Zyte’s Post

We’ve found similar results with our fine-tuning experiments at Jam & Tea. We typically prototype against GPT-4 and then fine-tune open weight models. I’ve been surprised at how well we can get those models to parity with GPT-4. #gamedev #ai #machinelearning #startups

Every time that I show these results, people are in disbelief. You can get #opensource #LLMs to perform equally or better than commercial ones by fine-tuning them. Open-source LLMs have been filling the gap with commercial ones over the last year at an astonishing pace, but everyone who has used them experienced issues like repetitiveness, not always following instructions, and they don't reliably perform as well as commercial ones. We fine-tuned 310 open-source models on 31 tasks, and discovered that you can get equal or better performance than commercial LLMs. Some of those datasets contain as little as 2000 data points, so the data requirement and the barrier for doing it yourself is very low. Moreover, the open-source models that we fine-tuned were all 2b-7b! The implication is that you can get that level of performance with substantially smaller, cheaper, and faster models than the commercial ones. This finding goes against the common belief that bigger is always better when it comes to AI, and more people need to know about it, so I'm happy to share it with you. We are running a webinar about this next week, link in the comments 👇 And we are publishing a white paper about all the experiments and results, but in the meantime you can find a summary from our Lora Land release in the comments 👇

In an era dominated by unstructured information, #parsing vital data trapped in complicated documents is crucial. However, messy layouts and formulas reduce parsing accuracy. By combining open-source/proprietary parsers and #LLMs, Proxet team successfully tackled the complexities. Our methodology signals a new direction: blending AI with existing tools increases speed, affordability and data extraction accuracy from unstructured sources. Early results show LLMs' potential to smoothly unlock documents' value and guide better decisions. I’ll add the link to the post in the comments section. Feel free to add your thoughts or ask questions. Stay tuned as we further develop this experiment! Do you want to learn even more about improving your data and AI strategy? Highly recommend following these thought leaders: Alex Wang Steve Nouri Vin Vashishta Louis-François Bouchard 🎙🎥 Greg Coquillo Bernard Marr , Sanjeev Mohan Ravit Jain Brian T. O'Neill Benjamin Rogojan and Andriy Burkov

Knowing HOW to work with the immense amounts of data efficiently is key these days. So proud to see the Proxet team is taking strides in exploring how to use existing and emerging tools to make better decisions faster. #dataparsing #llms #parsing

𝐃𝐚𝐲 𝟗/𝟒𝟎: 𝐂𝐮𝐫𝐢𝐨𝐬𝐢𝐭𝐲 𝐥𝐞𝐚𝐝𝐬 𝐭𝐨 𝐈𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐨𝐧𝐬, 𝐦𝐞𝐞𝐭𝐢𝐧𝐠 𝐉𝐞𝐫𝐫𝐲 𝐋𝐢𝐮, 𝐭𝐡𝐞 𝐯𝐢𝐬𝐢𝐨𝐧𝐚𝐫𝐲 𝐟𝐨𝐮𝐧𝐝𝐞𝐫 𝐨𝐟 𝐋𝐥𝐚𝐦𝐚𝐈𝐧𝐝𝐞𝐱! Our never-ending curiosity keeps sparking new inventions, always moving the limits of what we can do. One such breakthrough is LlamaIndex, an 𝐨𝐩𝐞𝐧-𝐬𝐨𝐮𝐫𝐜𝐞 𝐟𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 that empowers businesses to leverage their own data and large language models (LLMs) to create powerful applications. On the exciting 9th day of my visit to San Francisco, the hub of innovation, I had the incredible opportunity to meet the brilliant mind behind LlamaIndex, Jerry Liu. LlamaIndex is a framework that helps you transform your company's data into production-ready LLM tools. Here's what Llamaindex offers you: - 𝐃𝐚𝐭𝐚 𝐋𝐨𝐚𝐝𝐢𝐧𝐠: Load over 160 data sources and formats, including structured data (APIs) and unstructured data (PDFs, documents). Intelligent Indexing: Store and index your data for diverse use cases, ensuring efficient retrieval. - 𝐒𝐞𝐚𝐦𝐥𝐞𝐬𝐬 𝐐𝐮𝐞𝐫𝐲𝐢𝐧𝐠: Orchestrate production LLM workflows tailored to your specific data, maximizing its potential. - 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐌𝐚𝐝𝐞 𝐄𝐚𝐬𝐲: Evaluate your LLM application's effectiveness with built-in modules. Measure both data retrieval and LLM response quality for continuous improvement. This is just the beginning. Stay curious, delve deeper into generative AI, and harness the power of LLaMAIndex and other cutting-edge tools to propel your business forward. Keep learning and keep exploring the amazing possibilities that AI offers! --------------------------- If you want to unlock the potential of Gen AI for your business using these cutting-edge Gen AI tools then: - Book a call today at https://lnkd.in/dXBzQ9p7 - Explore our GenAI stack at https://aiplanet.com/ Here's to 40 days of learning, growing, and keeping you updated on all things SF! Stay tuned! #40daysofgenai #Llamaindex #innovation

Navigating the EU AI Act: Introducing the AI Act Requirements Graph As the EU AI Act moves closer to becoming law, understanding its complex web of requirements is crucial for anyone working with AI systems. That's why I'm excited to share my latest project: the EU AI Act Requirements Graph. This interactive notebook provides a clear, visual way to explore the relationships between the Act's key requirements for providers of high-risk AI systems. 🌐📊 The notebook features: 🔍 Interactive graph visualisation using yWorks' yFiles library: Easily zoom, pan, and click on nodes to dive into the details of each requirement and understand how they connect. 🎯 Subgraphs for targeted insights: Explore specific aspects of the AI regulatory landscape, such as requirements under a given Article or nodes linked by relationship types like "REQUIRES" or "SEE_ALSO". 🔑 Identify key requirements at a glance: The graph's centrality analysis highlights the most critical requirements, so you can focus your compliance efforts. 🖥️ Open-source and accessible: The project is available on GitHub, and the notebook can be run in Google Colab — no complex setup required. Simply clone the repo and start exploring! Whether you're an AI developer, business leader, policymaker, or simply interested in the future of AI regulation, the EU AI Act Requirements Graph is a helpful tool for navigating this complex landscape. I'd love to hear your thoughts on the project and how it could be enhanced to better serve the AI community. 🧑💻GitHub repo - https://lnkd.in/eQ5n8RAf 📓Colab notebook - https://lnkd.in/eBziKtyc Thanks to Phil Lee for his AI Act infographics, which were the inspiration for this project. #EUAIAct #AIAct #AIRegulation

🚀 Today March 7th, 2024 – Join us at 11,30 for an interesting workshop organized by the Orbis Project partners! 🤖📊 Titled "AI, Big Data, and Democracy Task Force," this workshop dives deep into the intersection of Artificial Intelligence (AI), Big Data, and democratic governance. In today's society, these technologies wield immense power, shaping our perceptions and interactions with democracy. However, they also pose challenges, potentially creating anti-democratic environments. 💡 The Task Force, comprised of projects like AI4Gov, ITHACA Project, KT4Democracy, and #ORBIS, tackles this challenge head-on. Operating within an interdisciplinary framework, it explores ethical and socio-technical considerations to ensure that AI aligns with human values and promotes positive societal outcomes. 🌐 Drawing from democratic theory, the Task Force emphasizes principles of participation, representation, and deliberation. It delves into ethical dimensions, prioritizing fairness, transparency, accountability, and privacy. Additionally, discussions revolve around governance models and innovation theories, seeking effective management and regulation of emerging technologies. #AI #BigData #OrbisProject #ResponsibleAI #HEurope #HorizonEurope Politecnico di Milano The Open University Anna De Liddo Lucas Anastasiou Ilaria Mariani Michelangelo Secchi Grazia Concilio https://lnkd.in/eeXyX3wQ

