Paul Bratcher’s Post

I'm doing research about whether we need to structure content to make generative AI more trustworthy and predictable. So far, I'm following the information scent that says 'Yes,' which is my own confirmation bias. There are at least 2 things I want to research to support this hypothesis: 1. Actual examples of how this has been tested and shown to be true. (I'd love more than just "we tested and it's true" types of articles. DM me if you have this info but it's not publicly available.) 2. I know some people think we can use AI/ML to create the structure already. Where can I find their work and research points?

A new study s looking at how to battle #hallucinations in #ai—specifically the chance that the model may not be consistent in its answers. And they use #ai to detect the inconsistencies. Sebastian Farquhar Jannik Kossen Lorenz Kuhn and Yarin Gal developed a way to looks at “semantic” variances rather than just sequences. It detected these inconsistencies almost 80% of the time. There’s a good summary of the study in Axios but the details of the study are in Nature My favorite quote in the study: LLMs are even better at “knowing what they don’t know”, they just don’t know they know what they don’t know. https://lnkd.in/eMJrM5Ye

What makes a good explanation? AI is often a black box, but luckily explainable AI methods exist to give you insight into the reasoning of an AI model. 🔍 You might have heard of LIME or SHAP as well-known explainability techniques that show which features were most relevant for a prediction. But how do we know how good, relevant and meaningful these explanations are?🤔 A few months ago I published the Co-12 properties: twelve desirable properties for explanation quality. Explanation quality cannot be captured in a single metric but involves evaluation of content-, presentation- and user-related aspects. Co-author Christin Seifert created this great visualization of our Co-12 properties, which makes it worth sharing again! Amazing right? 🤩 Wanna know more? Check out our open-access paper at https://lnkd.in/eGmvu_Mx

This is an interesting presentation of some of the assessments needed when you want to put a machine learning/AI models in production/ real-world environment.

Can you recognise text written by AI? I don't, but fortunately, there are tools for that. If you are not sure if the text is written by AI or a person you put it into an AI detection tool and voilá! You know. Or you know with 85% probability. That is the number the best AI detection tools are capable of. Better to say they are capable of today because strength on this battlefield constantly changing. If you look for a detailed comparison and description of AI detection tools check out this page. https://lnkd.in/e79_u_jc Maybe you wonder how AI can detect AI. Don't ask me. 😊But it seems that models tend to give somehow mediocre results. That means you can base your guess on two main characteristics, first, how surprising the text is (remember, AI is trained to guess the next most probable sentence ending) and second, how eloquent it is (AI models tend to average length). So, what do you think, was this text written by AI or a real person 😉 And am I a real person? Are you? 😊

Just came across this excellent article explaining AI and its capabilities. If you are wondering what AI is, how it work works, where it goes wrong, or if "Artificial General Intelligence" is possible, give this a quick read!

Crafting #humility in AI? This Axios article about how to -detect confabulations (aka computer-generated BS) via an algorithm -become better at setting expectations -convey #uncertainty in AI answers -teach users what a mistake looks like is very interesting as it shows pathways to address one of the persistent problems in generative AI that is confident but inaccurate answers. For humans providing confident but inaccurate answers other strategies may be needed. However creating environments (workplaces, families, circles of friends, organisations...) where uncertainty, humility and #curiosity are valued would probably go a long way. https://lnkd.in/evEsJmBg

We deceive ourselves into thinking that we can do multiple things at once, when in fact all the signs suggest that multitasking is a myth, and we’re just splitting the resources we have between lots of different activities.  I think the dominant feature of the AI age is that life in itself—if not the world in itself—has turned into a big distraction, but we’re only focused on what algorithms and artificial intelligence want us to focus on.  Has AI made people dumb? — Dr Tomas Chamorro-Premuzic #mckinseyinsights #ai What happens when we automate our most impactful and superior cognitive capacity—thinking—and we don’t think for ourselves?

Do you embrace AI but unsure where to begin? This is where we can help! 😎 It’s no secret that AI has quickly become the topic of conversation these days with some mixed reviews. This means it can be hard to look at the issue objectively. In the below article, our Principal Machine Learning Consultant Greg Duckworth and our Senior Machine Learning Consultant Damien Jade Duff answer the following questions, and more: · How are clients using AI? · Have clients' expectations changed as AI has become more mainstream? · How do you decide on the best solution for the client? · Once you’ve agreed on a solution, how do you prepare to implement it? Find out more about Daemon's tried and tested approach and how we can help you harness this new technology in the most effective way possible. Plus, find out what sets us from the competition 😉 Read the full article here: https://lnkd.in/eJmAjw5n

I’ve talked to hundreds of people about how to adopt generative AI...Most do it wrong. Here are 3 questions I ask all of them: 1️⃣ Have you seen a similar problem be solved with AI? If you haven’t seen one, be ready to invest in longer scoping and research time. An added cost many don’t think about. 2️⃣ Has a more simple solution been tried? Everyone wants to jump on the AI hype train but A LOT of times a simple rules engine, although not sexy, solves the problem. 3️⃣What does a perfect answer from an AI model look like? If you’re expecting the model to provide some exotic insight you can’t identify or define, pump the brakes. If you've got the next big AI idea or initiative, ask yourself these questions first. OR! leave a comment bellow and we can chat on it more!! #ai #aigatekeeper #generativeai