The Powerful Alternative To Fine-Tuning

In this video, Ian advocates for hands-on daily experimentation with AI, emphasizing that building powerful applications is now faster and easier than ever. He introduces Poetic, a startup pursuing the "holy grail" of recursively self-improving AI, but doing so far more cheaply than traditional billion-dollar model training. Poetic’s core insight is that fine-tuning is a losing game—new base models quickly obsolete old fine-tuned ones—so instead it builds a "harness" of code, prompts, and data that routes tasks across models and automatically improves with each new frontier release. This approach delivers impressive results, scoring 54% on ARC AGI v2 at $32 per problem, beating Gemini 3 Deep Think by 9 points at roughly half the cost, all with just seven researchers and under $100k per run. It also shifts the paradigm from human-led data analysis to having the AI itself identify failure modes and discover robust reasoning strategies. The company is not yet public, but startups can sign up for early access at poetic.ai, and Ian's final advice is simple: build with AI every day and push models to their limits.

• ▶ 0:00 The AI landscape is evolving extremely fast; Ian urges viewers to experiment with AI daily, noting that building things like an iPhone app with GPT-5 is now faster and easier than ever.
• ▶ 1:02 Poetic's mission is to build a recursively self-improving AI system — the "holy grail" — and to achieve it far faster and cheaper than the standard approach of training new LLMs from scratch, which costs hundreds of millions of dollars.
• ▶ 2:08 Poetic's value for startups is that it always outperforms the out-of-box frontier model without expensive fine-tuning, so your investment isn't "lit on fire" every time a new frontier model releases.
• ▶ 3:57 Fine-tuning is a losing game: new base models quickly outperform old fine-tuned ones. Poetic instead builds a "harness" on top of models that stays compatible and gets better with each new model, at far lower cost.
• ▶ 6:14 On ARC AGI v2, Poetic scored 54% at $32 per problem — beating Gemini 3 Deep Think's 45% (about $70+) by 9 points at roughly half the cost.
• ▶ 7:38 Poetic achieves these results with only seven researchers and under $100k per run; its core technology is a harness of code, prompts, and data that automatically routes tasks across language models without a human orchestrator.

• ▶ 9:35 The approach enables much more automated generation and optimization, making it quicker and cheaper than building custom agents, by optimizing the entire agent or specific components like prompts and reasoning strategies.

• ▶ 10:26 This is a different paradigm from RL: each model and the meta system have their own S-curve, which keeps shifting higher as both improve, until saturation or AGI/superintelligence is reached.

• ▶ 12:53 They are outsourcing dataset understanding to the AI itself—the AI analyzes the data, identifies failure modes, and discovers robust reasoning strategies, rather than requiring humans to know the dataset deeply.

• ▶ 15:01 Poetic hasn't been publicly released yet, but startups and companies with hard problems can sign up for early access at poetic.ai, and early partners will be the first to know when it's ready.
• ▶ 17:53 Ian's path to AI was indirect: after Google acquired his startup, he joined a robotics team in Google Research, learned "hardware is hard," then made a hard switch into machine learning research, spending a decade at Google and DeepMind.
• ▶ 18:29 His advice for engineers is to try things and do something with AI every day—build what you want, push models to their boundaries, and don't limit yourself; using AI can help make the world better.

• ▶ 0:00 Introduction and Poetic’s Mission (0:00 - 3:59) - Summary: Opening, welcome, Poetic’s mission, why startups want it, and the automatically generated system.
• ▶ 3:59 Benchmarks, Team, and Core Technology (3:59 - 9:35) - Summary: Fine-tuning obsolescence, benchmark wins, run costs, the seven-person team, bug routing, and core harness tech.
• ▶ 9:35 A Different Paradigm: Automation and Reasoning (9:35 - 14:51) - Summary: Automated generation, optimizing agents, S-curves, context engineering, and the DeepMind paper.
• ▶ 14:51 Startups, Background, and Closing Advice (14:51 - 19:36) - Summary: Poetic for startups, the YC batch, Ian’s Google background, advice for engineers, and closing remarks.