Two jobs we might think AI could do are trading the financial markets and reading medical imaging.
These two jobs look remarkably simliar on the surface. In the former, a trader reads a ticker tape and chooses to buy or sell. In the latter, a radiologist reads medical imaging and choosing to classify cells as benign or malignant.
We can the imagine trader or radiologist as working in a cockpit. They have monitors displaying the ticker taper or medical images. They have input devices to choose to buy or sell ticker symbols or to classify cells as benign or malignant. They sit in their cockpit, getting information from the displays and choosing which actions to pursue with their input devices.
(left) Steven A. Cohen at his trading desk at his fund, S.A.C. Capital Advisors, photographed for an interview for Vanity Fair. (right) An "operant conditioning chamber" for a pigeon, trained to detect breast cancer in a study at the University of Iowa. For more on these two specific images, go here
It's very tempting to think that we can replace traders and radiologists with LLMs. That we can just put an LLM in the cockpit and have them doing the job.
This idea of replacing the human in the cockpit with a computer is the same intuition that led to digital calculators, elevator control panels, artillery control systems, and CNC machines, replacing human calculators and elevators operators and augmenting artillery officers and machinists, respectively. Well that's interesting. Why were human calculators and elevator operators replaced, and why do we still have artillary officers and machinists? What tells us when a cockpit job will be replaced and when it will be augmented?
What were the computers unable to replace? What are machinist and artillery officers doing today? Why don't the equivalent jobs exist for human calculators and elevator operators?
Machinists need to program the CNC machines, inspect the quality of the parts, translate and clarify specifications, suggest and verify choices of materials and tooling. Artillery officers need to learn artillery tactics and strategy, understand the characteristics of their artillery, lead and coordinate their troops.
In other words, their role is now beyond the cockpit. The real world is much more complex then getting exactly correct specifications on machined parts or locations to hit with a mortar. If the world was simple like that than machinists and artillery officers would be automated away like an intelligence in a cockpit. But they aren't. They need to inspect, order, and perform maintenance to keep their machines running. They need to coordinate with and understand stakeholders, whether these be generals, enemy combatants, allied troops, clients, business interests, or coworkers. They are called upon to evaluate the purchase or design of new weapons or new machines. They need to quote probability of success or cost of manufacture.
To put it simply, the job of designing, building, and maintaining the cockpit, the job of sourcing and curating the information streams, and the job of contextualizing and communicating beyond the cockpit, is unable to be replaced by computers.
Cockpit jobs approximate some aspect of the real world and ask the intelligence inside to select from a menu of actions that can be controlled from the cockpit. When we try to automate a cockpit job we must ask:
- Who designs the cockpit? Who decides the cost-quality tradeoffs for the quality of information provided, the quality of the intelligence in the cockpit, the quality of the available inputs?
- Who builds the cockpit?
- Who maintains the cockpit?
- Who decides when it's time to do a redesign, rebuild, or maintenance?
- Who decides whether or not the job is needed, and when it's needed?
- Who flags when things goes wrongs, suggest improvements, and evaluates performance?
Who would be best suited to be the answer to these questions? It's often the intelligence in the cockpit. The person who is actually doing the job is best suited to answer questions about the jobs. It's hard to replace the intelligence in the cockpit because we need a human to be self-aware about what they are doing.
Whenever things change in the world and we need to rethink the cockpit, we need human intervention. This happens more often in more jobs then most would care to admit. After all, the existence of cockpit jobs are only a blip in our history. When a cockpit job is created, we have learned some modular job that needs to be done. As we define the scope and process of the task more and more, we been understanding what to automate. The people previously holding the cockpit job move up to managing the cockpit or out to translate their skillset to a similar but less defined job.
This evolutionary process is to be described for fully in a later part.