How to Compete When Cheap AI Models Converge

TL;DR

Use cheap AI models for routine execution, but reserve frontier models for discovering and prototyping valuable tasks that were previously impossible. As execution becomes inexpensive and widely available, differentiation increasingly depends on firsthand capability knowledge, expert judgment, and the imagination to ask questions that competitors have not yet considered.

Transcript

Here's something strange that everyone using AI is noticing right around the same time. The tools keep getting better. The prices keep dropping. You can produce more than ever and yet somehow everything's starting to feel the same. Your output, your competitors output, half of what's in your feed all looks really similar, right? Better tools, samey... Read More

Key Insights

• Routine AI execution is becoming commoditized because cheap and expensive models can deliver similarly acceptable results on familiar tasks such as implementing a predefined feature. When many users submit the same known work through similar tools and public playbooks, their outputs naturally converge.
• Model routing is a cost-control layer that should be used aggressively for ordinary execution. However, routing alone cannot remain a durable advantage because it is becoming table stakes, meaning competitors will also gain access to inexpensive models and comparable efficiency.
• The value of frontier models is clearest on difficult, previously unassigned problems. Mitchell Hashimoto reportedly spent two hours and $40 using a frontier model to optimize complex systems code, reaching performance that he said he could not have achieved independently.
• AI can only execute work that someone has first imagined. A backlog, sprint, prompt collection, or best-practices guide cannot automatically identify every newly possible task, so the user's ability to recognize valuable questions sets an important ceiling on AI returns.
• Imagination in AI work is developed through direct exposure rather than treated as an artistic gift. Hundreds or thousands of hours using models can create an intuitive understanding of where the capability boundary has moved and which previously impractical questions are now worth testing.
• A two-layer AI strategy combines cheap execution with targeted frontier exploration. Cheap models provide the engine for repeatable work, while frontier models help determine direction by testing hard questions that could change what the execution layer ultimately builds.
• Hyper-targeted porch marketing illustrates a newly possible business workflow. A frontier model can combine mapping, sunlight and temperature analysis, spatial reasoning, three-dimensional property models, customized visuals, and address data to identify prospects and propose relevant covered-porch installations.
• Successful frontier experiments can migrate into cheaper production pipelines. After a complex idea is proven, simpler steps such as merging an image into a mailer may not require the frontier model, allowing organizations to preserve novelty while reducing recurring execution costs.

Questions & Answers

Q: Why do cheap AI models match expensive models on routine tasks?

Cheap models can match expensive models when the assignment is familiar, clearly defined, and already widely understood, such as implementing a specified feature. The transcript argues that this result reflects the nature of the task, not a universal equivalence between models. Models have converged most strongly on work that many people already know how to request, so paying more may not improve ordinary output.

Q: When should companies use cheap AI models?

Companies should use cheap models for daily execution, repeatable production work, and well-defined tasks where testing shows that the output is acceptable. This approach controls costs without sacrificing useful quality. The mistake is treating inexpensive execution as the entire AI strategy. Cheap models should form the production layer, while harder exploratory questions receive targeted frontier-model attention.

Q: When is a frontier AI model worth the higher cost?

A frontier model is worth considering when the task involves difficult reasoning, uncertain feasibility, or a combination of capabilities that cheaper models cannot handle. In the transcript's central example, a two-hour, $40 frontier-model run optimized complex systems code beyond what Mitchell Hashimoto believed he could achieve alone. The potential value came from attempting a problem that was not routine.

Q: What limits the value a person can get from AI?

The transcript identifies the user's list of imaginable tasks as a major limit on AI value. Models can execute only work that someone thinks to request, so better prices, prompts, and tools cannot compensate for repeatedly applying AI to an unchanged task list. Greater returns come from identifying useful work that was previously impossible, impractical, or absent from existing plans.

Q: How can someone develop imagination for new AI use cases?

Imagination for AI use cases develops through sustained, hands-on interaction with capable models. The transcript argues that users cannot readily imagine applications for capabilities they have only read about. Repeated experimentation creates an instinctive understanding of the current capability boundary, helping experts notice difficult questions that have recently become feasible and are not yet included in standard workflows.

Q: What is the two-layer strategy for using AI models?

The two-layer strategy assigns inexpensive models to routine execution and frontier models to scouting, experimentation, and demanding reasoning. The execution layer keeps recurring work economical. The frontier layer tests questions that could redefine what should be built or automated. These layers complement each other because novel workflows can be discovered with stronger models and later transferred partly to cheaper systems.

Q: How could AI create hyper-targeted porch marketing?

The example combines Google Maps, geographic screening, sunlight exposure, summer temperature thresholds, three-dimensional structural models, customized visuals, and address data. The system identifies unshaded porches receiving all-day daylight in sufficiently hot areas, then creates personalized cards showing how a covered porch could look. The value comes from combining spatial and logical analysis into a specific business workflow.

Q: How can a frontier AI workflow become cheaper over time?

A frontier model can first prototype the overall idea, solve the hardest reasoning problems, and demonstrate that the workflow is viable. Afterward, the process can be separated into components and routine steps can move to cheaper models. In the porch example, difficult sunlight and spatial analysis may require frontier capability, while a simpler task such as merging an image into a mailer may not.

Summary & Key Takeaways

• Routine AI execution is rapidly becoming a commodity because inexpensive models can produce acceptable results on familiar, well-defined tasks. Model routing remains useful for controlling costs, but it is becoming standard practice. A durable advantage therefore requires more than completing an existing backlog faster or selecting the lowest-priced adequate model.

• Mitchell Hashimoto reportedly found that several models performed similarly on ordinary implementation work, yet a frontier model produced exceptional results when optimizing difficult systems code. That valuable assignment was not already documented or prioritized. It emerged because an experienced engineer recognized that a newly capable model might accomplish something previously beyond reach.

• A practical AI strategy has two layers: economical models handle repeatable production, while frontier models support scouting, experimentation, and difficult reasoning. Through repeated hands-on use, people learn where capabilities have moved. Successful frontier experiments can then be decomposed into workflows whose simpler components are transferred to cheaper models for ongoing execution.