Raphaël MANSUY’s Post

At Lab49 we are spending a lot of time thinking about what’s next for generative AI.    In a recent presentation, Microsoft strategist Chris Coulthrust made this point: the key to leveraging AI will be design. What Coulthrust meant was that while everyone right now is focused on the capabilities of generative AI models, soon our attention will shift to the design of applications as the most critical aspect of their performance. Conversational AI interfaces are the first new UI paradigm in over 60 years. That’s according to Jakob Nielsen and his partner Don Norman, who wrote the book “The Design of Everyday Things.” Every UX designer knows “The Nielsen 10” usability principles. Putting AI in that context underscores the importance and high stakes for business leaders of getting it right. In the financial sector, people are just starting to come to terms with how this will change not just the information that’s available to advisors, traders, and portfolio managers, but their workflow. We saw something similar with the emergence of mobile when information became available at our fingertips and was aware of our surroundings. Mobile allowed us to leverage push notifications which started changing behavior. Like mobile, AI is changing the way we do things and will fundamentally change our lifestyles.  It may require major revamps in the curation of data and the design of platforms to take advantage of the efficiencies that this new “democratization” of knowledge offers. Our belief at Lab49 is that companies should also be thinking about answering questions clients have before they have been asked. Right now, people “prompt” chat boxes by typing in queries and getting answers that summarize large volumes of data. In the future, the AI will prompt people with ready-made drafts, next steps, and new ‘magic wand’ interactions. We are seeing these features come up first in customer servicing applications, but it goes way beyond traditional call centers or robo-chats. Imagine you are a trader at a large financial institution. One of the key parts of your job is to provide customers with “market color,” or commentary on swings in the bond or stock market. AI models can monitor market fluctuations and combine those with transaction history, chat transcripts, and comments from sales notes stored in CRMs. LLMs can detect sentiment, tone, and other themes from the conversations. Now combine all of that with data from “indications of interest”. Those IOIs are the preliminary orders customers set up to buy or sell securities. That information would put any trader or salesperson in a position to generate market color, at scale, that is much more targeted and useful to their individual customer. You could imagine layering in data from news applications and even Twitter or Reddit. All of this requires a tremendous amount of coordination and design to be seamless. It also offers a huge competitive advantage. 

Automate Agentic System Design with AI: A new Approach for AI Workflow But How? Let us understand... 📌 Yesterday (On Aug 19th), A new paper released titled "Automated Design of Agentic Systems". (The paper link is attached in the Comments) - This paper introduces a new research field, Auto Agentic System Designs. - A new research area aimed at creating powerful agentic systems using a meta agent. - What is Meta Agent you may ask?  ↳ Simply, a meta agent is an advanced AI system that iteratively programs new agents. - The papers claim that this meta agent can save hours of manual tasks. - But how is this Meta Agent Built?  ↳ It is built using its own search and process algorithm along with an active interaction with a foundation model. - The search algorithm first discovers and evaluates the performance of the agents then it adds them to an archive of discovered agents and uses this archive to inform subsequent iterations. - The algorithm leverages Models for coding and refinement, ensuring the generated agents are error-free. 📌 But how powerful are these agents? - The new agents were put against popular frameworks like Chain-of-Thought (COT) and Self-Refine. - These techniques serve as the baselines for evaluating the performance of newly discovered agents. Results:  • The new agents improved F1 scores by 13.6/100 on the DROP dataset.  • It also adds that the Accuracy in math tasks has been increased by 14.4% on the MGSM Dataset.  • Finally, The new agents showed 25.9% and 13.2% accuracy improvements on GSM8K and GSM-Hard math tasks. 📌 Now to answer the question: How can you save hours with this new approach? - Considering the output displayed by the automatically created agent, this gives hope. - ADAS can save human effort in developing powerful agents. These can potentially provide more effective solutions than manual design. - Imagine developing niche Agents automatically according to your use case without any manual input. This can be the future of AI Automation. Another point to emphasize, - The discovered agents maintain superior performance even when transferred across different domains and models. Which makes it very sustainable. 📌 Personal Opinion: - I think it is an interesting new field that requires a good amount of research. - A few years (or even months) from now, we might be able to create better agents for our GenAI Applications. - These agents can significantly improve the way we build our GenAI Applications. But - Right now, we do not have enough developed use cases. However, I think this will change in future. I hope you liked it, Make sure to follow for more content like this! #llms #rag #agents #aiagents #research #aipaper

