Large Language Models (LLMs) entered the mainstream through 2023 with the public release of ChatGPT, breaking into the public consciousness and effectively taking over the “AI” term in mainstream discourse despite being just a type of it.

Because it’s a technology that essentially “hacks” into something as innately human as language, and is explicitly designed to imitate intelligence, not to mention the fact that it can “get the computer to do things”, careful consideration must be given when adopting it as an individual, as it carries trade-offs that might not be initially clear.

I am going to be sharing the rough thought process I’ve been developing as I see the technology mature and its adoption spread to individuals around me, both professionally and personally. I’ll be using Tasks as the lens to evaluate it, as something one is trying to get done for an outcome/goal. My focus will be on individual adoption, be it on a personal or a work level.

The cost of inference, however, is something I will not be taking into account, so bear in mind that even if a task is deemed to be a perfect fit for AI, it might not be cost-effective to do so.

I will also not factor in societal and ethical concerns, but stress that AI should not be used to “think” for you, or drive you, for regardless of the massive hype thrown by parts of the tech industry, it is but a tool and should always be viewed as such.

Before discussing my thought process on how to approach the types of tasks to delegate to AI, I will call out the risk profile of the tasks, as they apply to all of them. Simply put, they should be low-to-medium-risk tasks and you should still be able to do them manually, ideally with mitigations in place in case the AI gets something wrong. Delegating critical, high-risk tasks with potentially catastrophic consequences in case of failure should be avoided.

Types of Tasks, and how to think about them

Tasks where effort doesn’t change the outcome: Delegate away

These are tasks where no meaningful value can be added to the outcome by putting in effort, or the value lies purely in the fact that it was done. A thought model for these would be ‘If the underlying systems changed, would that make a difference to the outcome? If not, getting “good” at that particular system is therefore unnecessary’.

Examples

• Activity logging
• Collating statistics/numbers from different systems

Tasks that require expertise you have: Be conscious about how and what you delegate

These are tasks where expertise you have is required for the outcome, but might be done easier or faster with AI. This type of task poses the greatest risk for a knowledge worker and should be undertaken with the mindset that leveraging AI for those will likely result in impairment or loss of the given expertise, and therefore should be done consciously.

If the expertise is no longer necessary, or could be picked up without much trouble, Adopt, letting AI take over. If it is still needed, Split the work with AI on a ratio that allows you to maintain it, but reap the benefits of augmentation.

Examples

• Adopt: Using AI to write/manage shell scripts to update/fix text files en masse.
• Split: Defining infrastructure architecture, with and without AI

Tasks that you have knowledge of, but no expertise: Delegate and guide, but know you are not the expert

Here we reach an interesting point: leveraging AI to perform a task you have some knowledge or understanding of, through a related field or exposure, but no expertise. AI here can be a great enabler, as you should be able to better guide it, catch its mistakes, and assess the outcome.

I recommend, however, being extra aware that knowledge of a related field, or base-level understanding of it, doesn’t make you an expert in the actual domain, just because you’ve successfully used AI to achieve an outcome in it.

Examples

• A project manager using AI to write code
• A software engineer using AI to deliver infrastructure

Tasks that you have no knowledge of, nor expertise in: Private experimentation only

For tasks that you have no knowledge of, I recommend being very judicious and only using AI for private experimentation or prototyping. Do not delegate tasks to AI that you can’t independently verify the outcome of.

Tasks where the work is the goal: Just don’t

This should go without saying, but if the task itself is the actual goal, don’t use AI to complete it. I shan’t bother with examples, for they should be self-evident, and if they are not, you are clearly not the audience for this text.

#ai