning is a dead end: it costs millions, takes months, and gets wiped out the moment a new frontier model drops (e.g., fine-tuning on GPT-3.5, then GPT-4 obliterates your investment)
• 2.Poetic builds model-agnostic "harnesses" — code, prompts, and reasoning strategies layered on top of existing LLMs — that automatically get better every time the underlying model improves
• 3.On ARC-AGI v2, Poetic beat Gemini 3 Deep Think (45%) with a score of 54%, at half the cost ($32/problem vs. ~$70+), using the cheaper Gemini 3 Pro model
• 4.On Humanity's Last Exam (2,500 PhD-level questions), Poetic hit 55%, beating Anthropic's Claude Opus 4.6 (53.1%) — state-of-the-art as of last week — for under $100K in optimization costs
• 5.The biggest performance gains come not from better prompts (which help marginally) but from AI-generated reasoning strategies written in code — in one DeepMind experiment, this jumped performance from 5% to 95%
• 6.Poetic is a 7-person team; startups can sign up for early access at poetic.ai to get their existing agent optimized without rebuilding or retraining from scratch