I got the title from plugging in bullet points into ChatGPT and asking it to write a blog post. The post describes some of the things I did yesterday, which written in full would read as follows.
Today I attended the ML Applied Projects Group event, but it ended up being a 1:1 with Régis Schiavi since no one else was there (the new batch officially starts tomorrow). Régis described his recent focus on computer vision, since these models are much easier to understand and debug. While using transformers on vision models marginally helps performance, the paper he came across describes how vision models with transformers helps with portability across different use cases. I didn’t know that much about RAG, and Régis said it’s a bit like converting a document into vectors, and a new user query would fuzzy match against those vectors (say with cosine distance). An open question I had was whether creating a custom GPT a backdoor to avoid paying for ChatGPT fine-tuning.
Back on the Leetcode horse, I started with [Two Sum](https://leetcode.com/problems/two-sum/). My attempt solves for the given test case but does not generalize if the first item is not part of the solution. ChatGPT points this out and gets the working solution by updating the dictionary of remainder values.I then worked on [Move Zeroes](https://leetcode.com/problems/move-zeroes/). I solved this but moved the zeroes to the left instead of to the right of the array, as required by the test case. ChatGPT helps point this out and flips the equality condition for the right answer.
This was the prompt into ChatGPT.
Please write a blog entry based on the following bullet points. Each cluster of bullet points should be its own paragraph.
---
- 1:1 - Régis Schiavi
- Régis focusing on computer vision since models are easier to understand and debug
- Régis finds that transformer not help that much with performance, but helps with model portability across different use cases- I didn’t know that much about RAG, Régis basically shared how it involves converting a document into vectors, and allowing search to be done against those vectors- open question - is using a custom GPT a way to avoid paying for ChatGPT fine-tuning?- LeetCode - Two Sum https://leetcode.com/problems/two-sum/- solve given example but not generalize if index 0 not part of solution- ChatGPT gets working solution- LeetCode - Move zeroes https://leetcode.com/problems/move-zeroes/- solve but moved zeroes to the left instead of the right of the array, as per given test case- ChatGPT helps point this out
This was the response.
# Exploring Advanced Computer Vision and Coding Challenges: Insights from Régis Schiavi and ChatGPT
In the realm of computer vision and machine learning, Régis Schiavi, a notable expert in the field, has been making significant strides. Régis's focus on computer vision stems from its relative ease of understanding and debugging compared to other models. His insights reveal that while transformers may not substantially enhance performance, they do offer an advantage in terms of model portability across various use cases. This flexibility is crucial in adapting to different applications, a trait that makes transformers increasingly valuable in the diverse landscape of machine learning.
A particularly intriguing aspect of our discussion involved Régis's explanation of the Retrieval-Augmented Generation (RAG) model. He elucidated how RAG transforms documents into vectors, facilitating a more efficient search process against these vector representations. This approach not only optimizes the search mechanism but also enhances the overall functionality and application of the model in practical scenarios. A lingering question from our conversation was whether using a custom GPT model could be a strategy to bypass the costs associated with fine-tuning ChatGPT. This open-ended query hints at the ongoing exploration and innovation in the field of artificial intelligence.
Shifting gears to coding challenges, let's delve into the world of LeetCode, a platform known for its comprehensive range of coding problems. One particular challenge, "Two Sum" (found at [LeetCode - Two Sum](https://leetcode.com/problems/two-sum/)), presents an interesting scenario. The task is to solve a given example without generalizing the solution, particularly if the index 0 is not part of the solution. In tackling this problem, ChatGPT demonstrated its prowess by arriving at a working solution, showcasing the practical application of AI in problem-solving in programming.
Another intriguing challenge from LeetCode, titled "Move Zeroes" ([LeetCode - Move Zeroes](https://leetcode.com/problems/move-zeroes/)), posed a unique problem. The objective was to move zeroes within an array, but in an unconventional manner - to the left of the array, contrary to the standard test case which usually involves moving them to the right. This twist in the problem statement was astutely pointed out by ChatGPT, highlighting its ability to not only solve problems but also to recognize and adapt to variations in problem requirements. This interaction with ChatGPT underscores the versatility and analytical prowess of AI in coding and algorithmic challenges.
In conclusion, the insights from Régis Schiavi and the practical applications demonstrated by ChatGPT in solving coding challenges from LeetCode provide a fascinating glimpse into the evolving landscape of computer vision, AI, and programming. These developments not only push the boundaries of technological capabilities but also offer a window into the future of problem-solving in the digital age.
This is more fun than I thought.