This is a great summary of the sea change that’s occurred in programming due to AI agents. They were basically hype until December of last year and now they’re reality.

That you can give Claude instructions in English to build an app, configure various settings and take several actions which it not only executes perfectly but then figures out how to address errors on its own is magical.

It’s easy to dismiss this if you haven’t used the latest models but writing code is effectively free.

@carnage4life That doesn't match my experience. The LLM's will still report in error of what's been written. Shipping LLM code in production without thorough review and refactoring is a huge risk. But for prototypes, boilerplate, tasks which are well-solved in the training data, they can save some time.

@weyoun6 @carnage4life

I agree. I've found several pretty basic bugs in the system even though the code created the tests as well.

@carnage4life

Agreed. Its staggering.

I am doing in 2 days what a team of 4-5 people would take 2-3 months to build

UI, CSS, SQL, API, error handling, deployment, debugging, access control, vulnerability assessment, refactoring, Documentation, automated testing, git management, project management...

It is also scary as hell to think about where this is leading

@carnage4life

Yeah, the new wealthy and poor.
Karpathy could do in in 30min because he has unlimited compute, it would take me 30m+4h penalty box in between.

@carnage4life this. 100% this. This changes everything for programmers¹ in every industry.

​​

​​

¹ 𝘸𝘩𝘰 𝘢𝘳𝘦 𝘤𝘰𝘮𝘧𝘰𝘳𝘵𝘢𝘣𝘭𝘦 𝘴𝘩𝘪𝘱𝘱𝘪𝘯𝘨 𝘱𝘰𝘢𝘤𝘩𝘦𝘥 𝘥𝘰𝘨𝘴𝘩𝘪𝘵

@carnage4life I strongly disagree with your assessment.

I definitely think there is value in statistical models. Blindly trusting their outputs is not one of them.

I doubt there are many of those actually tasked with maintaining and building mission-critical, production-grade software who would agree with you. I'm sure their managers would though.

What you are writing is demonstrably false and a disservice to the people who have to spend significant time fixing vibe-coded security risks.

@carnage4life reminded of the ST:TNG S02E03 "Elementary, Dear Data"

"Geordi realizes that when he asked the computer to create the program he had asked for an adversary who could defeat Data, not Sherlock Holmes; as a result, the computer gave the holodeck character, Professor Moriarty, the intelligence and cunning needed to challenge Data, plus the ability to access the ship's computer."

@carnage4life giving a random LLM your username and password plus all of the other private data is not recommended and basically not safe.

You can create all these things and tell it a dumby U/P then ask the code to create a deployment file to allow you to deploy all the code by yourself.