Engineering Over Brute Force in AI
An analysis of why AI product success depends on rigorous engineering and evaluation rather than raw model capability. The discussion highlights the shift from brute-force compute to structured systems, the economic dynamics of open-source versus closed-source models, and the critical role of evals in managing non-deterministic AI systems.
The Shift from Brute Force to Engineering
The AI industry is currently dominated by a brute-force approach, where frontier labs throw massive amounts of compute and data at training runs to create more capable models. However, this strategy is reaching its limits. As Anchor Goyal, founder of Brain Trust, argues, the real differentiator in AI products is not the model itself, but the engineering surrounding it. Companies that succeed are those that build robust evaluation frameworks, feedback loops, and testing harnesses to manage the non-deterministic nature of AI systems.
Evals as the New Product Requirement
Evals are not just a testing tool; they are the scientific method applied to AI. They allow teams to form hypotheses about model behavior, test them against real-world data, and iterate based on quantitative and qualitative results. This approach ensures that AI systems are reliable, predictable, and aligned with business objectives. Without rigorous evals, teams are essentially guessing, leading to products that fail in production.
The Economics of Open Source Models
A significant trend is the rise of open-source models, particularly from Chinese labs, which offer comparable performance to closed-source models at a fraction of the cost. For high-volume, stable use cases, these models provide a clear economic advantage. However, adoption is limited by infrastructure challenges, such as rate limits and API reliability. As the industry matures, we can expect a shift toward open-source models for cost-sensitive applications, while closed-source models remain dominant for cutting-edge, high-stakes tasks.
The Role of Systems Thinking
AI is a systems problem, not just a model problem. The success of AI products depends on how well the system is engineered, including data preparation, context management, and state handling. Structured tools like SQL outperform brute-force approaches like bash in agent benchmarks, highlighting the importance of designing the right environment for the model. This systems thinking is essential for scaling AI in enterprise environments, where reliability and consistency are paramount.
Conclusion
The future of AI lies in engineering, not just in building bigger models. Companies that invest in robust evaluation frameworks, structured data access, and systems design will be the ones to succeed. As capital constraints tighten, the industry will shift from brute force to efficiency, making engineering the key competitive advantage.
Key insights
-
The success of AI products is determined by the engineering surrounding the model, not the model's raw capability. Evals, feedback loops, and testing harnesses are the critical components that ensure reliability and performance.
Impact: Companies that prioritize engineering over model selection will build more reliable and scalable AI products, gaining a competitive edge in the market.
-
Open-source models, particularly from Chinese labs, offer significant cost advantages for high-volume, stable use cases. These models can achieve comparable performance to closed-source models at a fraction of the cost.
Impact: Adopting open-source models can improve margins and reduce costs, allowing companies to scale AI applications more efficiently.
-
Structured data access, such as SQL, outperforms brute-force approaches like bash in agent benchmarks. This highlights the importance of designing the right environment for the model.
Impact: Engineering the right data access layer can significantly improve agent performance and reduce errors, leading to more reliable AI systems.
-
The brute-force approach to AI, where frontier labs throw massive amounts of compute at training runs, is reaching its limits. As capital growth stalls, the industry will shift toward engineering efficiency and optimization.
Impact: Companies that invest in engineering efficiency will be better positioned to adapt to the shifting landscape, reducing reliance on capital-intensive model training.
-
Strong typing and declarative state management are essential for building reliable AI systems. These systems approaches reduce errors and improve maintainability, especially in enterprise environments.
Impact: Implementing strong typing and state management can improve the reliability and scalability of AI systems, making them more suitable for enterprise use.
Action items
-
Implement a rigorous evaluation framework to test AI models and systems. Use evals to form hypotheses, test them against real-world data, and iterate based on results.
Impact: This will ensure that AI systems are reliable, predictable, and aligned with business objectives, reducing the risk of failure in production.
-
Evaluate open-source models for high-volume, stable use cases. Compare their performance and cost against closed-source models to identify opportunities for cost savings.
Impact: Adopting open-source models can improve margins and reduce costs, allowing companies to scale AI applications more efficiently.
-
Design the right data access layer for AI agents. Use structured tools like SQL instead of brute-force approaches like bash to improve performance and reduce errors.
Impact: Engineering the right data access layer can significantly improve agent performance and reduce errors, leading to more reliable AI systems.
-
Invest in engineering efficiency and optimization. Focus on building robust feedback loops and testing harnesses to manage the non-deterministic nature of AI systems.
Impact: This will reduce reliance on capital-intensive model training and improve the scalability and reliability of AI products.
-
Implement strong typing and declarative state management in AI systems. Use these systems approaches to reduce errors and improve maintainability, especially in enterprise environments.
Impact: This will improve the reliability and scalability of AI systems, making them more suitable for enterprise use and reducing the risk of failure.
Quotes
“I think evals are like the scientific method applied to software engineering with you know non-deterministic systems like AI systems.”
“The thing that actually matters, and the difference between the teams that build products that you and I would say work, like cursors product or products that just feel like shitty, is all the engineering work that goes around it to make sure that you can build a feedback loop from what's happening in production to what you're actually testing.”
“GLM five is three times cheaper than Sonnet. And then on their inference, Minimax's inference, mini uh M2.5 is three times cheaper than GLM5.”