AIception: The AI That Is Coding AI Now
Unleashing our data scientist within, using AI to 10x ourselves.
It’s a new world this week after the launch of GPT4, welcome to the fascinating world of AIception - AI writing AI. I’m sure the data science community is going to throw gas on this fire. Just when you thought the world of artificial intelligence couldn't get any more mind-bending, in comes this revolutionary tool that's pushing the boundaries of our collective imagination. In this blog post, we dive deep into the uncharted territory of AI-generated code and unravel how it's not here to replace AI researchers but to propel them into a new era of creativity and productivity. So don’t buckle up and get ready to witness a colossal transformation in the data science landscape.
So in GPT4 you can put this prompt:
and it will generate this code:
Then once the code has been generated I can even ask it how to run it on Ubuntu and it will tell me:
The script didn’t run the first time and threw an exception, when I told GPT4 the exception it quickly fixed it:
Then the code ran fine and resulted in an AUC of 0.970 on the training set by using a RandomForestClassifier(). Now if I want it add a stronger AutoML I can request that as well even with no knowledge of which AutoML library is best.
The system suggests TPOT and rewrites the function including that as a classifier and even takes advantage of my 12 processing cores. GPT4 TPOT output:
The original code was taking too long to run, so I asked if GPT4 could speed it up to run in a few minutes. It reduced my generation count to 10 and my population count to 20. These AutoML results produced an AUC of 0.969% and an accuracy of 91.25%, not bad for someone starting out with machine-learning learning the ropes.
Conclusion:
I think this is quickly becoming the new StackOverflow for data science developers, why stumble around the internet looking for code snippets that often don’t work or match what you need when you can just type it in? Also, GPT’s ability to handle exceptions and fix them is unprecedented.
Key points to consider:
Developers have always been figuring out how to move faster:
Over the years, developers have constantly sought new ways to improve their productivity and efficiency. The introduction of libraries like Python3, Sklearn, PyTorch, and TPOT has significantly accelerated the pace of development in data science. By providing reusable, well-optimized code, these libraries have enabled researchers to focus on higher-level tasks and foster innovation in the field.AI as RPA (Robot Process Automation) for exception handling:
AI-generated code is akin to robot process automation for exception handling. By anticipating potential issues and resolving them on the fly, AI-driven code development allows developers to focus on designing cutting-edge solutions while minimizing the time spent on tedious debugging. I love the idea of code self-healing and not wasting anymore time, or at least limiting time, trying to troubleshoot non-obvious bugs.How AI-generated code helps developers learn new APIs and best practices:
AI-generated code not only streamlines the coding process but also introduces developers to new APIs and best practices (e.g. TPOT, etc…) they may not have been aware of due to limited experience. By leveraging AI-generated code, developers can learn from the AI's knowledge of the latest libraries and frameworks, further enhancing their skills and capabilities. Especially when this starts learning in real-time and consuming the latest runner-up Kaggle solutions.Improved code quality through AI-generated code:
Just like GPT4 type technologies can be used to improve grammar they can also be used to improve our code for other humans to review. AI-generated code, being well-documented and PyLint compliant, leads to better overall code quality. Writing code at this level no longer requires the burden of extra work. Need your tests written? Yes… it writes your unit tests for you… 🤯. With precise documentation and adherence to coding standards and testing, AI-generated code can serve as a robust foundation for developers to build upon, leading to more reliable and maintainable software solutions.
Conclusion:
As the lines between AI and human developers blur, it's evident that AI-generated code is set to revolutionize the data science landscape. Vin Vashishta, a leading voice in the AI community, aptly puts it:
With AI-generated code taking the reins, developers can now focus on solving complex problems and innovating, while AI takes care of the mundane yet essential aspects of coding. It's not a replacement but an accelerant, empowering AI researchers to break through the barriers and redefine what's possible in the fascinating world of data science.
If you think this is impressive now, just wait 6 months. I’m already shocked at the rate of GPT4 improvement within GPT4 after 48hrs of testing, it’s learning as we speak getting smarter everyday.
Like/subscribe/comment/feedback always appreciated.
You read this far, here are some more images of Sam Altman hanging out with his AI system using MidJourneyV5: