Skill Acquisition

Some time ago, I stum­bled upon the Dreyfus Model of Skill Acquisition. It de­scribes the stages through which a stu­dent moves on the road to mas­tery of a cer­tain skill. That skill could be bak­ing, fenc­ing, pro­gram­ming, or even teach­ing.

I think that it stands to be a well-suited frame­work through which we can an­a­lyze cer­tain as­pects of the al­ways-evolv­ing pro­fes­sion of soft­ware en­gi­neer­ing.

The Model

The Dreyfus Model of Skill Acquisition breaks the process of learn­ing a skill down into six stages. My in­tu­ition is that these stages are not dis­crete blocks of time, but rather seg­ments on a con­tin­u­ous tra­jec­tory. That is, there is­n’t a spe­cific mo­ment we can point to and say, “that’s when stage one be­comes stage two!” In­stead, there are broad seg­ments of time that com­pose each stage.

Stage One

When a stu­dent first learns a skill, they al­most al­ways learn it from a teacher or from some kind of writ­ten re­source. When this hap­pens, they are given a spe­cific set of rules they need to abide by, re­gard­less of the sit­u­a­tion.

When you are learn­ing to ride a bike, you are told to keep ped­al­ing, keep your hands on the han­dle­bars, and turn by mov­ing the han­dle­bars in the in­tended di­rec­tion. These are ex­plicit rules that the learner must abide by in or­der to con­tinue.

Stage Two

In stage two, the learner be­gins to no­tice some nu­ances in when or how to ap­ply each rule. They see that some as­pects of the skill in ques­tion can de­pend on the sit­u­a­tion.

When you are learn­ing to ride a bike, you very quickly learn that you should not be ped­al­ing at the same speed all the time. If you want to go faster, you pedal with more rigor. If you are al­ready go­ing a speed you are com­fort­able with, you ease back. This is stage two.

Stage Three

At stage three, the stu­dent is fi­nally com­pe­tent enough that the core traits of the skill can be del­e­gated to the sub­con­scious, free­ing up the con­scious mind to fo­cus on higher level analy­sis and plan­ning.

When you have mas­tered the skill of bal­anc­ing a bike and pac­ing your ped­al­ing, you can fi­nally free up your mind to fo­cus on route. You can ex­pend en­ergy on fig­ur­ing out where and how you want to get to your des­ti­na­tion, and much less about how to stay up­right. If you en­counter un­usual ter­rain, you still need to fall back to the plain rules you were taught in the be­gin­ning.

Stage Four

Stage four is much like stage three, but with very lit­tle to no con­scious ef­fort re­quired for the atom­ics of the skill, even in un­fa­mil­iar sit­u­a­tions. Learn­ers are be­com­ing more com­fort­able do­ing higher level analy­sis sub­con­sciously.

It no longer takes men­tal ef­fort to plan bike routes or to nav­i­gate com­plex ter­rain.

Stage Five

All ac­tions in­volved in the skill are now mostly un­con­scious, even high level ones. Learn­ers are able to in­vent new be­hav­iors and ad­just to new con­di­tions on the fly with­out de­lib­er­a­tion.

Bikers in this stage can per­form tricks or nav­i­gate dan­ger­ous new ter­rain with­out prac­tice or pause. If asked to per­form a new trick or choose one of a num­ber of pos­si­ble routes through a com­plex course, they can do so with­out hes­i­ta­tion. If asked what went into their de­ci­sions, they will say that they sim­ply made the ob­vi­ous choices.

Stage Six

Stage six is mas­tery.

There is no as­pect of the tra­di­tional skill that can­not be done un­con­sciously. The con­scious mind is only needed to ex­plore fron­tiers of the skill that no one has ever done be­fore. Mas­ters can com­pose two or more as­pects of the tra­di­tional skill into more ab­stract never-be­fore-seen com­plex ma­neu­vers.

Bikers in this stage are no longer bik­ing. They view their skill as a craft. They seek non­tra­di­tional ma­chin­ery and seek to trans­form the bike + man sys­tem into some­thing much grander and more beau­ti­ful.

Observations

The ob­vi­ous ob­ser­va­tion here is the con­nec­tion be­tween the Dreyfus model and the Daniel Kahneman’s Nobel Prize win­ning work on his ver­sion of dual pro­cess­ing the­ory. In essence, both au­thors be­lieve that hu­man cog­ni­tion falls squarely into two bas­kets:

1. Fast in­tu­itive think­ing.
2. Slow log­i­cal rea­son­ing.

I think this is broadly true, and it’s a lens through which soft­ware en­gi­neers rarely view skill ac­qui­si­tion. Some as­pects of our pro­fes­sion are not so rig­or­ously in­tel­lec­tual, while oth­ers are.

The Industry

The vast ma­jor­ity of the soft­ware in­dus­try to­day seems split be­tween two ex­tremes:

1. You should use as much AI as pos­si­ble. You will be able to ex­per­i­ment more quickly and learn faster about your tar­get mar­ket and tech­nol­ogy. It will al­low you to do more with less.
2. You should use as lit­tle AI as pos­si­ble. It pro­duces poor code that can­not stand the test of time while also slowly erod­ing your in­nate abil­ity to pro­duce high qual­ity soft­ware.

I think this is a false du­al­ity, typ­i­cal of a pop­u­la­tion in stage one of the Dreyfus model. Both of these rules are en­tirely con­text free and non-sit­u­a­tional. Abid­ing by them too strictly demon­strates a novice-level un­der­stand­ing of soft­ware en­gi­neer­ing.

I be­lieve more highly skilled en­gi­neers would con­sider whether there are sit­u­a­tions where they should use an LLM, and sit­u­a­tions where they should not.

Of course, the avail­able tools are chang­ing so quickly, as they al­ways do, so it is a dif­fi­cult skill to learn.

Conclusion

While I did con­nect to LLMs in this post against my bet­ter judg­ment, I think this is a topic that gets quite un­der-dis­cussed. In­stead of solely think­ing about how LLMs are de­stroy­ing your crit­i­cal think­ing skills con­sider the nu­ances of skill ac­qui­si­tion on the whole. Per­haps, con­sider whether you are ac­tively learn­ing to in­tuit as­pects of your pro­fes­sion, rather than leav­ing it all to the con­scious mind. Are there ways you can train your­self to do this? I am not sure.