Unlocking LLMs Superpowers with RAG Pipelines 🔐 In a recent blog post, we dissected the intricacies of Large Language Models (LLMs) and laid out a clear blueprint to seamlessly integrate these AI powerhouses with your business-specific data. LLMs have revolutionised the way we think about problem-solving in AI. They've shifted the dynamic from highly specific, data-intensive models to more adaptable, general-purpose frameworks. With these flexible tools, businesses can bypass the heavy lifting of data labelling and large-scale data science teams, leading to cost-effective, rapid application development. The beauty of LLMs lies in their ability to transcend industry barriers, offering out-of-the-box functionalities that can be fine-tuned to specific business needs.Redactive emphasises the right approach to LLMs—leveraging general model functionality and enhancing it with prompt engineering for a tailored result that truly enhances productivity and growth. Retrieval Augmented Generation (RAG) provides a streamlined method to infuse LLMs with business-specific context. It offers a simpler approach to introduce additional information to AI applications without the complexity of traditional data pipelines. RAG sidesteps limitations related to the volume and structural nature of information that can be prompted to an LLM. Redactive simplifies this further. Forget about the complex data management that comes with chunking and vector databases. Our managed RAG solutions ensure easy, secure integration with your document databases, effectively turning your proprietary information into the superfuel for LLMs. Read the full article below to dive deeper into the transformative potential of LLMs with RAG pipelines 👇 #LLMs #RAG #RAGpipeline #ArtificialIntelligence #DataIntegration #RedactiveAI #Technology #Innovation

The iQuantNY blog was a great success illuminating what's actually in open data. Inspired by that, I had a little fun creating a quick gpt-enabled blog post tongue-in-cheek-titled aiQuantCA. It's a little analysis of groundwater levels from the data.ca.gov open data portal. It's... kinda interesting? But then not really useful per say. It certainly has the first pass appearance of looking smart. Sometime AI can be helpful with tasks, sometimes it feel like this sort of open ended inquiry isn't exactly increasing the signal to noise ratio. https://lnkd.in/giUktv4e Thoughts for future refinements: I could see AI particularly succeeding in more semi-structured tasks like monitoring a certain dataset and flagging outliers based on certain criteria. This type of open ended inquiry is getting shinier as gpts progress but then isn't really insightful per say.

Last week, I attended the #TheAISummit in London, where I participated in talks from business leaders across various domains and saw tech demos from some of the world's largest companies. Unarguably, the main focus of the summit was Natural Language Generation, specifically LLMs, and the hardware needed to run and build these models. Overall, the insight I gained from these talks and demos is that companies, generally, have no idea how to turn LLM use cases into consumer-facing products. Most use cases described were internal tools or small customer-based tests where the implications of LLM usage hadn't been thought through with a user-based focus, ultimately causing problems. Companies were generally more concerned about the risks and "safety" around LLM usage rather than the actual customer use case itself. In the case of regulators, they seemed to ignore the advances in AI and maintained that their regulations already accommodate these advances. It's clear that many companies are putting other machine learning projects on the back burner to pursue what's hot right now, rather than evaluating what their users need and what the best tools for the job are. Alternatively, regulators seem paralysed by fear due to the overwhelming scale of the task ahead and their inability to keep pace with progress. I've said it before, and from what I've seen, my opinion hasn't changed: companies need to go through a data science and user experience-focused process and use the tools that will get them to their goals rather than the tools everyone is talking about. There is definitely a trade-off here where "Generative AI" is a buzzword for businesses and share valuations, but it's important to remember that a text template using a backend database query to fill in the gaps is also "generative", doesn't cost millions to build, and doesn't use the power consumption of a small country to achieve the same end goal in some cases. My general advice for business leaders is to get the experts in the door, actively listen to what they have to say, and remember that just because something is or isn't an LLM, it doesn't mean your objectives will be fulfilled any quicker. For data scientists, it's important to remember that business leaders operate under a different set of constraints, so listen to them, try to understand why they might need to list "Generative AI" as one of their solutions, and build something that allows them to do that (maybe text templates 🤔 😉 ). As always, let the data and analysis do the talking. If the cost, efficiency, and results trade-off point to an LLM, use it. Otherwise, don't! #LLM #AI #DataScience