Zach Harrison

Door to door salesman → self taught developer → engineer. Building software after work. Usually thinking about AI, products, or dogs

[Zach Harrison @Zach__Harrison]

I’ve had a weird career path.

I started in door to door sales straight out of high school, moved into car sales, then taught myself to code during the pandemic.

Now I work in tech at https://t.co/nJT6ldEXs1.

My next experiment: can I build a real software business after work?

[Zach Harrison @Zach__Harrison]

@icreatelife

[Zach Harrison @Zach__Harrison]

@DWKemp @kapilansh_twt This is spot on

[Zach Harrison @Zach__Harrison]

@theo UX for sure. I feel like all the interfaces written by AI have that cheap SaaS feel to them.

[Zach Harrison @Zach__Harrison]

@firesidealpha I don’t think software IP is rendered worthless because someone built an open source copy cat of the adobe suite.

[Zach Harrison @Zach__Harrison]

@rynorhn This market is moving so fast.

In a few months from now, we’re all going to be talking about cancelling our Claude subscriptions because of some new model that comes along.

[Zach Harrison @Zach__Harrison]

Vibe coders and traditional engineers have different strengths and weaknesses, but I think both sides are missing the same thing.

Judgment and taste. Knowing what’s worth building.

Vibe coders move fast and aren’t afraid to experiment. But too many are racing to the bottom, churning out the same AI slop without asking whether anyone needs it.

Traditional engineers understand systems and know how to make them reliable. But most have zero business sense. They wait to be told what to build, then obsess over how to build it.

You can ship something nobody wants in a weekend or spend six months engineering it beautifully. Either way, no one is paying for it.

As AI handles more of the code, I see the technical work moving up the stack. Understanding how the system fits together, making tradeoffs, and catching mistakes becomes a bigger part of the job.

Both sides have something to learn from each other. I think that starts with being brutally honest about our own knowledge gaps and learning how to use today’s tools to fill them.

[Zach Harrison @Zach__Harrison]

@signulll Maybe I’m just too much of an ADHD tech bro. But personally I get a dopamine hit knowing I never have to do any of those things myself.

[Zach Harrison @Zach__Harrison]

@CoFoundersNik Makes you realize how brutal OpenAI’s business model at the moment.

[Zach Harrison @Zach__Harrison]

@nickvasiles This idea is a really interesting opportunity. What sort of workflows are you guys using autonomous AI agents for ?

[Zach Harrison @Zach__Harrison]

@dannypostma I agree. I also think engineers will need to learn to go up the stack and learn the skills of a becoming a Product Manager

[Zach Harrison @Zach__Harrison]

I never went to university, but I think this misses a huge part of its value.

The networking.

You’re paying for access to classmates, professors, and alumni. Some of those relationships can shape your entire career.

Even if an LLM can give you a better education at home, knowing the right people still gets you places that knowing the material won’t.

[Zach Harrison @Zach__Harrison]

@marclou @mariannehere I’ve walked past restaurants using AI food photos that barely look edible. If the photo puts me off eating there, you’ve paid for an ad that makes your business look cheap.

People get excited about what a tool can generate and forget to ask whether it actually looks good.

[Zach Harrison @Zach__Harrison]

I accidentally burned $200 making AI slop with Opus 5.5.

My X feed was flooded with cool looking AI generated videos and web demos with crazy animations, so I decided to get in on the action and try making something myself.

The goal was to create an impressive, visually appealing animation of a dog walk from both the owner’s and the dog’s perspectives.

Keep in mind I have zero experience developing animated shorts. So I was doing this all on vibes.

The plan was simple. Tell it what I want, give it maximum autonomy, and get the heck out of the way.

I was excited to try this, so I logged into my Claude Console and loaded up my account with $50 worth of credits. That should be plenty, right?

I decided to quickly run a small test to create an 8 second video of a cartoon character throwing a football. The video was unimpressive, but it worked. I checked my credits and I was at $48.59. We’re good to go.

The prompt was basically: make a high quality animated short entirely in code, build the characters, environments, soundtrack and effects, then render the final video.

Then I added this:

“Use maximum effort on this task.”

“Do not optimize for token conservation.”

“Continue improving the project until you genuinely believe the film is polished enough to publish publicly.”

This is where I fucked up.

I had basically told Opus to keep burning tokens in perpetuity until it decided the video was good enough.

Shortly after I kicked off the prompt, I realized this was going to be a lot more expensive than I expected.

The original $50 disappeared fast. Then I added another $20. Then another $20. Then another.

At that point the sunk cost fallacy kicked in. Hard.

Naturally, I added another $20.

Eventually I threw in another $60, followed by $35 more.

Do I cut my losses and leave the blackjack table, or do I double down?

Screw it. We’re already this far in the hole. Might as well let Claude cook.

By the end, I had purchased $225 worth of API credits and burned almost $200 of them making this thing.

What I wanted was a visually stimulating video that was highly entertaining and a little educational. I wanted to see the world from the dog’s perspective and experience the colors and smells from its point of view.

That was a far cry from what I got. First of all, the sound makes no sense and doesn’t match the visual. The video is choppy and awkward. The movements are way too jerky, and it totally dropped the ball on the dog’s POV, which was the main thing I wanted to show.

Lesson learned. If you don’t know what the hell you’re doing and tell a coding agent to burn tokens, it will. And you’ll pay the price. Literally.

I was served a $200 slice of humble pie, and it tastes bad.

I’m going to audit the code and see if there’s anything worth reusing, but either way I learned my lesson. I’ll be a lot more careful before flipping the effort gauge to max and telling a model to go nuts.

Hopefully you guys can learn from my mistake and at least get some entertainment out of my $200 AI slop.

[Zach Harrison @Zach__Harrison]

Meta AI in Instagram’s search bar is the most useless AI feature I’ve ever seen.

[Zach Harrison @Zach__Harrison]

@RyanSael @OpenRouter @alexatallah @bcherny This is so fucking cool. Congrats man. Watching someone go from 200 followers to millions of views just by building stuff they think is fun is incredibly inspiring. Well deserved.

[Zach Harrison @Zach__Harrison]

This weekend’s hobby project was playing around with OpenAI’s new GPT Live API.

I built an AI host for a fake Italian restaurant. What do you think of talking to a restaurant website like this?

[Zach Harrison @Zach__Harrison]

Am I the only one who still doesn’t understand how AI usage limits are calculated ?

[Zach Harrison @Zach__Harrison]

@boristane What’s more valuable right now: a brilliant engineer with zero product sense, or a crappy engineer who understands design, product, and people?

[Zach Harrison @Zach__Harrison]

@SoveyX What are you guys using to make such good looking videos???

[Zach Harrison @Zach__Harrison]

@gabriel1 True. I think it’s fairly easy to see who’s just trying to game the algorithm, and who actually has something to say.