The Powerful Alternative To Fine-Tuning

Fisher says Poetic aims to improve how language models solve hard problems without training new models from scratch. Its harnesses combine models with code, prompts, and data, and can be adapted as newer models become available. He argues this can avoid the cost and obsolescence risks of fine-tuning.

Fisher describes Poetic’s results on ARC-AGI-2: 54% verified performance at $32 per problem, compared with Gemini 3 Deep Think at 45% and about $70 per problem. He also says Poetic reached 55% on Humanity’s Last Exam, above the cited 53.1% result from Claude Opus 4.6. The optimization for that run cost less than $100,000, and the company had seven research scientists and engineers.

Fisher says Poetic’s meta-system automatically generates or optimizes systems for difficult tasks. Customers can bring an existing agent for optimization of prompts, reasoning strategies, or other parts. He says improvements can come from reasoning strategies written in code: in an earlier DeepMind project, adding these strategies raised performance with Gemini 1.5 Flash from about 5% to 95% on the hardest task.

Fisher says Poetic’s system examines task data and decides whether it needs more context, generated examples, or other changes. For one benchmark, it produced prompts that Fisher says a person would not typically write, including a simple example that appeared incorrect; the team left the output unchanged. He describes this as shifting dataset analysis and failure-mode discovery to the AI system.

Poetic had not yet released its product at the time of the conversation, but Fisher invited companies with difficult reliability problems to sign up for early access at poetic.ai. He also recounts moving from a mobile tools startup and computer security into machine learning research at Google and DeepMind. His advice to engineers is to use AI regularly, test its limits, and try building things they want to make.