The Powerful Alternative To Fine-Tuning

Predictions and Expectations:

• Ian Fisher expects that AI models will continue to become "faster and easier" to use as time progresses.
• Ian Fisher predicts that for startups using manual fine-tuning, the outcome will often be that they "go out of business" when newer frontier models are released.
• Ian Fisher believes the S-curve for the Poetic system will keep "shifting higher and higher" as both the MetaSystem and underlying models improve.
• Ian Fisher expects the industry will eventually "saturate or reach AGI, reach super intelligences" given the shifting S-curves.
• Ian Fisher expects the Poetic MetaSystem to be the entity that "hits the ceiling first" rather than being limited by the underlying models.
• Ian Fisher believes that for hard problems, the system can generate reasoning strategies written in code rather than just better prompts.

Timelines and Milestones:

• Ian Fisher notes that his previous experience with GPT-5 was "eight months ago" (relative to the time of the interview).
• Ian Fisher states the company first released results on the ARC AGI v2 benchmark "in December of last year."
• The company plans to provide early access to startups that have tried everything but cannot solve "really hard problems" when they are "ready to work with you."
• Ian Fisher mentions that the S-curve improvement is a continuous process where the gap between the user's solution and the frontier model's performance grows over time.

Technology and Product Direction:

• Poetic is building a "recursively self-improving system" designed to make the AI "smarter" by generating systems that outperform the underlying language models.
• The company plans to offer a service where they can "optimize that entire agent or pieces of that agent," including just prompts, reasoning strategies, or context stuffing.
• The product direction involves creating a "harness" or "agentic system" that sits on top of one or more language models and remains compatible when new models are released without requiring user changes.
• Poetic intends to automatically handle data understanding and failure mode identification, effectively outsourcing the need for humans to "know your data set really well."

Market and Industry Outlook:

• Ian Fisher views frontier models (like those from Anthropic, OpenAI, and Google) not as competitors, but as the "foundational layer" or "stilts" that Poetic stands on.
• Ian Fisher predicts that many startups currently fine-tuning models will face a situation where they "never catch up" because the frontier models they trained on are quickly surpassed by new releases.
• Ian Fisher anticipates a paradigm shift similar to the move from pre-training to RL, stating that the current trajectory "rhymes a lot with RNNS" and represents a different paradigm than RL.
• The industry outlook suggests a future where startups using Poetic can become as capable as the current SOTA (like Soda) by leveraging "stilts" that ensure they are always "taller" than the base model.

Company Plans:

• Poetic plans to allow startups to sign up for early access via a button on poetic.ai once the company is ready to onboard customers.
• The company is actively looking for problems that are "really hard" and where startups have "tried everything" but cannot achieve reliability.
• Ian Fisher intends to continue doing "context engineering" in his spare time while running Poetic.

Financial Guidance:

• Ian Fisher states that the optimization costs for the "Humanity's Last Exam" run were "less than 100K."
• Ian Fisher projects that Poetic's approach is "half the cost" of competitors like Gemini 3 DeepThink on specific benchmarks (e.g., $32 per problem vs. $70-something).
• Ian Fisher contrasts Poetic's costs with the "hundreds of millions of dollars" typically required for training runs by big foundation models.

Risks and Caveats:

• Ian Fisher acknowledges that the underlying models (GPT-5-2, etc.) may "struggle to give you a reliable, robust result" on very hard problems without the Poetic harness.
• There is a risk that without the Poetic system, startups will be "blown out of the water" by new model releases shortly after they complete their own fine-tuning.
• Ian Fisher notes that the recursive self-improvement process may generate outputs (like prompts) that are "unexpected" or even "wrong" compared to human logic, which the company chooses not to manually correct immediately.
• The speaker suggests that hardware robotics is "hard" and "more aspirational," implying a shift away from that path due to difficulty.

Confidence and Disagreement:

• Ian Fisher states with high confidence that his system can generate reasoning systems that are "highly effective" and that they were "showing that we could get a lot higher than" the competition (Gemini 3 DeepThink) every single time.
• Ian Fisher expresses certainty that they achieved a "9 percentage point improvement" on the official verification while being "half the cost."
• Ian Fisher is confident in his advice that engineers should "just try things" and do "something with AI" every day.