Writing Code with GPT 4o and Claude 3.5 & 3.7 Sonnet
Over the past week I have spent a great deal of time kicking the tires on OpenAI’s GPT 4o and Anthropic’s Claude Pro 3.5 and 3.7 Sonnet models from the perspective of writing code. The punchline is that GPT 4o and Claude 3.5 and 3.7 are great code-writing productivity enhancers. Autonomous AI agents they are not.
Like the title of the old Phil Collins song, we still have a “long, long way to go” before fully autonomous AI agents are capable of replicating platforms such as TikTok in 30 seconds as former Google CEO Eric Schmidt said was “right around the corner” in the summer of 2024. That statement was pure hyperbole and is in part responsible for the AI valuation bubble we are in, although that bubble appears to be collapsing.
However, there is significant economic value to be captured between here and there.
To set the table, while I have built countless financial models and analog databases, I have not written code since my college days. Not to fear, GPT and Claude will write the code for you, so long as you can describe the output you wish to create and are comfortable iterating as code errors present themselves. Not to fear here either, as users may feed erroneous code and developer console error messages to either GPT or Claude - each will peruse the shared input, identify the errors, and rewrite the script for you to where you simply have to copy and paste the new script into the destination file and deploy it.
I imagine within several years AI agents will autonomously write blocks of code, perform quality control, and deploy said code to terminals and third party environments. That alone will be a significant win, although it will not be enough to justify today’s AI valuations. For a present day functional comparison, the user must deploy the code and add it third party environments in order to generate the desired output.
As I recently wrote, experienced software engineers can be significantly more productive today versus a year ago given the latest advances of frontier models.