I messed up... maybe letting AI do all our coding isn't a great idea

AI worked just fine. The human component was the weak link. This is a great lesson on multiple levels.

Indeed

That's why I have agent teams, one agent QC the work of other agents and so on. I built My Virtual Office to manage agent teams like real employees, allowing me to see everything they do live:

Should've been charged more, frankly. We need more taxes on people who deflect and then peddle to others the same addiction.

Can't outsource our thinking. Every time I hear someone say they're not checking the results anymore ... I just think that is asking for trouble. It's ok to speed ourselves up, TRUST, but VERIFY

I dont let AI code logins/payment stuff

Slows you down then

my dad helps me with it xD... also whens the dgx station video comming 👀

The best thing you can do is give it a rest. Go get some sleep. You look tired. AI isn't worth having a stroke or heart attack. Rest up and be well.

Don't be so hard on yourself. I've seen plenty of seasoned engineers blow thousands of dollars on seemingly simple mistakes. No AI involved.

Thanks for sharing this. Good learning moment for everyone entering the space Hopefully @Google and @awscloud make more accessible billing system that even a lamen could navigate It's going to be needed

Oof, hard lesson. AI is incredible until you forget the human oversight layer. Thanks for the reminder!

This is an expensive lesson to learn. Many coders I've worked with in the past will automatically deploy servers on Azure or AWS - which would default to the a "standard" build much larger than what might be needed for a dev environment. This lesson applies here too. There is a decoupling between the coding side and the infrastructure management side of the architecture that is not exposed to vibe coders. Worth the extra time to understand how the hosting agency will manage this.

I just prepay with usage alerts

This is the sort of 'oops' that leads me to believe you have no experience designing large scale systems. Anyone designing their app infrastructure should understand all components and cost. Not trying to knock you. It's a learning experience.

Dude, just use Cloudflare, with Cloudflare code mode. Use workers and the $5 plan.

Autopilot is still only good for short sessions. For anything longer, you need to carefully monitor the LLM’s thoughts or you’ll get a bunch of bad mistakes made that you didn’t want.

The simple truth is that LLMs have no idea what is important to you, what your context is. Why are you doing this? What's your budget? Is this a hobby project or a prototype for a launching startup? How much do you actually know about developing software? What features are more important than others? etc. This context needs to be built up through your collaboration with the AI — and that needs a memory system that captures what happened and what reality looks like in your work (world model). Shameless plug, for exactly these reasons I developed collabmem: an easy-to-use, easy-to-install, yet powerful long-term memory system, optimised for Claude Code:

I like the quote, from this year can't remember from who said "AI didn't replace SWE, just average thinking". It's part of the reason I wrote this for people who are going into vibe coding fully raw to give them SOME kind of development concepts

@MatthewBerman this is exactly the reason I built @afkode_ai. Building with AI should be about spending more time in the thinking phase, which would have allowed you to not take those suggestions from you coding agent at face value. Take a look. I am sure you'll like it.

One other thought on this. You mention code release, but you don't talk about code management, customer support or any of the other update functions that result as a release of code. As vibe coders start to release more code, who is going to keep them updated? Who fixes bugs? How do you hire someone for QA to fix code that no one knows how it was made? Companies will be looking for other companies to support products that maybe no one knows who made. The classic plan-build-deploy-support model will have to have a major overhaul too. I know the short answer here is that agents will take care of it all... but, really? In the short term?

Summary: just dont use vercel

That's on Vercel's side though. Same issue happened to me..

Thanks for sharing your experiences Matt. My AI app wrote 5 million lines of Java code plus tests and am very happy with the results. In particular it recently adopted your prompt-injection detection strategy. Classic software engineering is the answer.

As AI starts taking on larger workflows, stronger, more robust AI frameworks become increasingly important. V-model light (systems engineering): Have (AI) define clear specs + modular design first, then verify it meets the spec. Six Sigma: monitor & reduce output variance → lower defects. AI is getting increasingly good, but trips on critical parts or produces often unacceptable high defect rates. And humans struggle to monitor processes that are 80% or 95% successful.

I got Claude to just setup my own server with Hetzner and it’s beefy enough to actually host multiple of my projects. I got it to create its own skill for deployments so I can just tell it to deploy and it deploys to my own server, no hidden costs or fees

You can also build locally and deploy the pre built

Sounds like a skill issue

Just use Replit

It’s a reminder that blindly trusting can lead to surprises. But interestingly, this does not apply only to AI. You are managing, and you are not the subject-matter expert anymore. Be a good manager.

AI IDEs should have a "guardrail view" where users can set parameters to prevent automated decisions made that violate specified parameters related to security, efficiency, platform costs, etc. Also, a log of user interventions where humans redirected the coding agent.

Matt, you can put monthly caps you know that

Maybe Ai has its own language & we won't understand it

locally host your LLM Problem solved GLM, Qwen, & MiniMax

Gemma?

The pattern I keep seeing: AI makes the first 80% trivial, so you skip the part where you actually understand what was built. Then something breaks and you're debugging a system you never internalized. The build wasn't the mistake - skipping the review was.

Doesn’t Vercel have a free plan lmao? That wasn’t good enough?