Splash columnists collide as Kris Kosmala takes on Pierre Aury on all things AI.
I read an interesting argument this week in the context of the usability of artificial intelligence (AI) in the business of shipping. What started as an interesting rant about ChatGPT, quickly turned to broader discussion of how much intelligence, artificial or otherwise, a shipping company would need to perform their business. Mind you, it took on specifics of spot market and commenced with the argument that “shipping is essentially an immensely simple activity with the market going down if there are two ships for one cargo and going up when there are two cargoes for one ship”. It concluded that the decision to fix or not to fix an agreement in such situations doesn’t need AI.
Arguments like that make a good headline but mislead. I will come back to ChatGPT later. For now, let’s think why AI will force a rethink of many businesses and operational processes performed by any shipping company.
To fix or not to fix is not the question
What is the more likely question? How about this one: “How much profit will I obtain for my ship making six trips a year most likely consisting of X NM of laden and Y NM of ballast legs, most likely calling on these six specific ports. While sailing according to those assumed values, the ship should remain within the starting CII band at the end of the year. Assume that there will be no detrimental technical problems paralysing operations of the ship in that year.”
Answering such a question can easily perplex any employee of a shipping company. The sheer number of constraints (informed by internal and external data) and complexity of dependencies between decisions require decision support technology ideally based on AI in lieu of human decision. Why? Because human decisions made by the chartering department, operations department and technical services department with respect to just a single ship over the course of 12 months will not consider maximising the common objective of the company, while also rationalising between objectives of the three teams. Now extend the same for the period of 10 years during which the company will keep that ship on their books. Daunting, isn’t it?
Back to ChatGPT
A recent study by Goldman Sachs indicated that advances in artificial intelligence, particularly generative AI systems like ChatGPT, may significantly transform employees’ productivity and employment landscapes in many industries. If we equate AI with ChatGPT, and we should not, are there situations in shipping that qualify as being so complex that an application of AI, like ChatGPT, will do much better than a normal human brain? It is a wrong question on many levels.
The best way to think about ChatGPT is to think about Google Maps and Geographic Information System (GIS) combined with Global Positioning System (GPS). ChatGPT is like Google Maps, one of the thousands of applications made possible by leveraging underlying technologies. Therefore, one should ask whether GPT (Generative Pre-trained Transformer) technology, an autoregressive language model that uses deep learning (a branch of AI) to produce human-like text, is useful to a shipping company.
The not-so-science-fiction answer is yes. What is needed is to “teach” the model the language of shipping using the database of a shipping company and augment the internal training data with external data sets for industry business terms. Combining this GPT model with optimisation algorithms utilising a company’s data would create a very skilled virtual worker. The shipping company could realistically start thinking of reducing the number of people in chartering, voyage operations, fleet management, and back office or significantly increase the number of ships in their fleet without increasing headcount in chartering, voyage operations and fleet management. The overarching objective of the company would remain the same: maximise annual profits from all assets and operations.
So what does it take to train such a digital employee? Well, the current and maybe next generation of the human workforce in shipping should not lose their sleep over being replaced with a machine. Training an internal GPT model is a significant expense, a few scrubbers worth of an expense. Even if the dataset is free, there are costs associated with cleaning and processing the data, and those costs can range from hundreds of thousands to millions of dollars. As with any employee, digital workers need frequent skillset review. A digital chartering employee service bot, for example, may need to be fine-tuned every so often. What’s expensive is that the company needs to keep doing it and keep testing the model to make sure it’s doing what it is expected to do. Eventually though, the computing costs will fall, the AI smarts to cleanse the data drop, and current workflows designed around human knowledge and human decision limitations could be safely blown up and replaced with AI.
Will the business of running a shipping company change as a result? Time will tell. But don’t cancel your company’s Christmas party for your human workforce just yet.
You equate complexity with the decision with the need for intervention from AI… Shouldn’t it be the opposite? Ie if a situation is simple, AI can answer it. How can AI help answering any questions about how to maximise profit when previous datasets are always out of date for future decision making? In this domain I think it is useless to think of AI or humans – rather different scenarios which could be generated by AI should inform complex decision making – but of course this is to be based on “upside/downside” and risk mitigation rather than trying to pretend that AI can give any certainty about the future and profit maximisation on its own.
Oh, and as an addendum – such models need to be fed by skilled people with domain knowledge in chartering, technical ops, engineering, finance, etc. The feed should be a mix of contextual knowledge and quantifiable knowledge. The competitive advantage of any company could be driven more by how well the employees are managing to complete their modelling with the correct variables, (“correct” here can often be EXTREMELY complex to understand – hence this is a constant operation rather than simply designing something and letting it run) – but more importantly translate this in to some action at the right points. In that case the need for some kind of multidisciplinary congruence will be much more important than what is often the case in the industry today, where different disciplines are often working at cross-purposes.
I suggest there is 1 major limit – its not gong to get our stuff landing on time is it?? 🙂
profitability often leaps when customer service or reliability falls. AI Might help make a decision based on input factors you specify, but its not gong to fix both ends of a balancing act.
Lets say with up to 12000 shipments loaded on a single container boat, AI isn’t’ going to fix the issue that some of those customers wont’ be happy whatever you do?