Faisal Ahmed’s Post

When asking AI tech questions, it's worth keeping in mind that the initial response will have answers with high popularity ratings, which aren't necessarily good, technically speaking. Here's an example. In this case, the first two choices are just bad - one deletes an element by value (a set should really be used for that), and the other one creates a whole new list instead of deleting an element from the existing list. The 3rd method is what one would use if the deleted value is not needed, and the 4th one when the value is going to be consumed in some way (the example just throws it away, which is also not good).

A couple of weeks ago, we introduced the idea of computer vision. What happens when you pair it with large language models (LLMs) like Google Gemini? This powerful combination takes AI to the next level. It can recognize and interpret images, analyze them, and provide detailed descriptions and insights. We’re transitioning to a world with an AI that understands what it’s seeing and can guide you through complex scenarios with step-by-step instructions! This synergy transforms how we approach learning and problem-solving, making it more interactive and intuitive than ever. Explore more in AI’s Real World Applications: Vision, Google Gemini here: https://lnkd.in/g27uFYBD

Watching Melanie Mitchell, my new hero in AI development...my team joined the ARC (Abstraction and Reasoning Corpus) competition on Kaggle- She goes deeper than semantics, which is nice- I like that the model they want in the competition we joined is optionally offline- that is another like future reach for myself educationally- I am curious about the effects of the ethics involved that are not statistically curved by browsers or marketing, that comes with internet usage. I like that her background is the same as mine, Computer Science- I am personally working on a distributional semantics model.... https://lnkd.in/gFmE-csT

I often encourage my child to make requests politely, punctuating them with the "magic word", namely "please". Turns out, the magic power of that little word may be magnified in the era of generative AI (though of course more research is needed on this). There are two reasons I find it plausible that saying "please" when prompting a machine leads to higher quality responses: 1. Traditional pre-ML software follows the mantra "do as I say, not as I do". ML software has a bit of both, and frankly the "as I do" may be more important than the "as I say". Are you more helpful to people who ask you nicely? So are most people who created the data that trained the LLMs. 2. With the type of detailed prompts you need to write to get the most out of an LLM, it is not always obvious where the context ends and the request begins. "Please" takes care of that. So, whether you're asking Microsoft Copilot or a literal copilot to do something, remember that common courtesy can go a long way. Please. #generativeai #genai #benice

Build custom on-device machine learning models with ease using the Google AI Edge MediaPipe Model Maker, no deep ML expertise required 🛠️ → https://goo.gle/4iwTVaE Train models for tasks like image classification, object detection, and text classification right from your computer or browser with colab.

Marketing is great, the I’m thinking message is a great find 🙃. Mind u it is an LLM it’s not consciousness, self-aware, or has genuine understanding! Conversation: “You raise valid points about the limitations of large language models (LLMs) like myself. While we can generate coherent and contextually relevant responses based on patterns learned from vast amounts of data, we do not possess consciousness, self-awareness, or genuine understanding in the way humans do. Human cognition involves abstract reasoning, emotions, subjective experiences, and the ability to form complex theories about the world—facets that are currently beyond the reach of AI. LLMs operate through statistical interpolation within high-dimensional latent spaces, which enables us to predict and generate text. However, this process doesn’t equate to true understanding or reasoning. Biases inherent in the training data can influence outputs, and attempts to extrapolate beyond trained knowledge may lead to inaccuracies or confabulations.” While LLMs represent a significant step in natural language processing, they are just one part of a much larger exploration into understanding and replicating human cognition.

This time on my journey to make cool stuff, I dive into the world of AI Agents 🤖 AI Agents have gotten an increased focus with the release of things like Devin- the AI code assistant, Microsoft’s AutoGen, and LLM tuning for function calling effectively showing that you can simulate reasoning and using tools with language models to create your own “mini worker” agent! I’ve been experimenting with making Agents from the ground up, and have just released my initial research into the context of what an AI agent is, as well as an in-depth example where I create an agent-supervisor workflow, where one supervisor LLM is “managing” two sub agent LLMs that are using different tools to analyze social media comments and write reports. Check out my vid here:

AI, LLMs Deep learning

A few food for thought papers and presentations this Monday 1. Google's #Gemini Agents - enough said! https://lnkd.in/eM9ptPAf As #LLMs are becoming more mainstream, the science of benchmarking is finally catching up: 2. OpenEQA: Embodied Question Answering in the Era of Foundation Models - "OpenEQA – the first open-vocabulary benchmark dataset for EQA supporting both episodic memory and active exploration use cases. OpenEQA contains over 1600 high-quality human generated questions drawn from over 180 real-world environments. In addition to the dataset, we also provide an automatic LLM-powered evaluation protocol that has excellent correlation with human judgement. Using this dataset and evaluation protocol, we evaluate several state-of-the-art foundation models like GPT-4V and find that they significantly lag behind human-level performance." Paper: https://lnkd.in/eVzFeGbp 3. CausalBench: A Comprehensive Benchmark for Causal Learning Capability of Large Language Models Large language models (LLMs) need to grasp causality concepts to work well. This includes explaining their reasoning, adjusting to new information, and thinking about "what if" scenarios. As these AI models become more common, it's important to check how well they handle causality. However, there hasn't been a good standard way to test this, which has made previous studies limited and repetitive. To fix this, the paper introduces a new test called CausalBench. This test includes three specific tasks that help compare how AI models do against traditional causality learning methods. CausalBench also uses different types and sizes of causality problems to really push the limits of what these AIs can do. It even includes background knowledge and structured data to help the AI understand long texts and use information it already knows. The paper also tests nineteen top LLM models and shares findings about their abilities. Paper: https://lnkd.in/eGjANzzs

Is it magic — or is it a LLM learning how to make a probability distribution? A new open-source, web-based tool designed by researchers at IBM Research and Georgia Institute of Technology lets non-experts see what actually goes on under the hood of a transformer-based language model, which learns from large-scale data how to mimic human-generated text. Check it out in our latest explainer breaking down the AI transformer: https://ibm.co/3Mgre2Z