AI 10x-ing Math
GeeksOnSkates wrote such a good set of questions, it’s worth a whole new post to address.
>> I like how you broke it down in terms of math. Like let's say coding had levels like RPGs, and you're a level 1 coder (I'm sure you are much higher than that, but I'm just following your logic here). Then let's say I'm a level 10 coder. If AI truly is a "multiplier", then you using AI and me not using AI puts us at the same level (10). The question is, is AI really a 10 * you speed multiplier?
First, let me commend you for asking legitimate questions and not just having a hot take. I can see how my 10x math could lead to conclusions I didn’t mean.
Here’s what I meant - everyone can benefit from using GenAI to code. But, the gap actually increases the better you already are at coding (not just coding but the whole software development stack including planning and architecture). I’m not suggesting that giving a 1x coder GenAI they become as good as a native 10x developer. They have the opportunity to be more productive and quicker than without GenAI tools. But - this is perilously close to “I don’t know code at all, and I vibe coded an app”. Which happens. And — it’s where the AI slop comes from.
I’m saying - don’t look at the person on the bottom of the rung of software development who’s using AI tools to form your opinion of the value of these tools.
When the 10x experienced and highly talented coder takes the time to learn these tools, figures out the optimal workflows, experiences anti-patterns and develops workarounds. LOOK OUT - we are talking about 10x-ing someone who was already 10x so you get 100x.
It’s not meant to be “math true”. It’s meant to talk about scaling the intelligent, talented and diligent practitioner. Such people are already posting their experiences on X. I just watched the Primeagen Youtuber guy so many of us love and he’s hated GenAI from the beginning. But he has a role for it now. He still hand codes everything that’s important to him, but because of GenAI he can do so many more projects that would never be done because they aren’t important enough for his time and best hand coding efforts. But the tools enable him to get these projects to life. Others like DHH who has created Ruby on Rails, Omarchy, Bandcamp and more — a very talented coder, is full on board and he too didn’t like it at first.
I’m saying - expose yourself to THESE people to help form your opinion and expectation. And while they are influencers with media channels, there are a LOT of people who have already climbed the “learn to code” to elite levels who are now climbing the “use GenAI in elite ways” and THEY are the ones that you will be competing for jobs with.
>> I ask because the difference between slop and not-slop often comes down to the time you spend reviewing code.
Another excellent point. While AI can also help in code review - and it takes time to learn how to do this most effectively - the increased output of GenAI means MORE time needed to test and approve the code. The daily life of someone using GenAI to code will have a lot more time in planning and architecting, and in testing and evaluation because you are doing more work and the human coding part will have shrunk considerably.
But there are already commercial offerings to help with this. It’s a known issue and can be a failure mode for sure. But guess what - poor testing, poor peer review - is a failure mode in software development period.
It’s kind of like the GIGO problem that GenAI also has. It’s not that it isn’t a problem, it’s just don’t hang your hat on that problem and close your eyes to how that problem is mitigated against allowing productive use of GenAI.
>> Reviewing the code is the solution. Which returns me to my original question: is AI really gonna do a "speed *= 10;" to programming work?
Reviewing the RESULT of the code for sure. With well defined tests, you can determine if the code is working without having comprehensive knowledge of every line of code. AI Slop is almost entirely the product of people who DON’T do a quality check. It’s like the lawyers submitting briefs to the court that the AI writes and the AI makes up citations. The AI is a tool. There needs to be procedures followed.
When it comes to AI coding - there is MORE testing to be done because you are creating more code. GenAI tools can help but this is the human in the loop point that should ALWAYS remain. While GenAI is collapsing the coding time it takes to create code, that’s always and forever only been PART of what needs to be done. If someone thinks “the code compiled, I see a website, ship it” - that’s a human failure.
I’m going to compress his next point. This is my simplification.
>> I’m using AI and finding it takes longer to produce code to my quality. You say it should be like putting rockets on my skates, what gives.
You are simply incompetent my friend. I’m kidding. Have you ever transitioned a “Waterfall” team to agile? Companies heard that agile was THE WAY to reduce time to market. And so they gave out mandates “we are going agile”. Some even hired Agile coaches to teach the team. And what happened? Invariably - and I mean EVERY SINGLE TIME - productivity went down. It takes time to learn the new way, to understand the WHY of the new way, to adapt the new way to your org’s particulars. It’s project 3 or more likely 5 where you see the productivity gain. Fact check me. I’m not quoting articles I’ve read, I’m giving “lessons from the trenches”.
Adopting GenAI for coding is in the EARLY adopting phase. The people seeing the “rockets on skates” are not the people who adopted it last weekend. Give yourself time. Keep tabs at least somewhat on the rapidly changing landscape (no one can be on top of all of this as it’s evolving so fast). Find the low hanging fruit and benefit from it.
You’ve seen me writing on this topic for three years. I still feel like I’m behind the curve. I spent last weekend coming up to speed on the Ralph Wiggum loop (more on that later). I found it powerful but “not fully baked” - more like a promise of the future. But I’m not alone in exploring and people are WAY ahead of my adoption and seeing more benefit.
Those who thought “if you use AI you’ll never learn anything” boggle my mind. I’m in full on self education mode. I’ve seen GREAT benefit, but I’m not personally yet at the 100x level. But I’m working to get there.
>> Or maybe I'm just lighting the wrong end of my rocket-skates. 🤣 Okay for real tho, how do you get to where it's speeding you up like that?
I’ve listened as you’ve shared your journey for some time. You are clearly open minded, you are experimenting. You are doing the things that need to be done. If you want inspiration, follow the right people on X. I might not be a credible source as I don’t ship code the whole world uses. And yet, I see time and again in my own work - a marked increase in capability, quality and velocity.
The people to listen to are those who ARE shipping real code, who are known to be able to ship quality code they wrote themselves, and are finding the place that GenAI has in their workflows.
The future of senior level development work in a GenAI world is that you, GeeksOnSkates, will be the team lead of a bunch of AI agents. It is SO MUCH like leading a human team it’s not funny. All my human skills transfer. I am the product owner, the taste maker, the systems architect, the detail plan designer, and I give my team great guidance on what we are striving for. They code but they get stuck, I get involved enough to point out where the problem lay. And then I tell them to test their own work. Don’t give me code that doesn’t even compile. You keep at it until the code compiles. Write positive and negative test cases and then code them. I shouldn’t be testing at all until you, my AI agent, have cleared testing to this level. But THEN, I am the UAT person. I am the go/no go. It’s my responsibility to ship code that works and is secure. I am the one throat to choke. BTW, I could expand this paragraph with more nuance, but you get the point if you’ve ever led teams.
