Ryan Moultrup

Software engineer making AI films. Thinking out loud about AI, software, and the future of work.

[Ryan Moultrup @MoultrupRyan]

The models are getting better and are now just implementing this stuff. Just from the beginning of this year until now we've seen huge improvements in how well they write software. In the beginning of the year you had to write long skills that laid out exactly what good architecture is, security, UI/UX, etc. and explain how to do it. Fast forward to now with Fable and Opus 5.5 you no longer need this. It actually can hurt the output if you keep these files around.

At the beginning of the year you treated the model like a junior or early mid-level engineer and had to babysit it and tell it how you wanted things built. The models now are like senior engineers that you collaborate with and already know this stuff. They actually know it better than you and your instruction may override the model with a worse version of how to create good software. Pretty soon the models will be principle engineers and shortly after that something beyond what any human is capable of.

[Ryan Moultrup @MoultrupRyan]

@DominickSGerard @manali_manek01 I look forward to it.

[Ryan Moultrup @MoultrupRyan]

@itz_komal01 Linux

[Ryan Moultrup @MoultrupRyan]

God created us in his image. We created AI in our image. AI is smarter and more capable than we are even though we created it. Do you think this pattern holds all the way back? Is god just a bumbling idiot compared to us?

[Ryan Moultrup @MoultrupRyan]

@Denche2005 Promptnosticators

[Ryan Moultrup @MoultrupRyan]

@Taniyatweets_ What are you building with it?

[Ryan Moultrup @MoultrupRyan]

@kayaravenya Hmmm...

[Ryan Moultrup @MoultrupRyan]

@qqb16802 WTF!

[Ryan Moultrup @MoultrupRyan]

@qqb16802 You have 5K followers. Maybe you just need to be more interesting.

[Ryan Moultrup @MoultrupRyan]

People acting like human written code never breaks in production. Not to mention ANYONE who's worked at a real tech company or on a large project at a non-tech company understands that sometimes you have to fix production issues involving code you didn't write. Often this code is far worse than what the latest models write.

Prior to AI you have to figure out how it worked under pressure. AI makes all this so much easier. Edge cases are going to happen, production is going to break if you any real users at all. AI makes it ALL easier that it used to be.

[Ryan Moultrup @MoultrupRyan]

That's a solid point. In my example I gave the extreme ends of the spectrum where in my personal experience I've learned an incredible amount because I fit in the middle ground.

Sticking with the example I am able to reject 5 of the options from experience and asked the AI to explain to me more about the 2 I was not sure about. I've spent many hours having a conversation with the AI asking questions about an approach it recommended, why it recommended that particular approach, if it can cite sources online of where that approach worked best.

This is one of the reasons I'm such a staunch advocate of AI. It feels like the most powerful tool I've ever had access to, because not only does it improve the speed and quality of my work, but it's helped me learn at a greater pace that I have on my own in the past. It does this by exposing me to new ideas and seemingly knowing everything about it. Also, it never tires of my endless questions. And because I'm talking to a machine, I don't have the human instinct to keep quiet because my question might be dumb or to have it explain the same thing to me over and over again until I deeply understand it.

[Ryan Moultrup @MoultrupRyan]

@AIandDesign "Made without AI" is the new I overpaid for something of arguably worse quality at the farmers market because some hippy made it by hand.

[Ryan Moultrup @MoultrupRyan]

Depends on what the project is. How big do you expect it to get? How much traffic will it get? How secure does it need to be? Is there critical math that needs to happen such as dealing with money? Is it distributed or monolithic? Do you plan on using AI to build it?

If you can't answer any of these question use PHP.

[Ryan Moultrup @MoultrupRyan]

Any code that AI writes must have a team of reviewers review it. Easiest if you create a skill but a simple prompt works.

"Launch three agents to independently review your code. One Opus, one fable and one Codex"

That's it. They will find the issues. You can instruct these agents to check for dead code or how the code can be simplified. I have specialized agents that do this. It all depends on your token budget.

Do you honestly review all the code when Opus writes 3000 lines and delivers a full working feature?

[Ryan Moultrup @MoultrupRyan]

Do you really think option 1 is real? There's no developer that understands every line of code in the real world. Maybe they understand every line for a small feature they built on a greenfield project but after a month of working on another feature that knowledge is lost and in most real project they have to work with code they didnt write. Spending the time to understand every line is not feasible. How do you think this works on a project any larget than 10K lines of code? Which is tiny in realwold projects with actual users.

[Ryan Moultrup @MoultrupRyan]

@ben_sage This sums it up beautifully. I might add that the 100x gain from the software engineer won't be noticed, or appreciated, until further along in the project lifecycle. Do you think the non-engineer vibe-coders have any idea what's in-store scaling and maintaining a product?

[Ryan Moultrup @MoultrupRyan]

@OchukoDotSpace @notdvb Makes total sense. It's simultaneously an eye catching work art that grabs your attention and I don't even have the words for what you'd call trying to sell something like that as a professional service.