DEUS: human(ity)-centered AI

Do you really know what your AI systems are doing? Your model doesn't operate in a void. It impacts your business and society in unpredictable, and occasionally harmful, ways. We are excited to announce the launch of our platform, Seia, which changes all that. • Get actionable insights on the performance of your models. • Manage risks quickly. • Identify opportunities to maximise revenue • Prepare for incoming regulations, like the EU’s AI Act. Maximise your model's ROI and compliance with a human-first approach with Seia. Sign up for a private beta here: https://lnkd.in/eFBmBB9b #deus #ai #artificialintelligence #datascience #engineering #design #strategy #seia #adaptableai #beta

The latest issue of our newsletter is out! 🚀 This week we dive deep into Moshi🎙️, Kyutai's open-source counterpart to GPT-4o and Project Astra. Here's why it's turning heads: ✅ Runs on-device with just 7B parameters ✅ Processes voice and text in real-time ✅ Incredibly low latency (0.2-0.3 seconds!) ✅ Developed in just 6 months by a team of 8 Many analyses seem to get stuck on Moshi's "quirks" like inventing personas and getting stuck in loops. All these are missing the bigger picture - Moshi is not supposed to be an assistant, it's simply a prototype that can accelerate the way we interact with machines. 🤔 What does this mean for the future of AI? 1. Edge computing possibilities 2. Privacy-focused applications 3. New horizons for AI-human interaction Curious to learn more? Check out our newsletter for an in-depth analysis.

Feeling lost in the dramatic power struggle, back-stabbing, and absurdism surrounding AI? Some call it the AI reality show, others Game of Thrones. We prefer the term “AI Telenovela”, since it encapsulates the general vibe best. To that end, we’ve created a new series, depicting all recent developments as a Greek tragedy—the OG telenovela. In this first instalment, we’re focusing on #GPT4o. Here’s what happened to GPT-4o since its announcement 2 months ago: 1️⃣ Sam Altman tweeted “her” insinuating that the voice Sky is similar to the one Scarlett Johansson played in the movie with the same name. 2️⃣ Scarlett Johansson released a public statement saying OpenAI reached out to her and she rejected their offer twice. OpenAI swiftly took the voice down. 3️⃣ Altman made rounds on podcasts and conferences, claiming this was not intentional. 4️⃣ The voice actress behind ChatGPT's 'Sky' voice came forward, saying she was hired months before Altman contacted Scarlett Johansson. 5️⃣ The French research lab Kyutai released its open-source voice model that rivals GPT4o, named Moshi, and will make its weights public in the coming weeks. 6️⃣ OpenAI still hasn’t released GPT4o’s voice modality. The visual overview was created by AI (with some help from us). 🤖✨ Curious to learn more about everything else that happened in AI in the last month? Check out and subscribe to our newsletter: https://lnkd.in/d5sRvQ6U! 📩 #AI #TechNews #MachineLearning #ArtificialIntelligence #OpenAI

Worried about cloud costs? See how a simple idea can make a difference! 💸 Paulo Ferreira, our Engineering Lead, tackled the challenge of FinOps analysis for companies with a straightforward solution by creating the Cloud Cost Summary Tool. This innovative tool empowers our finance and operations teams with insights into cloud expenses, making cost management more efficient and transparent. In his latest Medium post, Paulo delves into the development process, the key features, and the significant impact of this tool on our business operations. Curious to learn more? Read the full article here 👉 https://lnkd.in/dnXX3JnX

We just published our June Iris Insights newsletter! In this issue, you'll find: 🌊 An overview on this months' three main stories: The AI Act was approved by the European Parliament Google's struggles with AI Overviews, and Apple's announcement of Apple Intelligence 🧭  Mini overviews on AI agents and Conversational AI 👩🔬 A selection of tools, tutorials, datasets, and repos to experiment with 🔮 Interesting AI developments from academia, industry and politics And last but not least, we're introducing a new section! We affectionately call it the AI telenovela, as no other term properly reflects the dramatic flair, back-stabbing, and absurdism that defines almost everything that’s been going on around AI monetisation efforts. You can subscribe to the newsletter on Linkedin, or if you prefer to have it in your inbox, check out our site: https://iris.deus.ai/

Can you actually make real impact with LLMs? LLMs are making headlines, but their impact comes once they become part of real-world applications. With OpenAI's GPT store, Microsoft's copilots, and the rise of coding agents like GPTEngineer, it's clear LLM-powered agents are becoming instrumental to the wider adoption of AI. Why use agents? Experiments carried out by Andrew Ng and his team show a dramatic performance increase when wrapping an LLM in an iterative agent loop. For example, GPT-3.5 accuracy for code-writing tasks jumped from 48.1% (zero-shot) to a remarkable 95.1% with an agent workflow (28% higher than GPT-4). These tools are by no means even close to replacing humans, but they can automate and support us in carrying out a bunch of repetitive tasks. Over the past few months, we've been exploring and developing our own AI agents, both independently and in collaboration with clients. We've prepared a quick primer, including main modules to use, frameworks for building them, and methods for measuring their value to organisations. You can download the overview in high-resolution here: https://lnkd.in/ddc-yQni Curious to learn more, reach out to us! #LLM #GPT4 #copilot #AIagents #Automation #techInnovation #futureofwork

