Working with AI
Capitalism changes what AI tooling means for engineers. Here is what I see from inside an AI-forward team.
Engineers have always built things to make our own lives easier. There is a saying that the best person to have on a team is a lazy, annoyed engineer. They find a repetitive process, fix it, ask for no kudos, and quietly make everyone around them more efficient. Sometimes that fix is glue work, sometimes it is a new tool, and sometimes it is a whole service that automates something nobody else wanted to think about. Most of that work goes unnoticed, and it compounds anyway. Teams get faster, companies get cheaper to run, and the person who built the thing has already moved on to the next one.
LLMs are that same pattern at a much broader scale, and the scale is what makes this cycle feel different. Capitalism has strong opinions about efficiency gains. Every company in the world is right to try to make their teams faster right now, because leaving that kind of value on the table is how you get out-shipped by someone who did not.
The trouble is that the math does not add up yet.
Take Jensen Huang at his word. He says a 500k engineer needs 300k in tokens, which is 800k per engineer per company. That number is more than a lot of senior engineers make. It sits in the range of an SDE2, or a strong SDE1. The industry is not paying that today, and I am not convinced it ever will at that rate. We are in the middle of a bet that no one has priced correctly, and most of the market is acting as if the price is already settled.
In the meantime, companies are getting handed an easy justification for downsizing. "AI will fill the gap." It will not. AI can write code, and it can make things show up on a screen, but it cannot connect the dots, come up with the idea in the first place, or be left alone to run. Not at this stage, and probably not for many years. The ceiling on what AI does well is still a human who knows what needs to be built and why.
I work at an AI-forward company, and honestly it is refreshing. We use it as a velocity booster on the stupid parts of the job. Tasks that used to eat two hours ship in ten minutes, and the redundant work that slowly wore people down is mostly gone. That is a real win, and frankly it should have been automated a decade ago.
What we do not honestly know yet is the economics. Nobody has solved per-token cost against real engineer output. Companies will claim cost savings because AI pushed up PR count or lines of code, and anybody who has spent time on a team already knows how that story ends. Put a metric in front of people and they will hit it, and then they will game it. That is not a scandal, it is just how incentives work, and no one should act surprised when it happens.
The thing I am actually worried about is entry-level jobs. The easy tickets used to be the ramp. They were where a new engineer learned how the codebase fits together, how the team really makes decisions, and how a production system behaves at 3am when the dashboards look weird and nothing matches. AI can close those tickets now, but it will not close them with the context or the learning attached. The bug still gets fixed, but the on-ramp that the next engineer was supposed to walk up quietly disappears.
So here is where I land. Five years from now, the engineers who understand how things really work are going to be worth more, not less. Knowing how systems connect, how a team actually ships, and how to tell a real problem from a gamed metric is the part AI cannot do for you. It is very good at raising the floor on raw output. It does nothing to lower the value of judgment, and if anything it makes good judgment the scarce thing. That is a trade I am happy to take.