Jim Dowling

*Real* testing of LLMs—for “Deception” If you want to understand how LLMs are really tested and red-teamed, this is a pretty good discussion. https://lnkd.in/dTtcN-bQ

21

3 kommentarer

🌟 Exciting News! 🌟 On October 24th, I will present at PyData Copenhagen event at Microsoft Denmark "Building and Validating AI Applications in Regulated Industries." We'll explore how to build AI agents that are safe, effective, and truly make a difference in regulated industries like healthcare. Special thanks to Ali Reza Farahnak and Anders Bogsnes for organising the event. 🔍 Session Highlights: - Integration of AI in regulated sectors - Risk identification and assessment during design phases - Validation and testing strategies - Live demonstrations and code examples - Case study focus: A web and mobile application for the cerebral palsy community, designed to ensure easy access to reliable information and comprehensive support. 👥 Engage with us: Beyond this event, me and my partners Fireraven and Holistic AI are keen to collaborate and consult on your upcoming projects. Whether it's developing AI strategies or addressing specific challenges, let's work together to make AI innovation accessible and impactful. https://lnkd.in/d4U-kkS3 #PydataCopenhagen #ResponsibleAI #AIinRegulatedIndustries #Innovation #AI

35

8 kommentarer

💡 Join us for one of the most exciting AI & quantum debates of the year at TechBBQ! On September 12th, I’ll be moderating an AI & innovation geeks dream panel debate on Denmark's AI Vision: Supercomputers, Innovation, and the Startup Ecosystem – and I’d love to get your input! Together with an incredible lineup of experts, we’ll explore Denmark's proactive approach to AI, powered by Gefion, one of the world’s most advanced NVIDIA AI supercomputers. This facility will drive research and innovation across healthcare, life sciences, and the green transition, while unlocking opportunities for startups and entrepreneurs. Here’s who will be joining the debate: 🎙️ Emma Adolfsson (Partner AI & Data, EY) A digital strategist and advocate for the ethical implementation of AI, Emma brings her expertise in navigating the intersection of AI and data to foster responsible innovation. 🎙️ Nadia Carlsten, PhD (CEO, Danish Centre for AI Innovation) With over 20 years of experience in AI, quantum computing, and machine learning, Nadia now leads the Danish Centre for AI Innovation, which operates Gefion, Denmark’s first NVIDIA AI supercomputer. 🎙️ Erik Balck Sørensen (EIFO Investment) A serial entrepreneur and investor, Erik has been a key player in Denmark’s startup scene, co-founding companies like Billetto and Downtown. He’s passionate about how AI can empower new ventures. 🌐 What would YOU like to ask these leaders about AI, supercomputing, or Denmark’s national AI strategy? How will this AI infrastructure shape industries and support entrepreneurial ventures? What technological breakthroughs are you most excited to see in healthcare, sustainability, or the startup ecosystem? Whether it’s healthcare innovation, sustainability, or how AI is reshaping industries, we want to hear your thoughts! 💡 Drop your questions in the comments and I’ll bring them to the debate! 📅 Date: Thursday, September 12th 📍 Event: TechBBQ Looking forward to your ideas and seeing you at this exciting discussion! #AI #QuantumComputing #TechBBQ #DenmarkAI #Innovation

57

3 kommentarer

💥 How AI is saving billions of years of human research time 👉 https://lnkd.in/dgv_rxzK Some of the topics explored: - The sheer scale of new knowledge about our protein universe that we now have access due to AlphaFold, with models replacing the need for real world experimental work - The potential to bring together the exploding world of AI to the increasingly costly world of drug design - Why I believe that new breakthroughs in machine learning have allowed us to produce an AI Analogue of biomolecular systems in the human body - How could we can use AI like AlphaFold 3 to modulate the function of proteins involved in cancer - How we are moving to a world where AI agents can begin to design new drugs purely in-silico, on a computer #machinelearning

3

We introduce Patched MOA (Mixture of Agents), a novel inference optimization technique for large language models. - Patched MOA significantly improves gpt-4o-mini's performance, surpassing gpt-4-turbo on the Arena-Hard-Auto benchmark. - Achieves superior results at approximately 1/50th the cost of gpt-4-turbo. - Applicable to various LLMs and tasks without requiring fine-tuning. - Outperforms other tested methods like Best of N and Monte Carlo Tree Search.

