The AI freight train is running in full effect. Demand is through the roof, new models keep besting the benchmark and the arms race to be the leader requires more data, compute, and human capital to keep up. The question that is arising from each new model is becoming, is AI truly that useful?
AI has jumped leaps and bounds from the release of ChatGPT in 2022. Now beyond just an answer to questions, or writing basic texts, Large Language Models (LLM) like ChatGPT, Claude, Gemini, Grok, Perplexity, and Meta AI are at a level capable of replacing human brain intelligence. For a well defined prompt, they can scour the training data set, run searches on the internet, and craft a narrative that can be short or a full research report. What would take hours if not weeks of effort from humans, can now be offloaded to GPUs and return in minutes. Content and knowledge is no longer the limiter. In fact LLMs are now a similar to TikTok and cable programming; so much content can be generated tailored to what you want, there is an overflow of information to process.
Beyond just information, AI now drives coding on a daily basis replacing hours of manual programming and turning a complex language based implementation into mere minutes of code generation that requires human validation for correct interpretation of the prompt. Claude led the path early with its coding AI, and its Claude Code app, and now Open AI has Codex , while Gemini has Antigravity. All are the programming interfaces that use English as the programming language and saved into the MD files. The MD files are converted into application code by the AI and the whole repository continues to remain with github. Agile sprints were held up due to the number of developers on the project and now projects can run multiple agents to generate code in parallel with the limiter now being token costs.
The shift now moves from developing the code to actually validating that
- The prompt was written well enough for AI to understand it correctly.
- The output matches the functionality the user was describing.
- Prompt updates for bug fixes while making sure the correct parts do not get touched.
- Ensuring the end to end functionality of the app.
With coding, a developer would unit test and run a basic smock test of the feature integrated into the platform. Now with so much changing developers would now have to validate end to end functionality. People can review code as it is generated but like reading a book, reading does not mean understanding. When additional functionality comes in, or there is indeed a bug, the developer is far removed from the code and has to fix it via the prompt and run through the whole process of ensuring the fix in one area does not disturb the running application as a whole. This is far more difficult with AI based coding.
Looking ahead, we will see what AI coding agents do when they are updated and trained on AI generated code. So much talk has been made of AI slop, and hallucinations, that there is a concern on the quality of the data in future runs. When application’s do not run well, or easily identified errors exist, the quality of AI driven coding comes up. If AI is used to detect AI coding and filter it from training, will there be any code left if all code has been generated by AI? Would this result in the drive back toward human powered AI code?
The AI driven future is truly the wild west and we are in for interesting times!