Ian Fischer, co‑founder and co‑CEO of Poetic, explains that his company builds “recursively self‑improving” AI reasoning harnesses that sit on top of one or more large language models (LLMs) and automatically generate systems that outperform the base models without the massive cost of training new LLMs from scratch. Unlike traditional fine‑tuning or reinforcement‑learning approaches—which require hundreds of millions of dollars and become obsolete when a newer frontier model appears—Poetic’s harnesses are cheap to optimize (often under $100 k) and remain compatible with each new model release, delivering continual performance gains.
The team demonstrated this advantage on benchmarks such as ARC‑AGI v2 (raising scores from ~45 % to ~54 % at a fraction of Gemini 3 DeepThink’s cost) and, more recently, on Humanity’s Last Exam, achieving 55 % accuracy—about two points above the prior state‑of‑the‑art (Claude Opus 4.6 at 53.1 %). Poetic’s “meta system” automatically discovers effective prompts, reasoning strategies, context‑stuffing, and rerouting of LLM calls, effectively outsourcing data‑understanding to the AI itself.
Although the technology is not yet publicly released, interested startups can sign up for early access at poetic.ai. Fischer’s own journey—from a YC‑backed mobile‑tools startup, through a decade at Google/DeepMind researching machine learning, to founding Poetic—illustrates his belief that engineers should constantly experiment with AI, push its limits, and build the tools they envision. He encourages founders to apply to Y Combinator and to treat AI as a readily available lever for making the world better.
1. Ian Fischer is the co‑founder and co‑CEO of Poetic.
2. Poetic builds recursively self‑improving AI reasoning harnesses for large language models.
3. Ian Fischer previously spent a decade as a researcher at Google DeepMind.
4. Ian Fischer founded a mobile‑devtools company through YC years ago.
5. Poetic’s system can generate problem‑specific systems that outperform underlying language models without the massive expense of training new LLMs from scratch.
6. The generated systems remain compatible with new model releases without requiring changes.
7. Training a new LLM from scratch costs hundreds of millions of dollars and takes months.
8. On the ARC AGI V2 benchmark, Poetic’s system achieved a 9‑percentage‑point improvement over Gemini 3 Deep Think (45 % → 54 %) at about $32 per problem, while Gemini 3 Deep Think cost around $70‑something per problem.
9. Poetic’s approach was half the cost of Gemini 3 Deep Think on ARC AGI V2.
10. On Humanity’s Last Exam (2500 expert‑written questions), Poetic scored 55 %, almost two percentage points higher than the prior state‑of‑the‑art (Anthropic Claude Opus 4.6 at 53.1 %).
11. The optimization cost for Poetic’s Humanity’s Last Exam run was less than $100 k.
12. Poetic’s team consists of seven research scientists and engineers.
13. Poetic’s meta system can automatically generate reasoning systems, prompts, and strategies for hard problems.
14. The system can optimize prompts, reasoning strategies, context stuffing, summarizing, or reranking depending on the task.
15. Using Poetic’s reasoning strategies improved performance on a hard task from 5 % to 95 % with Gemini 1.5 Flash.
16. Poetic’s system does not require users to deeply understand their dataset; the AI itself identifies failure modes and robust reasoning strategies.
17. Poetic offers early access via poetic.ai for startups with hard problems.
18. Poetic seeks startups that have tried existing methods and need something more reliable.
19. Poetic’s approach is distinct from traditional RL and context engineering.
20. Poetic’s system can be applied to various tasks, with performance gains varying per problem.
21. Poetic’s system enables continuous optimization for new models without losing prior investment.
22. Poetic’s system avoids the “bitter lesson” cost of repeatedly fine‑tuning models.
23. Ian Fischer used GPT‑5 to build an iPhone app in a weekend, something he hadn’t done in a decade.
24. YC’s next batch is accepting applications at ycombinator.com/apply.