11

Does AI (and AI dev) threaten the open web? Mathias Biilmann Christensen, CEO & Co-Founder of Netlify, makes a strong case why that threat is real, and why it’s important to defend it. It was a pleasure hosting Matt on The AI Native Dev podcast, going deep on many topics, such as: - AI agents have become a new type of Netlify user - how do you adapt to that? - How AI use of a non-AI dev tool (Codemod) boosted their productivity the most - The connecting lines between AI music generation and AI dev New tech paradigms are easier to grasp in closed, opinionated environments, controlled by one entity, than in an open and composable ecosystem. AI is clearly following that path, but where is it leading? Are we ok with all AI dev capabilities coming from our one vendor of choice, or insist on being able to assemble it from best of breed? These are just a few bits of a great episode. Matt’s authenticity and clear passion for AI and the open web meld perfectly with his deep insights, so much that the episode is a bit longer than planned - I just couldn’t stop the conversation! #Netlify #AIWeb #AINativeWeb #AINativeDev #MattBiilmann

61

3 kommentarer

Mistral just released their first code model: Codestral. Fluent in 80+ programming languages it outperforms LLaMa 3 on HumanEval. The 22B open-weight generative AI model is explicitly designed for code generation tasks. It helps developers write and interact with code through a shared instruction and completion API endpoint. - New endpoint via La Plateforme: https://lnkd.in/gKV-YT_B - Try it now on Le Chat: http://chat.mistral.ai licensed under the new Mistral AI Non-Production License, which means that you can use it for research and testing purposes. Codestral can be downloaded on HuggingFace: https://lnkd.in/gzWeHEyw ↓ Are you technical? Check out https://AlphaSignal.ai to get a weekly summary of the top trending models, repos and papers in AI. Read by 180,000+ engineers and researchers.

212

10 kommentarer

Meta's LLama 3 is the first real open source LLM competing with OpenAI and GPT-4 or outperforming it in certain benchmarks. Does this mean that OpenAI is losing the game? In some areas that might be true, but for any person speaking Finnish or some other Finno-Ugric language, or any other rare language, GPT-4 still outperforms all other closed or open models. The current benchmarks do not measure these real-world use cases: they measure how well the models perform some expert tasks or answer tricky questions. And that's why most users will still stick to OpenAI or other models that are the most useful for everyday pragmatic use.

16

3 kommentarer

GPT-4o was released yesterday. It is claimed to be both faster and cheaper than its predecessor. Is it worth switching to the new model? At DEMA, we have compared GPT-4-Turbo and GPT-4o on our test suite for generating specific API queries from user questions in the domain of data analysis. From our initial testing, we have observed: ~3x Faster Inference: Significantly faster response times across a wide variety of prompts. A slight trade-off in precision with prompts containing more elaborate instructions. Is it worth the switch? It probably depends on the use case: For simpler tasks, GPT-4o performance seems to (at least) match its predecessor while being both faster and cheaper. For more complex tasks, some differences start showing, and careful evaluation and testing need to be carried out before adopting the new model. We'll be investigating ways to maintain and improve accuracy while enjoying the benefits of much faster inference. Exciting times ahead, as this enhancement promises significant improvements in user experience.

41

4 kommentarer

New Drama in OpenAI 🚀 Ilya Sutskever (co-founder & Chief Scientist) and Jan Leike have quit OpenAI The Two most respectful AI scientists in the world lead the OpenAI Superalignment team, which was set up to ensure that AI systems will be safer. It is not the most confident thing, and then two foremost safety scientists leave the company. I remember when the AI field didn’t have drama; it just had interesting papers and excitement. I hope Ilya creates his new company; I speculate that he will join Elon Musk because Elon convinced Ilya to join OpenAI. #ai #openai #samaltman #llm

