LLMs are more like E-bikes. More assist makes you go faster, but provides less exercise.

In 1980, Steve Jobs spoke,

What a computer is to me is it's the most remarkable tool that we've ever come up with, and it's the equivalent of a bicycle for our minds.

I don't think that LLMs are like bicycles for the mind. They are more like E-bikes.

Greenways and E-bikes

A few nights ago, I was walking with my wife on the greenway. It's illuminated, so we were easily visible. We were absorbed in our conversation when an e-bike startled us, appearing around a curve and speeding toward us at about 25 miles per hour (~40 km/h). With just a couple of seconds to react, I pushed my wife out of danger and into the damp grass. As the two-tired vehicle passed, I turned toward its rider and yelled, "SLOW DOWN!". He muttered an unintelligible, unconcerned reply and motored on. As he disappeared from sight, I was angry and my wife was distraught.

The town where I live has excellent greenways, and I spend a lot of time running and walking on them. Seriously, it's thousands of miles each year, and over the last five years, I've noticed an increase in use of electric vehicles on our greenways: I've seen e-things of all kinds and sizes: bikes, scooters, skateboards, longboards, and unicycles.

I suppose people believe that the absence of a gas engine makes a vehicle acceptable on the greenway. The greenway is an outdoor space reserved for humans to exercise and enjoy nature. E-bikes that go too fast deter pedestrians from using it as intended.

I am not against electric vehicles on the greenway when riders use the throttle responsibly. Electric assist is a fantastic gateway to a healthy aerobic exercise regimen for those who could otherwise not even start. For example, my older colleague's e-bike gives him the confidence he needs get out in the first place.

LLMs are like E-bikes

In software engineering, LLM use is like riding an e-bike on the greenway. Appropriate use comes down to how much "throttle" you use: how much cognitive effort you offload to the bot. Engineering is an intellectual and collaborative space where human minds work hard to solve business and technical problems. When you drop a mostly LLM-generated contribution into a chat or pull request, you zoom by on your LLM-moped, disrespecting and alienating those who are engaging their neocortex without assistance.

One of the rudest encounters I've had in my career was when a former colleague contradicted me in a technical discussion with an LLM-generated reply. Their screenshot of ChatGPT was lazy and reinforced an incorrect belief. It amounted to nothing more than an appeal to authority: the AI said it, so it must be true ². The asymmetry principle ³ helps explain why it was so frustrating: my counterpart effortlessly generated the LLM rebuttal which took me time and effort to parse and refute.

At my work, the ratio of LLM-generated code in pull requests is rising. The interesting thing about these is the author often doesn't seem quite as, well, authoritative. It seems that when a robot has written the code, a higher portion of my feedback is ignored or answered unsatisfactorily. Since the point of a pull request is knowledge sharing and responding to feedback, it's a bit of a waste of time and feels more like a rubber-stamp request.

I'm not pointing fingers: In mid-2023, I completely over-relied on ChatGPT and shipped code I didn't understand to production. It took me three weeks to fix a defect that came up. Pretty embarrassing.

Summary

You can depend on LLMs too much. Use them in a way that doesn't atrophy your mental fitness and doesn't exasperate your colleagues.

References

1. What's Wrong with this Picture?
2. Appeal to Aithority
3. Brandolini's Law