AI Essentials Sept 2024 – Rewriting the rules for Software development

The roles of the traditional software developer will change. What is astonoshing is the speed of this shift.

Written by

Mattias Skarin

Published on

2024-09-23
BlogAI

Instead of writing code, the software developer craft will be shifted to :

  • Build data platforms for AI training (real and synthetic data)
  • Fine tune existing models, handle the AI solution lifecycle
  • Design tests to validate the applications are safe (not harmful)
  • Write code to handle edge cases (driven by Safety or lack of data)

While we saw the light of ChatGPT in 2020, it was easy to imagine that this would be a helpful reserch and coding tool to a developer. What was harder to imagine is that AI would replace software developers. Well, now in 2024 that is happening.

Yes, it is still in early experimental stages. But if took four years to get here, give the amount of $$$ poured in this field, where will be say four, eight years from now? My prediction is that the roles of the software developers will shift (see above).

In case you don’t belive me, take a look at these three projects.

LlamaCoder, open source web app that can generate an entire app from a prompt

TogetherCompute built LlamaCoder, open source web app that can generate an entire app from a prompt

https://ai.meta.com/blog/together-ai-llamacoder

Cognition on the path of building AI software engineers

Cognition has built Devin, a coding agent that codes, deploys the Linus infrastructure and resolves bugs. Worth paying attention is attentin and progress of how to evaulate the results of their AI engineers.

“The production version of Devin trained using this methodology performs 74.2% on our internal cognition-golden benchmark without ever having seen the evaluation tasks during training”

Check it out:
https://www.cognition.ai/blog/evaluating-coding-agents

But the most astonishing leap is this.

AI breaks new ground by simulating Doom without a game engine

Take a look at the video below. It’s real time capture from Doom produced by GameNGen. It runs a playable version of Doom without a game engine. Yes, it only runs for a short time due to memory shortage, and it was only trained on Doom. But imagine if it was trained on multiple games and game types?

Read about the Gamegen architecture on github:

https://gamengen.github.io/