← Back to context

Comment by Capricorn2481

3 days ago

> junior developers with no experience are going to need more handholding

Unlike AI, which gives me fake methods, broken code, and wrong advice with full confidence.

I just “wrote” 2000 lines of code for a project between Node for the AWS CDK and Python using the AWS SDK (Boto3). Between both, ChatGPT needed to “know” the correct API for 12 services, SQL and HTML (for a static report). The only thing it got wrong with a one shot approach was a specific Bedrock message payload for a specific LLM model. That was even just a matter of saying “verify the payload on the web using the official docs”.

Yes it was just as well structured as I - someone who has been coding as a hobby or professionally for four decades - would have done.

  • That's great for you. I ask Sonnet 4 to make a migration and a form in Laravel Filament, and it regularly shits itself. I'm curious what those 12 services were, they must've had unchanging, well documented APIs.

    • That’s the advantage of working with AWS services, everything is well documented with plenty of official and unofficial code showing how to do most things.

      Even for a service I know is new, I can just tell it to “look up the official documentation”

      Using ChatGPT 5 Fast

      AWS CDK apps (separate ones) using Node

      - EC2 (create an instance)

      - Aurora MySQL Serverless v2

      - Create a VPC with no internet access - the EC2 instance was used as a jump box using Session Manager

      - VPC Endpoints for Aurora control plane, SNS, S3, DDB, Bedrock, SQS, Session Manager

      - Lambda including using the Docker lambda builder

      - DDB

      - it also created the proper narrowly scoped IAM permissions for tfe lambdas (I told it the services the Lambdas cared about)

      The various Lambdas in Python using Boto3

      - Bedrock including the Converse and Invoke APIs for the Nova and Anthropic families

      - knowing how to process SQS Messages coming in as events

      - MySQL flavored SQL for Upserts

      - DDB reads

      In another project the services were similar with the addition of Amazon Transcribe.

  • > I just “wrote” 2000 lines of code for a project between Node...

    I think I wrote -200 lines of code on my last PR. I may be doing something bad for that number to be negative.

    • The difference is probably that I only do green field POC implementations as a solely developer/cloud architect on a project if I am doing hands on keyboard work.

      The other part of my job is leading larger projects where I purposefully don’t commit to pulling stories off the board since I’m always in meetings with customers, project managers, sales or helping other engineers.

      I might even then do a separate POC as a research project/enablement. But it won’t be modifying existing code that I didn’t design.