← All episodes

The Powerful Alternative To Fine-Tuning

| 8 products mentioned
Watch on YouTube fine-tuning alternatives ai reasoning systems recursive self-improvement prompt optimization benchmark performance startup ai strategy model-agnostic approaches

Ian Fischer, co-founder of Poetic, discusses a paradigm shift beyond fine-tuning: building recursively self-improving AI reasoning harnesses that automatically optimize prompts, code, and reasoning strategies on top of existing LLMs. Rather than spending hundreds of millions retraining models from scratch, startups can use Poetic's system to outperform frontier models at a fraction of the cost, achieving state-of-the-art results on benchmarks like ARC-AGI v2 and Humanity's Last Exam while remaining compatible with future model releases.

Key takeaways
  • Fine-tuning is a losing strategy because massive investment becomes obsolete when new frontier models release; recursive self-improvement harnesses solve this by staying compatible across model versions.
  • Poetic's system generates AI reasoning harnesses (code, prompts, data, and reasoning strategies) automatically, reducing optimization costs to under $100k—dramatically cheaper than fine-tuning's hundreds of millions.
  • Reasoning strategies matter far more than prompt engineering alone: in one example, manual prompt optimization achieved 5% performance, but adding structured reasoning strategies jumped results to 95%.
  • Poetic achieved 54% on ARC-AGI v2 at half the cost of Google's Gemini Deep Think (45%), and 55% on Humanity's Last Exam compared to Anthropic Claude's 53.1%, demonstrating the "stilts" approach.
  • The meta-system learns what works for a specific problem without human intervention—it decides when to add context, generate examples, or modify reasoning—inverting the traditional ML requirement to deeply understand your dataset.
  • Startups should experiment daily with AI tools, push capability boundaries, and focus on unsolved hard problems rather than waiting for perfect conditions; the pace of change makes execution more important than planning.

Recommendations (2)

GPT-4
GPT-4 uses

"Last summer, I took a weekend and used GPT-4 to help me build an iPhone app. I hadn't done that in a decade. And yeah, it's so fast and so easy."

Ian Fischer · ▶ 0:13

Gemini 1.5 Flash

"We got like to 5% performance with Gemini 1.5 flash this was a while ago and then when we added on the the reasoning strategies we went from 5% to 95%"

Ian Fischer · ▶ 14:09

Mentioned (6)

ARC-AGI
ARC-AGI "I remember when you first came out with your paper in December of last year, uh you shot to the t..." ▶ 5:12
Gemini 3 Deep Think
Gemini 3 Deep Think "Gemini 3 Deep Think had just come out. Uh and they were, you know, really quite uh dramatically a..." ▶ 5:42
Humanity's Last Exam "So recently you guys just announced some incredible results for humanity's last exam. Can you tel..." ▶ 6:39
Claude Opus 4.6
Claude Opus 4.6 "They got 53.1% and we got 55% on it." ▶ 7:13
GPT-3.5
GPT-3.5 "You fine-tuned, you know, like 3 years ago on top of GPT 3.5 or whatever and then GPT-4 comes out..." ▶ 4:08
DSPy
DSPy "DSPy is this very popular paper everybody's kind of implementing that that will get you some perf..." ▶ 14:33