I think todays LLMs have a very non-unintuive effect on the information worker job market.
LLM models make "deisgn taste" both in product design, and software achitecture massively more valuable, becuase they massively reduce the cost of everything else, while doing comparatively little for product-design or software-architecture-design.
They can make a pretty button, sure. They can implement a clean version of an algorithm from a research paper. But they can't (as of yet) decide what software to build, why you should build it, what parts of the software to build for flexibility in the future vs practicality now. What algorithms will be the best fit for the real world needs, none of it.
And what they do in those areas is worse than not-knowing. They confidently expound garbage. Absolute drivel. And I laugh. As someone who is a lifelong programmer and entrepreneur, selling software when i was in high-school, selling a company to Google when I was 27 years old, and currently collaborating with AI like a madman to produce lots of useful software, I'm here to say that 2-4-8x productivity is not good, it's just the start. And the next tier happens when you have someone better at software development than the AI keeping it in check, micromanaging it if you will.
I'm currently hovering between 40-50x productivity (on a LOC basis, which I know is bad, but it's the only objective mesure I have), over 16+ weeks, with Opus 4.6 (nothinking!) and now Fable (low thinking). I get there not doing a few things differently:
1. I frontload the design-time thinking and elaboration. I absolutely use agents in design (usually 2 of them, plus me). However, I don't let agents design my software. I use them as researchers, as sounding boards, then I make the call when I see how my goals fit in. Then I have them write a design spec, and review it, and I review it; then i have them write a "concise phazed bulleted checkbox implementation plan" then i review that, and once it seems coherent, I put that into a new session that has VERY LOW thinkikng budget, so the model can stop second guessing everything we've already pinned down in the thinking/design phase, and just implement it.
2. I stay connected and read most of the verbal chat output and careful code details to see the big picture... No trees of subagents. I watch and manage every agent chat, all the time. 4-6 simultaneous windows of agent chats (in my own self authored harness) that I watch like a hawk, because they type 100x faster than I do, but they are much worse coders and so I have to catch their mistakes before they compound and eat the codebase. (This is a large interconnected codebase, not a pile of isolated separate things)
3. I design the process to get from A to B... the ordering of the feature development, just like I would do myself, to minimize architecture risk, get the unknowns out of the way, feel the thing taking shape before the complexity layers in..Then build harnesses with as much AI visible iteration loop as is practical for the part we are working on. I use third-party oracle references when I can (like pixel correct font and shape tests against Skia, instead of just eyeballing) And I decide the level of detail where this is useful, then put the LLM in a place where it can succeed.
Good to see Mustafa Suleyman striving to prove that he's much of an idiot as he is a worthless person.
https://www.businessinsider.com/deepmind-mustafa-suleyman-go...
I think todays LLMs have a very non-unintuive effect on the information worker job market.
LLM models make "deisgn taste" both in product design, and software achitecture massively more valuable, becuase they massively reduce the cost of everything else, while doing comparatively little for product-design or software-architecture-design.
They can make a pretty button, sure. They can implement a clean version of an algorithm from a research paper. But they can't (as of yet) decide what software to build, why you should build it, what parts of the software to build for flexibility in the future vs practicality now. What algorithms will be the best fit for the real world needs, none of it.
And what they do in those areas is worse than not-knowing. They confidently expound garbage. Absolute drivel. And I laugh. As someone who is a lifelong programmer and entrepreneur, selling software when i was in high-school, selling a company to Google when I was 27 years old, and currently collaborating with AI like a madman to produce lots of useful software, I'm here to say that 2-4-8x productivity is not good, it's just the start. And the next tier happens when you have someone better at software development than the AI keeping it in check, micromanaging it if you will.
I'm currently hovering between 40-50x productivity (on a LOC basis, which I know is bad, but it's the only objective mesure I have), over 16+ weeks, with Opus 4.6 (nothinking!) and now Fable (low thinking). I get there not doing a few things differently:
1. I frontload the design-time thinking and elaboration. I absolutely use agents in design (usually 2 of them, plus me). However, I don't let agents design my software. I use them as researchers, as sounding boards, then I make the call when I see how my goals fit in. Then I have them write a design spec, and review it, and I review it; then i have them write a "concise phazed bulleted checkbox implementation plan" then i review that, and once it seems coherent, I put that into a new session that has VERY LOW thinkikng budget, so the model can stop second guessing everything we've already pinned down in the thinking/design phase, and just implement it.
2. I stay connected and read most of the verbal chat output and careful code details to see the big picture... No trees of subagents. I watch and manage every agent chat, all the time. 4-6 simultaneous windows of agent chats (in my own self authored harness) that I watch like a hawk, because they type 100x faster than I do, but they are much worse coders and so I have to catch their mistakes before they compound and eat the codebase. (This is a large interconnected codebase, not a pile of isolated separate things)
3. I design the process to get from A to B... the ordering of the feature development, just like I would do myself, to minimize architecture risk, get the unknowns out of the way, feel the thing taking shape before the complexity layers in..Then build harnesses with as much AI visible iteration loop as is practical for the part we are working on. I use third-party oracle references when I can (like pixel correct font and shape tests against Skia, instead of just eyeballing) And I decide the level of detail where this is useful, then put the LLM in a place where it can succeed.