We just released our comprehensive guide on conversational AI! The guide is packed with insights from our research and work across the Finance, Real Estate and Energy sectors, and it's designed to help you create meaningful interactions with your customers. Inside you can find: 📈 Measurable Impact: Discover how conversational AI can deliver up to 30% higher customer satisfaction, 40% cost reduction, and 50% faster response times. 🚀 Real-World Examples: See how conversational AI is used for search (think Amazon's Rufus), customer service (e.g., Klarna's AI assistant), and task automation (such as Microsoft’s Copilot). 🖌 Actionable Design Principles: Understand the three key principles for crafting impactful conversations: show & tell, give options, and ask questions. ❓ Considerations for Building Your Solution: Explore various development options and strategic considerations to get you started. You can download the guide on our site (link in comments). #conversationalAI #AI #LLM #futureofwork

87% of customers prefer brands that offer seamless conversational experiences. Companies like Walmart, Alaska Airlines, and HSBC Bank are already leveraging large language models to provide their users with options like conversational search. LLMs, however, are only one component of what makes a successful conversational experience. You need an approach that takes into consideration: 🖌 Design Principles: Employ the principles of "show and tell”, “give” and “ask" to design rich, informative dialogues (https://lnkd.in/ddgXhZg3). Brand values: Establish "branding guardrails" to ensure LLM responses align with your brand identity. 🏗 Build Grounding: Ensure responses are verifiable, accurate and up-to-date (e.g., use Retrieval-Augmented Generation). Prompt Engineering: Craft prompts to guide the LLM toward precise and contextually-relevant answers (e.g., use meta prompting). Parameter Optimisation: Tailor the LLM to your specific business needs. Infrastructure: Ensure your systems are ready to maximise LLMs’ potential. ✅ Compliance Data Privacy: Protect your organisation and customers’ data by adopting a federated approach. Regulatory Compliance: Adhere to relevant legal and regulatory standards, especially for high-risk systems (e.g., as defined by the EU’s #AIAct). 📈 Improvement Monitoring & Feedback: Regularly track performance, collect user feedback, and refine the system (e.g., using approaches like RLHF and DPO - links in comments). Scenario Testing: Test diverse scenarios for reliable AI performance (e.g., employ purple teams for robust validation (link in comments)). Does that mean you always need to build a custom solution? No. Prompting foundational models like Google’s #Gemini or OpenAI’s #GPT4 is a cost-effective and easy to integrate way that can expedite the development of a conversational AI proof of concept. They do, however, lack the flexibility to fully adapt to specific business needs and can sometimes lead to scalability issues. A middle ground exists. You can fine-tune pre-trained models to your specific organisational context. Techniques like #metaprompting, few/zero-shot learning, #LoRa, domain/task-specific fine-tuning, or #RLHF can help you to do that. Curious to learn more about designing and building successful conversational experiences? Comment “GUIDE” below and get a free copy of our #Conversational #AI #Guide!

Highlights from last night’s Apple WWDC! 🚀 Apple is stepping into generative AI with #AppleIntelligence, a privacy-focused system integrated into iOS, iPadOS, and macOS. Here are the key points, announced last night; 1. On-Device Intelligence: - Inspired by #Microsoft's Copilot , it uses pre-trained models for tasks like sketch-to-image conversion and generating summaries. - Available on A17 and M-series chips, starting with iPhone 15 Pro. 2. Privacy Cloud Compute: - Securely handles larger workloads with Apple silicon-powered servers. - Ensures data privacy with third-party audits, setting Apple apart from many competitors. 3. Public Generative AI: - Connects to larger models like ChatGPT (GPT-4) with user permission for data sharing. While promising a significant leap, especially for Siri, initial availability is limited. Full adoption across the Apple ecosystem is likely to take some years. Want to find out more about applying AI in your organisation? Contact us on [email protected] #AppleWWDC DEUS: human(ity)-centered AI #AI #Privacy #Innovation #TechNews

Imagine cycling 254 km from Porto to Santiago de Compostela, pushing your limits, and relying on your team to reach the finish line. That’s exactly what our colleagues did on the Camino de Santiago! 🚴 Their journey wasn’t just about the destination, but also the hilarious moments along the way. Like Paulo “shortcut” through the beach and forest 🏖️ and Márcio realizing that climb was not super smooth for his legs! And talk about dedication—they truly went the “extra mile for the company”, literally! Their adventure reminds us that teamwork is essential in our pursuit of innovation. By combining our diverse strengths, we can push the boundaries of AI and create groundbreaking solutions. 💡✨ Let’s draw inspiration from their epic ride and continue to foster a culture where collaboration fuels our innovation! 🚀🙌 #ai #teamwork #innovation