[iamthemonster]

I'm an engineer in the oil and gas industry. Some of my job involves messy judgement calls that I would never involve LLMs in, but some of my job boils down to integrating different items of data that exist across documents, drawings and a few databases that are in different formats and don't cross reference each other. At times I have used LLMs as a kind of "highly enhanced search engine" to do semantic search across documentation of every different types. My alternative was opening each document and using ctrl+f, along with my intuition of knowing what document titles to search for.

For a more concrete example, I have an interface to the data that comes from every sensor on the oil processing facility. It has a built in "AI" (I try not to use that term!) but it has a feature where I ask how to process data in plain language and it'll give me the calculations, then it'll also provide a plain language summary of all the calculations I conducted. That saved me 10 hours of work.

I am a negative nancy on LLMs in general but I still passionately believe that they're a tool which every white collar employee will need to learn to use effectively.

I cringe when I hear engineers say "I didn't know the answer so I asked ChatGPT" but I also do worry that I could be significantly outperformed by another engineer with 10 years less experience in engineering and 1 year more experience in judicious use of LLMs.

[iridione]

What's an example of a messy judgement call?

[iamthemonster]

Sorry I didn't see your comment until now, apologies for the late reply.

A classic messy judgement call would be:

1. Input information includes some word of mouth info that I have no reason to doubt, but also absolutely no way of verifying against field data

2. A single piece of equipment is not functioning - the plant is reasonably safe to operate with the failure, but how safe? Are the other relevant protective layers in place and effective in the relevant scenarios?

3. If I decide to implement a really robust and good-quality solution that'll stand the test of time, will it actually take so long to implement that I would have been better off with something simpler but less robust?

4. Is my decision making process clearly communicated enough for the decision makers involved? Which installation manager is on shift?

5. If the regulator audited my decision making process would they raise a recommendation? So what if there's a 0.1% risk that they'll raise a recommendation as long as people are safe?

These kinds of thought processes are where I add value as an engineer in a way that's irreplaceable by LLMs. I just hope that LLMs can really improve how quickly I can access data to make my human decision-making better based in fact.