Deterministic Software + Probabilistic Intelligence
What I learned after trying to make the combination scale. I thought I had found a simple formula. Deterministic software handles what we know how to do. Probabilistic intelligence handles what requires judgment, interpretation, and adaptation. Put them together, and suddenly software can do things that previously required a person. For the past month or so, I've been testing that idea in…
Deterministic software excels at tasks with clear procedures, while probabilistic intelligence shines in situations requiring judgment and interpretation. Combining the two could enable software to perform tasks that previously required human intervention. The author tested this concept by building tools that automated predictable transformations and utilized large language models (LLMs) for interpretation. Over time, the tools evolved into software with departments, responsibilities, and specialized AI workers.
However, the author encountered a limitation when pushing the system too far. The probabilistic component of the AI system is not free, and giving it more context doesn't guarantee correct usage. The author experimented with task-driven development, an MCP server for ChatGPT conversations, and sub-agents to improve context management.
These improvements revealed more issues, including the computational cost of understanding information and the excessive token usage when agents performed extensive reasoning and communication.
The author concludes that a powerful system is achieved by combining cheap certainty (deterministic software) with expensive uncertainty (probabilistic intelligence). However, the cost of state reconstruction and inference must be considered. The key insight is to determine what actually needs intelligence and avoid spending inference on deterministic tasks.
When a probabilistic system discovers a procedure that could be turned into deterministic software, it may be worth optimizing that process instead of relying solely on inference.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
