LLMs produce proper English and offer reasonable follow-ups. Funnily enough, recent models are less good at this, but a good harness and skillset can do wonders.
This is in contrast with my colleagues from xxxxx [0], who appear to be basically immune to learning to read, write, speak, or understand English any reasonably, despite working at a multinational company explicitly requiring so. They also keep insisting on pointless meetings where I get to tell them the exact same things I already told them in writing, except now it is across both of our accents, and through their terrible microphone and internet connection. Mind you, I do listen myself back sometimes, so this is not me being gleefully self-unaware either.
Nothing against them personally, but I truly despise working with these people. They have zero edge over even something like Luna, and will 100% be replaced wholesale once management stops pretending, unless they start using LLMs hard. Something they're unable to do, for the same reasons they're terrible to work with. [1]
And there's no amount of doorman fallacy invocations that will save this. They're cheap, sure, but again, even the cheap models meet or beat them. They've been meat proxies even before LLMs, and the amount of instructions they need to be given will make even the dumbest LLM fly through things. Even though they're hired in swarms, not unlike how people try to save terrible LLMs, that only takes you so far. They don't grow, or grow at glacial speeds. They're cooked.
If you can offload a task to Sonnet, you met them in price, and beat them dozens of times over in quality. Very inhumane way to put it, but it's the reality already. And the bar is rising.
[0] Censored at will because this is not a race thing, but bottom of the barrel HR practices and market realities.
[1] In true terrible management fashion though, it is entirely possible that they'll simply fire the more capable and more expensive people instead, falsely expecting that these other folks will wonderfully use LLMs to just figure things out. Now that will be quite the fireworks show!
IMHO neither deep codebase familiarity nor communication skills are durable advantages. Arguably they already aren't. Claudish is a problem for Claude but I've seen no signs of equivalent problems with GPT 5.6. It communicates clearly. The ability to find things and use the existing codebase well is largely dependent on how well structured it already is. I've mostly used AIs with codebases that I wrote myself and they're well structured. Some have plenty of internal dev-facing documentation too. I'm often surprised at how well the models use the internal abstractions - easily as well as I would have. Except over time I'll forget the details of codebases I don't work deeply on, whereas the model is rediscovering each time by using its far superior reading speed, so its ability to use the details won't degrade and mine will.
I currently do add value by guiding the models when they overlook a better way to do things. But it feels like writing code by hand is going to go the same way as writing assembly language by hand already did. There will be rare cases where it's necessary for some reason, but coding will steadily become seen as some sort of dark wizardry only a handful of old codgers know. Eventually the art will be lost entirely, a bit like how military archery and spoken Latin were skills Roman warriors took for granted but today ~nobody can do them.
So where's the durable advantage? I see a few candidates:
1. AI bubble pops and the rate at which the big model firms ship features takes a dive. Working around model/tooling limitations and optimizing costs ends up becoming a source of sustained value.
2. Business change consulting. Models are passive, so you need to think of a question to ask. Someone actively using their technical knowledge to scout out opportunities and come up with creative solutions might have value for a while, not because AI couldn't come up with these ideas given the right prompt but because it just won't be given the right prompt without you.
3. Building "AI native" companies. Big technology transitions often leave existing institutions behind because they can't adapt, either culturally or in terms of skills. The internet was one example of that - companies like Amazon should theoretically have had no chance given the existence of many highly competitive retail firms - and the inability of Swiss watch companies to deliver smartwatches is another specific example. So it's possible that AI will be like that and fully absorbing the consequences won't be possible for existing firms with large labour forces, opening up an opportunity for a wide range of new startups to come in and take shares of mature markets.
A lot of engineers are gonna hate this, but the domain that LLMs will have the hardest time dominating is people skills
And when they do dominate that we’re going extinct within 10 years so it’s a good bet to work on those.
Romance chatbots say what?
LLMs produce proper English and offer reasonable follow-ups. Funnily enough, recent models are less good at this, but a good harness and skillset can do wonders.
This is in contrast with my colleagues from xxxxx [0], who appear to be basically immune to learning to read, write, speak, or understand English any reasonably, despite working at a multinational company explicitly requiring so. They also keep insisting on pointless meetings where I get to tell them the exact same things I already told them in writing, except now it is across both of our accents, and through their terrible microphone and internet connection. Mind you, I do listen myself back sometimes, so this is not me being gleefully self-unaware either.
Nothing against them personally, but I truly despise working with these people. They have zero edge over even something like Luna, and will 100% be replaced wholesale once management stops pretending, unless they start using LLMs hard. Something they're unable to do, for the same reasons they're terrible to work with. [1]
And there's no amount of doorman fallacy invocations that will save this. They're cheap, sure, but again, even the cheap models meet or beat them. They've been meat proxies even before LLMs, and the amount of instructions they need to be given will make even the dumbest LLM fly through things. Even though they're hired in swarms, not unlike how people try to save terrible LLMs, that only takes you so far. They don't grow, or grow at glacial speeds. They're cooked.
If you can offload a task to Sonnet, you met them in price, and beat them dozens of times over in quality. Very inhumane way to put it, but it's the reality already. And the bar is rising.
[0] Censored at will because this is not a race thing, but bottom of the barrel HR practices and market realities.
[1] In true terrible management fashion though, it is entirely possible that they'll simply fire the more capable and more expensive people instead, falsely expecting that these other folks will wonderfully use LLMs to just figure things out. Now that will be quite the fireworks show!
> You have to beat the models at something
And for the rest of your career* you'll be running on a treadmill trying to stay ahead of them!
(* probably like two years if you're one of the lucky ones)
Finally someone is asking the right questions.
and then teach ai how to beat you
IMHO neither deep codebase familiarity nor communication skills are durable advantages. Arguably they already aren't. Claudish is a problem for Claude but I've seen no signs of equivalent problems with GPT 5.6. It communicates clearly. The ability to find things and use the existing codebase well is largely dependent on how well structured it already is. I've mostly used AIs with codebases that I wrote myself and they're well structured. Some have plenty of internal dev-facing documentation too. I'm often surprised at how well the models use the internal abstractions - easily as well as I would have. Except over time I'll forget the details of codebases I don't work deeply on, whereas the model is rediscovering each time by using its far superior reading speed, so its ability to use the details won't degrade and mine will.
I currently do add value by guiding the models when they overlook a better way to do things. But it feels like writing code by hand is going to go the same way as writing assembly language by hand already did. There will be rare cases where it's necessary for some reason, but coding will steadily become seen as some sort of dark wizardry only a handful of old codgers know. Eventually the art will be lost entirely, a bit like how military archery and spoken Latin were skills Roman warriors took for granted but today ~nobody can do them.
So where's the durable advantage? I see a few candidates:
1. AI bubble pops and the rate at which the big model firms ship features takes a dive. Working around model/tooling limitations and optimizing costs ends up becoming a source of sustained value.
2. Business change consulting. Models are passive, so you need to think of a question to ask. Someone actively using their technical knowledge to scout out opportunities and come up with creative solutions might have value for a while, not because AI couldn't come up with these ideas given the right prompt but because it just won't be given the right prompt without you.
3. Building "AI native" companies. Big technology transitions often leave existing institutions behind because they can't adapt, either culturally or in terms of skills. The internet was one example of that - companies like Amazon should theoretically have had no chance given the existence of many highly competitive retail firms - and the inability of Swiss watch companies to deliver smartwatches is another specific example. So it's possible that AI will be like that and fully absorbing the consequences won't be possible for existing firms with large labour forces, opening up an opportunity for a wide range of new startups to come in and take shares of mature markets.
None of these are coding.
[dead]