LIMITATIONS OF LARGE LANGUAGE MODELS IN ENGINEERING DESIGN

In 2023, services like ChatGPT, Gemini, Copilot, Stable Diffusion, Dall.e and many others drew the attention of both the general public as well as the engineers and scientists to artificial intelligence (AI) and in particular to large language model (LLM) branch of it. Even the AI community was caught by surprise how well those services performed. The success gave rise to many speculations as to how these systems could assist in creative engineering design.

The discussion departs from the ontological foundations of what exists on “stage” in engineering and architectural design and what is happening on that “stage”; thus examining the “what exists” and “what happens” when we build. Since Aristotle, a well-established ontological foundation consists of a trinity of (1) real world stuff, (2) ideas in the mind, and (3) symbols that is used to communicate the ideas we might have about real world things.

The first limitation of LLM-based AI is that its world is symbolic. It consists of a huge array representing words, that is symbols, and is totally blind to two other apexes of the Aristotle’s semiotic triangle – real world “stuff” and ideas. The latter become words or drawing only after we want to communicate but not while we think and solve problems.

The second theoretical foundation is the Form-Function-Behavior model of designing. It sees designing as a process in which we define the form so that the product would have the required function and exhibit the expected behavior. Humans learn designing by reducing experiences into abstract theoretical constructs or by learning theoretical constructs directly. After having learned the abstract ideas, designer creates specific instances and he does this iteratively, to satisfy behavioral and functional requirements. For example, we have an abstract knowledge of beams, and we bridge a river by such and such reinforced concrete structure.

The second limitation of LLM based AI is therefore that it does not draw its intelligence from a capability to “reduce” experiences into abstract theories and specify abstractions into something concrete but “solves problems” intuitively. The third limitation then is that while humans explain correctness by referring to abstractions, AI is not able to so, because it was not deriving a solution from an abstraction.

LLM based AI is very good at continuing a pattern that is given as a prompt and can create texts and pictures with ease. As engineering drawings, both DWG and IFC, are in fact text, it can very well create CAD and BIM models and we will be seeing plenty of research down that line. However, these models will not be based on any deeper insights; they will be just good looking patterns. This works well for creating essays, summaries, or even building creative architectural ideas. It does not work for engineering design which is (1) an iterative process that (2) constantly moves back and forth on a reduction/induction, abstraction/specification axes. Abstractions and specifications happen on conceptual rather than on a symbolic level even though they are communicated with symbols. However, it is exclusively the symbolic at which LLM based AI operates.
This is why, we argue, engineering work will be harder to automate using AI than artistic architectural design, not to mention plain language based tasks in law and organization. This runs contrary to the long standing assumption that it is the architects that are doing the difficult creative work and engineers just making sure buildings do not crumble under their own weight.