Benefits for architecture firms to adopt AI models:

Information Retrieval: Generative AI could access and utilize archived drawings to assist with project documentation and offer high-fidelity exploration/solutions with AI generated details.

Project Throughput: Fee burns can be limited by reducing unnecessary feedback loops. Post occupancy services could also be offered with customized dashboards and services.

Competitive Advantage: Employing Generative AI for cognitive offloading by treating human cognition as a scarce resource, reserving it for higher priority situations. Gen AI will also provide decision enhancement, human-level judgment, creativity and appraisal.

Services and Products: Tools used internally are already minimum viable products that can go-to-market and be passive sources of revenue and marketing. Product sales could offset downturns within the AEC industry.

Drop-in Tools: Pursuing in-house generative AI tools will yield new products throughout the CI/CD pipeline which can be integrated into operational workflows. Smaller packages that are paired with existing tools that can easily be adopted with very little training.

Core needs to integrate:

Transformer Architecture: Multimodal Learning will require transformer architecture to process text and image token embeddings in parallel either through Encoder-only, Encoder-Decoder or Decoder only models, i.e. KOSMOS-2.5. Transformer architecture will need to be developed in parallel with Vector Databases.

Vector Database: Pre-learning can use existing data that is publicly available. To prevent AI hallucinations and be firm-specific, fine tuning will require archived drawings to be structured and indexed into vector databases that store positional embeddings for fast retrieval, ie Rotary Positional Embeddings (RoPE) is an option to increase compute efficiency and reduce hardware demands.

Model Training: Memory requirements to train foundation models are determined by the KV cache (Key-Value Cache) within the Self-Attention Layer. Due to the large amounts of tokens that will need to be decoded while processing architectural drawings, memory usage is expected to be high. Training on 3rd party Cloud/GPU servers would be recommended.

The flexibility and advantage of DIYers :

The explosion of publicly available AI tools and tutorials makes everything accessible. A quick stop at YouTube or TikTok and anyone can design their own renovation project. This would undoubtedly save time and money!

Imagine shopping for a home and wanting see what a kitchen and bath renovation would look like. Generative AI allows you to create designs on-demand to see what options are available. Or if you don’t know exactly what you want, go through a quick generative design exercise and AI may be able to spark ideas.

By embracing Generative AI as a Do It Yourselfer, you would be taking the procedural steps usually associated when hiring professional design firms and making the process fit around your schedule and budget.