Ten Wild Examples of Llama 3.1 Use Cases Meta’s recent release of Llama 3.1 has stirred excitement in the AI community, offering an array of remarkable applications. This groundbreaking model, particularly the 405B variant, stands out for its superior performance and open-source accessibility, outpacing even top-tier closed models. Here are ten wild examples showcasing the versatile use cases of Llama 3.1, from enhancing personal gadgets to innovative AI deployments. Efficient Task Automation: Llama 3.1 405B can be harnessed to teach the smaller 8B model how to execute tasks perfectly, reducing costs and latency. This setup allows users to train the 8B model to handle various operations, providing a cheaper alternative without compromising performance. Introducing `llama-405b-to-8b` Get the quality of Llama 3.1 405B, at a fraction of the cost and latency. Give one example of your task, and 405B will teach 8B (~30x cheaper!!) how to do the task perfectly. And it's open-source: https://t.co/H5590RiFhc pic.twitter.com/UyVJtFZH6V — Matt Shumer (@mattshumer_) July 26, 2024 Personal Phone Assistant: By turning Llama 3.1 into a phone assistant, users can enjoy quick and accurate responses to queries. This integration utilizes Groq’s API, demonstrating the model’s ability to provide instant intelligence, making daily tasks more manageable and interactive. I turned Llama 3.1 into my new phone assistant. It can answer anything, and look at how fast it does it using Groq’s API pic.twitter.com/dmlQ2gzSfu — Alvaro Cintas (@dr_cintas) July 25, 2024 Local Deployment of Chatbots: Building and deploying a chatbot that learns from user interactions is now possible in under ten minutes using Llama 3.1. This setup facilitates the creation of a personalized conversational agent that becomes more knowledgeable and efficient with each interaction. meta just released llama 3.1 you can now build & deploy a quick chatbot that learns more and more about you as you talk to it. in literally less than 10 minutes. here's how. there's no better time to build and ship stuff. these open models are incredible! pic.twitter.com/9fX0MMABNt — Dhravya Shah (@DhravyaShah) July 23, 2024 Distributed AI Clusters: Through the @exolabs_ home AI cluster, Llama 3.1 405B can be distributed across multiple devices, such as two MacBooks. This configuration enables users to run complex AI models efficiently at home, showcasing the model’s scalability and flexibility. Llama 3.1 70b beamed to my iPhone from my @exolabs_ home AI cluster of 2 MacBooks and 1 Mac Studio My own private GPT-4 assistant at home / on the go pic.twitter.com/0svmX39y4E — Alex Cheema – e/acc (@ac_crypto) July 26, 2024 Streamlit App Integration: With minimal code, users can create a Streamlit app to chat with Llama 3.1 8B locally via @ollama. This setup emphasizes the ease of integrating advanced AI into user-friendly applications, making sophisticated AI accessible to non-experts...

Choose the Best Tools for Developing Outstanding Applications Using AI

Nice short summary of how AI is creating interfaces.

The integration of Artificial Intelligence (AI) and Machine Learning (ML) into mobile applications is revolutionizing the way businesses operate and interact with their customers. Whether you are looking to develop chatbots, implement predictive analytics, or create personalized user experiences, OZVID Technologies has the expertise to guide you through every step of the process. #ozvid #chatbot #ArtificialIntelligence #developchatbot #MachineLearning

In an era where digital transformation is paramount, our recent survey findings showcase the expanding role of Gen AI chatbots. The results offer an overview of current user experiences and future expectations, showing a notable increase in daily usage and diversification of applications. This evolving reliance on AI underscores the need for continuous advancement and user-driven optimization. Learn more on our blog: https://bit.ly/4a8ObPS #GenerativeAI #TechTrends

Did Microsoft uncover the killer UI for generative AI last week? Satya Nadella showed a slide with a 'universal interface' in his keynote speech last week. (page 6) What does this mean? The name speaks for itself. It's an interface that understands all types of content, in real time, in order to communicate back to you, in real time. Funnily enough... they don't have a working demo of a system that adopts it. Google on the other hand, released Astra which epitomises this idea of a 'universal interface' In a world where people are plugging LLMs into everything, will the universal interface be the killer app for AI? For those of you in the coal furnace you've probably noticed that people are realising that: • You don't need to adopt a new tool to get value from AI • You can plug it directly into an existing tool or workflow This is where people seeing the most value from generative AI because: • It can be done quickly; • at a low cost; • and has minimal impact to users because they keep using the same interface that they did previously. This is where intelligence as a service (iQaaS) comes into play. It plays into the idea that the interface is irrelevant. The universal interface proves that. This approach is very different to the SaaS revolution in that it focuses on: - Augmentation over transformation - Enhancement over replacement - Impact-driven practice over standard practice - Agents over apps What is the highest impact UI you're seeing? Do we think that any of these will create the killer apps or are their more cool new interfaces to come?

Yesterday I saw yet another provocative poll about using AI to create user stories. People in the comments section were quick to dismiss the idea (I guess more out of fear of being replaced than rationally thinking about it): "User stories are about USERS!". The arguments are mostly that AI cannot replace a user and I agree, but we can still use AI to create user stories from real ... user stories. I have a space in my app where I often ask people what they want me to do next. It's a small sub-set of my users who comment in there (I would say not your average commuter type). So I took the comments and gave them to ChatGPT (3.5 because I am just prototyping here) and asked it to create me some user stories with acceptance criteria. Here's one example it created together with the user's message from the original chat: User Story 2: As a user with enthusiast mode, I want a dedicated section on the home page to view the entire fleet of different bus companies individually. Acceptance Criteria: 1. The enthusiast mode section should show a list of bus companies with their complete fleet details. 2. Users should be able to view information about each bus company by sliding through the list. 3. The section should be visually appealing and easy to navigate. Example from the conversation: "I think it would be a little cool if possible to have an extra section on the home page for people who have enthusiast mode on. Where use can see the whole fleet for every company individually." - VampireThunder I actually think it did a pretty great job of this - I understand exactly what it is that my user wants. If I was a PO/PM tasked with creating some user stories from a conversation, I would probably come up with exactly the same (without acceptance criteria number 3). For me, this is helpful in unleashing my creativity. I know have AI involved in both gathering the information and feedback from customers and in creating the product (GitHub Co-Pilot has boosted my productivity significantly - I am doing much more with way less mental effort). There are loads of possibilities with this - I can connect it to my user reviews and let it find common issues or find out what users really appreciate about the app. I could let it crawl my emails and find more missing user requests I might have forgotten about. I could even let it talk to my user's directly, asking more probing questions and getting more details that I might have forgotten to ask about. One thing hasn't changed though - my user stories still originate from my users and I am suggesting that AI can help with MORE user engagement, not to remove it.

Despite encountering issues such as bias and inaccuracies, Gen AI users are optimistic about the technology, with a majority acknowledging improvements in response quality. This year's survey highlights the complex relationship users have with AI, balancing skepticism with a hopeful outlook towards future enhancements. Read the survey for key considerations in shaping the next generation of AI tools: https://bit.ly/4a8ObPS #GenerativeAI