Coffee and AI with Sohini, Episode 2: chatGPT for Ed-tech, LLMOps, AutoGPT, Responsible AI explained
We learned about what chatGPT is in the last episode also linked in this episode. Now chatGPT is a virtual companion that has been trained on billions of public domain web pages that has the capability of making our code and writing better and respond to complex questions in a human-like manner. If you are tuning in for the first time, I highly recommend you to listen to that first for more context.
Today’s episode is geared towards the recent rush to define clear use cases for chatGPT applications and products.
MIT tech Review: ChatGPT is going to change education, not destroy it
https://www.technologyreview.com/2023...
ChatGPT and alike models have also raised the question “Are they Responsible AI?”. The obvious answer is No, since there are times (significnatly less) where the responses go off the rails and are unacceptable. For instance in the college application example, the chatGPT was given options of two public universities, one wil 24K students and another with the 26000 students. The response it returned was that the larger public university with 24000 students was better. This makes the responses as unstable and unexplainable, which leads to it being a not-responsible AI. The hope is with appropriate prompt engineering and guardrails this problem can be solved.
This brings us to the next latest update, the AutoGPT on March 30, 2023 by the company Significant Gravitas Ltd as an open source product. The autoGPT is trained with instructions on top of the chatGPT4 to execute complex tasks rather than the constant interactions as needed by chatGPT4. So lets say you want to write a new childrens book. For AutoGPT, you give it a set of resources to do its search (like google and chrome) and say that you want to make a story with a little boy named Max and bananas. You can specify the number of sentences for the story etc. This becomes a rough outline. Then you tell AutoGPT to save this story in a file called “story.txt” and then generate 8 possible titles for the story and save that as a file “names.txt”. AutoGPT will then execute this all and result in 2 separately saved files. For the same task if you had used chatGPT, you would have had to enter a command for each task, like generate a story, save it, generate titles, then save titles etc. AutoGPT on the other hand eliminates the need to enter each instruction and operates at a task level. However, researchers have found autoGPT to have bugs. Two major concerns raised are: 1) It cannot remember the steps from one task to another, something that humans can easily adapt to, 2) It can get argumentative if you dont agree with the outcomes. So AutoGPT is a step in the right direction but far from being an actual product. Also, it is NOT an Artificial General Intelligence (AGI). An AGI would be capable of performing at a human level and more.
Segment 2:FAQs:
Updates on Tools and softwares:
https://dinov2.metademolab.com/ that can do 2D camera depth estimation, semantic segmentation and instance segmentation all without any new annotated data. This would mean that the tool is already trained by millions of well annotated data so far. Try it and report back if it does not perform the way intended.
Comment on new Data sources for enthusiasts to begin their Data Science Journey:
Kaggle is a great resource. Latest multi-modal source: ego4D https://ego4d-data.org/
How to transition from Data Science/ ML Researcher to Stack Developer.
For this I find the cloud provider specializations to be much more specific, updated and relevant to learn end to end ML system design. So AWS, Azure or Google Cloud certifications for MLOps, Full stack engineering will be crucial. You will need to learn apart from backend (Python), frontend webapp development, PHP, Java, Flask, Fastapi and Kubernetes deployment and monitoring solutions.