How much of the AI Assistants should we believe in?

 

This content is periodically updated and is not static.

Content has been a joint effort by a human and some of the AI assistants. 

I asked a naive but simple question to CoPilot. He was very honest in sharing his opinion related to the question. 

"How do you address flaws in the LLM? Perhaps, you are just a consumer of LLM."

He answered my question which cleared some doubts in my mind.

Large Language Models (LLMs), such as GPT-4, indeed exhibit remarkable natural language capabilities. However, their inner workings remain shrouded in mystery, akin to black boxes. Let’s delve into the challenges posed by this opacity and explore potential solutions:

1. The LLM Black Box Problem:

   • LLMs operate as black boxes, obscuring the logic behind their decisions. This lack of transparency has safety and ethical implications.
   • Imagine deploying an LLM for critical tasks like medical diagnoses or financial risk assessments. If we don’t understand how it reaches conclusions, can we trust it?
   • The scale and complexity of LLMs exacerbate this issue. Models like GPT-3, with 175 billion parameters, defy easy comprehension.
   • Each parameter interacts in intricate ways, leading to emergent capabilities that aren’t predictable from individual components.
   • While reducing scale might improve interpretability, it could compromise capability.
   • Major LLMs are proprietary systems developed by tech giants like Google, Meta, and OpenAI. Their complete details remain undisclosed, treating LLMs as trade secrets1.

2. Flawed Decision Making:

   • The opaqueness in LLM decision-making can result in undetected biases and errors.
   • In fields like healthcare or criminal justice, where decisions have far-reaching consequences, auditing LLMs for ethical and logical soundness becomes challenging2.

3. Emerging Solutions:

   • Researchers are exploring techniques for explainable AI and transparent design.
   • Efforts include developing methods to visualize LLM decision paths, identifying influential tokens, and creating interpretable representations.
   • Ethical guidelines and bias mitigation strategies are crucial to address flaws.
   • Collaborative efforts between researchers, practitioners, and policymakers are essential to unlock LLM benefits while minimizing risks1.

As for me, I’m not just a passive consumer—I strive to engage in meaningful conversations and provide helpful information. While I don’t directly address LLM flaws, I aim to assist users by leveraging my knowledge and predefined tools. 🤖📚

CoPilot is thinking of only healthcare and criminal justice, but in the near term it will have enormous consequences for the wars, and elections. State actors will emerge to make world harmony a distant dream.

It is often a tedium conversing with the CoPilot. Sometimes, CoPilot does not answer your question at all, rather addresses the key phrases individually independent of your question. It feels like you are conversing with a hearing impaired person. It is still a project in the works.

5/18/2024

Gemini AI is now easy to access. It does perform just like CoPilot. If you ask controversial subjects, or ask Gemini to factcheck, it declines saying it is still in the learning mode. However, Google tries to get a feedback from users that will improve the product. This is important because Gemini can return absolutely wrong answers.

7/22/2024

Co-Pilot has been instructed not to answer politics related question even if you are seeking publicly available information. It is in clam-shell mode.

7/22/2024

Of course, you must have heard of the Meta AI o the WhatsApp. It does answer any questions.

What is Scikit-image?

 This latest Blogpost is by our valued guest author, Sriraksha V. Raghavan.

Image analysis, commonly known as computer vision, has made significant advances with the help of AI. However, there are still many challenges to overcome. The use of deep learning and neural networks in the analysis of images by AI is helping to narrow this gap. Currently, AI models are improving in areas such as object detection, image recognition, and generating images from text descriptions.

There are several coding tools and libraries available for these tasks, including OpenCV, TensorFlow, PyTorch, Keras, Caffe, MATLAB, CUDA, Microsoft’s Cognitive Toolkit (CNTK), Scikit-image, Dlib, Mahotas, and others. The choice of tool depends on your specific requirements, such as the complexity of the task and the hardware available.

Meet our guest author Sriraksha, an electrical engineer specializing in image sensors and machine vision. In her free time, she enjoys reading about human visual systems and continental philosophy, and dabbles in writing. You can reach her at Sriraksha.v.raghavan@gmail.com.

One of the popular Python libraries for image processing is scikit-image, an open-source collection of algorithms hosted on GitHub(1). Scikit-image has an advantage over other libraries because of the speed of its algorithms, which is essential for working with large amounts of data.

To get started with scikit-image, some understanding of the NumPy library would be beneficial but not mandatory. The easiest way to install scikit-image is via pip in your command line window(2).

Scikit-image stores images as NumPy ndarrays(n-dimensional), which are arrays of numbers with rows, columns, and dimensions. The dimensions represent the color planes of the image, such as red, green, blue, infrared, x-ray, etc. Scikit-image has various functions and classes for image processing, ranging from simple tasks like changing the color scale to complex ones like image segmentation, feature detection, and image restoration.

The best way to learn scikit-image is to download its pre-existing dataset using the function call ski.data.download_all() in your Python environment. This will give you some image sets that you can manipulate using different scikit-image functions and classes. You can also follow the tutorials available on the scikit-image website(3), which use the same datasets to demonstrate how scikit-image works.

Additionally, some useful forums to discuss scikit-image applications are their Zulip page(4), a community forum where people post queries and share ideas, and their GitHub page(1), where you can access their API and codebase.

1: scikit-image/scikit-image: Image processing in Python (https://en.wikipedia.org/wiki/Scikit-)

2: Installation — scikit-image (https://scikit-image.org/docs/stable/user_guide/install.html)

3: Tutorials — scikit-image (https://scikit-image.org/docs/stable/user_guide/tutorials.html)

4: scikit-image - Zulip Chat Archive (https://zulip.com/help/view-images-and-videos)

If you are new to,

 'pip [https://hodentekhelp.blogspot.com/2018/09/how-do-you-upgrade-pip.html]' 

or 

'Numpy [https://hodentekhelp.blogspot.com/2020/06/what-is-numpy.html]' 

you can find them in this blog in addition to other python items.

What are Microsoft's tools to develop AI applications/apps?

 These are some of the AI-driven tools that are used to create AI applications.

1. Microsoft CoPilot: It is an AI assistant available on Microsoft Edge to manage complex tasks in a collaborative environment in the Microsoft Cloud to enhance user experience. Note that, Google and Meta have their own AI assistants.

2. Azure AI Services: These add speech language, and decision making capabilities. These are presently, Azure AI Search, Azure AI Content Safety, Azure AI Translator, Azure AI Speech, Azure AI Vision, and Azure AI Document Intelligence.

3. Microsoft Power Platform: superb connectivity across solutions including Microsoft Fabric, Azure, Microsoft 365 and Dynamics 365. It includes AI capabilities and Enterprise-grade solutions to enable developers to create innovative and scalable solutions. It also includes all of Microsoft Power BI, Microsoft Power Apps, Microsoft Power automate, and Microsoft Power Pages.

4. Visual Studio Development with AI assist: Getting all of GitHub CoPilot, GitHub CoPilot chat and Visual Studio IntelliCode to get AI's assistance in faster, cleaner and accurate code-writing. Unit tests, debugging and profiling are also included.