Schlafender Hase’s Post

You no longer need to be a machine learning shop to join the #AI party. #LLMs are ready for #biotech today. Find out how generative AI and LLMs can be used for more than drug discovery, throughout the entire R&D lifecycle. Read the guide from our AI experts Helen Liu-Mayo, Janet Matsen, and Alan Pierce: https://lnkd.in/eq_5GvMF

Medprompt: a new frontier in AI prompting! Initiated by a Microsoft team of researchers, Medprompt was designed to improve GPT-4's precision with medical-related questions. It combines dynamic few-shot learning, where relevant examples guide the AI's learning process, with an internal dialogue strategy that pushes the AI to elaborate its thought process. This, coupled with an ensemble method that integrates varied AI responses, transforms GPT-4 into an analytical and reflective entity. The scope of Medprompt has since expanded far beyond its medical roots. Applied to the MMLU benchmark—a diverse measure of AI's understanding across multiple domains—Medprompt+ propelled GPT-4 to a striking 90.10% accuracy. This is a testament to the adaptability and resourcefulness of the methodology. Diving into the specifics, GPT-4 achieved a solid 68.4% in math and an impressive 87.8% in complex problem-solving tasks like HumanEval. These figures represent more than mere performance; they symbolize the depth and flexibility of Medprompt's application across different knowledge areas. Go check out the prompt base repo on GitHub (https://lnkd.in/di4je7ie) for more details about prompting techniques and tools.

Have you considered how Machine Learning is reshaping industries? From enhancing healthcare precision to revolutionizing automotive safety, the impact of AI and ML is profound. Our latest blog post delves into this transformative journey, exploring how these technologies are setting new standards in ethical responsibility and business strategy. Unearth the detailed insights and implications in our blog. https://lnkd.in/e-Teva4Y

Unlocking the Power of Retrieval-Augmented Generation (RAG) in AI/ML As AI and ML shape the future gradually, among the recent innovations that create a landmark revolution within the AI paradigm is Retrieval-Augmented Generation, RAG. RAG is essentially an integration of two emergent AI models: 1. Retrieval models: These do exploratory searches of gigantic datasets to fetch the information needed; 2. Generative models: They then synthesize responses using that data. So, the result is higher-accuracy, more context-aware answers – from smart customer service to the most complex applications in research. At smartData Enterprises Inc., we're already taking those AI/ML advances and applying them in real-world client solutions. In healthcare, in e-commerce – endless possibilities.

AI is a myth. AI is a money grab. Machine learning on the other hand is powerful and is the tool we should be talking about. machine learning is the engine that powers AI. We must change the narrative if we seek an ROI. ML should be the the new buzzword acronym. Here's a practical example my Company just trained a ML model that can predict cash collections, the dollar amount and the timing of the cash collection, within 99% accuracy in the healthcare industry. This of course has applications in every industry. We will be releasing demos soon. stay tuned. And remember, Opilio is the only healthcare business that is a champion of the providers.

Check out our new article.

AI is involved in all of our work. We're still navigating how we use it, how we govern it and how we deploy it across organisations. I love playing with it, but can I integrate it into my daily work yet? Partly but I always feel I'm cheating. Notebook LM by Google is great - especially the Deep Dive tool. While it's a bit cheesy, it shows the potential of future applications. There's a lot to learn and UK Research and Innovation have now provided some guidelines for the use generative AI in application, preparation and assessment. Go check it out. https://lnkd.in/e9SBKwf4

Unfortunately, I've had to use MyChart through my local provider a lot lately. While I don't appreciate the circumstances, I really love having access to these medical records. I am mostly impressed with the value of the medical notes. I like seeing the lab readings and test results but often what helps me the most are the notes.   All of that is valuable data. More specifically, it’s all unstructured data.     The volume and velocity of unstructured data in healthcare is almost impossible to comprehend, let alone act on in a systematic sustainable way.     But that changes with generative AI.     Generative AI is language-based. These large language models (LLMs) are literally built for unstructured data. And despite its conservative reputation for picking up new innovations, we’re seeing the healthcare industry embrace these tools and develop meaningful AI strategies that have the potential for business impact and patient outcomes.     I can’t think of a more exciting time to be building healthcare technologies with these GenAI based tools now — complementing the work we've been doing for years with ML based tools with structured data.

Just written part 2 of "How to build AI that actually works" with Yousef. This covers performance evaluation - how do you know if your AI is working? TL;DR: Start by building a dataset, not a model. Engineers often start by building a model. They try some queries, make tweaks and, when it looks about right, ship it! Then they get into trouble when their AI doesn't do what they expect in a live setting. They end up playing whack-a-mole and feel like they're going in circles, without making their AI more reliable overall. You can avoid this fate if you instead start by building a dataset. You first put in some time to define what "right" means across a range of cases. While this can be tricky, particularly with language models, the payoff is worth it. Once your dataset is in place, you'll know: i) how reliable your AI is overall ii) the impact of any changes you're making iii) specific areas where its performance is weak You'll then be able to iterate quickly, and build confidence that your AI is actually working. Full post:

Generative AI is driving the AI revolution, which is built on foundation models. Foundation models are offered through commercial sources or available as open source. But these pre-trained models won’t help you realize your AI strategy alone. Read why you need a platform that can train, prompt-tune, fine-tune, and serve these models for your unique use case and with your data.