All right, so I should probably just start out by saying that this is where openAI support told me to look for information on this topic. Hopefully someone on here doesn’t mind answering questions, or dealing with complicated problems.
I’m working on Hephaestus AI, an initiative designed to help non-technical users build custom, task-specific AI systems. The central idea is to automate the development process for specialized AI solutions. For example, if multiple restaurants want to automate inventory tasks (like counting boxes), they could pool their resources, split the cost of development, and each retain full control of their own data.
However, I’ve been encountering a significant challenge I call “drift.” When an AI agent attempts to build another AI, small inaccuracies (hallucinations) tend to compound over time, leading the final product to deviate from the original vision. Metaphorically, you start with plans for a human skeleton but end up with a hybrid that’s part-human and part-gorilla. This becomes especially problematic when large amounts of code are generated, as it’s difficult to maintain a consistent structure all the way through.
Another hurdle is merging multiple task-specific AIs into a single, cohesive system. If a business creates two or three specialized AI models, it would be ideal to combine them seamlessly into one larger model without having to rip apart each block of code and reassemble everything piece by piece. I’m looking for an approach that remains flexible yet rigid enough to avoid drift, handle large-scale code generation, and allow for straightforward integration of multiple AI services.
I’m open to the possibility of open-sourcing the initial phase of this project, especially since so much will rely on API calls anyway. Ultimately, the goal is to democratize AI development so that organizations and individuals can build—and jointly fund—tools they otherwise would struggle to create on their own.
Given these challenges, I’d love insight into how best to structure an AI architect system to mitigate drift and simplify combining multiple AI modules. Any advice or resources you can suggest would be incredibly helpful.
Thank you in advance for your time and expertise!