146

6 kommentarer

Join us on Dec 18 for the Stockholm AI, ML and Computer Vision Meetup at Fotografiska. Reserve your spot: https://lnkd.in/gwWcmcTK Alessandro Pieropan of Univrses will be presenting “How to Maintain a High Quality Map in a Scalable Way.” In this talk he’ll discuss one of the biggest challenges in HD mapping, which consists of keeping a high quality map up to date while reducing the human supervision close to zero. In this presentation, he will show the methodologies used at Univrses to build a scalable solution that allows to keep a high fidelity digital twin of the city while ingesting a continuous stream of data sent by a fleet of vehicles of arbitrary size. Got an interesting use case or research project you’d like to share? DM me about speaking opportunities at the Stockholm Meetup! #computervision #ai #artificialintelligence #machinevision #machinelearning #datascience #stockholm Daniel G. Robert Wright

9

Bitmovin's Split-and-Stitch encoding disrupted the status quo for video-on-demand processing, and enabled huge leaps in turnaround time improvements. We've spent the past decade refining and enhancing our split-and-stitch algorithm to increase efficiency and improve quality with innovations like Smart Chunking. Check out our latest blog to learn more about how Bitmovin's split-and-stitch rises above the competition and can speed up your encoding workflows without sacrificing any quality. https://okt.to/bX2Kn6 #videoencoding #Encoding

16

1 kommentar

I ❤️ TensorFlow. It’s one of the most versatile solution in developing and deploying deep learning models in the real-world. Here are 15 operations within TensorFlow I use often in my deep learning workflows👇 1. tf.keras.layers.Dense() # The core building block for NN 2. tf.keras.optimizers # Optimization (e.g., Adam, SGD) 3. tf.reduce_mean() / tf.reduce_sum() # Common math operations. 4. tf.data.Dataset # Create data pipelines for training 5. tf.GradientTape() # Automatic differentiation. 6. tf.keras.losses # Loss functions 7. tf.nn.softmax() # Essential for classification tasks 8. tf.keras.layers.Normalization() # Normalize continuous data 9. tf.cast() # Convert data types 10. tf.keras.layers.Dropout() # Reduces overfitting 11. tf.where() # Filter data 12. tf.argmax() / tf.argmin() # Common math operations 13. tf.reshape() # Reshape dimension of data 14. tf.keras.metrics # Model evaluation 15. tf.keras.layers.Embedding() # Creates embedding vectors Know these functions for your day-to-day jobs in ML engineering and data science (specialized in deep learning), and, for ML coding interview questions! 👉 What are your must-know TensorFlow (or PyTorch) operations? Drop a comment👇 👉 I will be teaching 𝗗𝗦 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗕𝗼𝗼𝘁𝗰𝗮𝗺𝗽 (https://lnkd.in/eiA5Ntdp) next week! Join the June cohort before it starts! Ace upcoming interviews with these👇 📕 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗣𝗿𝗲𝗽 𝗖𝗼𝘂𝗿𝘀𝗲𝘀: https://lnkd.in/gzgB-dHT 📙 𝗝𝗼𝗶𝗻 𝗠𝗟𝗘 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗕𝗼𝗼𝘁𝗰𝗮𝗺𝗽: https://lnkd.in/e6HbN6dy 📗 𝗔𝗕 𝗧𝗲𝘀𝘁𝗶𝗻𝗴 𝗖𝗼𝘂𝗿𝘀𝗲: https://lnkd.in/g82dMJ77 👉 Smash 👍 and follow Daniel Lee to land dream data & AI jobs 🚀

156

5 kommentarer

I ❤️ TensorFlow. It’s one of the most versatile solution in developing and deploying deep learning models in the real-world. Here are 15 operations within TensorFlow I use often in my deep learning workflows👇 1. tf.keras.layers.Dense() # The core building block for NN 2. tf.keras.optimizers # Optimization (e.g., Adam, SGD) 3. tf.reduce_mean() / tf.reduce_sum() # Common math operations. 4. tf.data.Dataset # Create data pipelines for training 5. tf.GradientTape() # Automatic differentiation. 6. tf.keras.losses # Loss functions 7. tf.nn.softmax() # Essential for classification tasks 8. tf.keras.layers.Normalization() # Normalize continuous data 9. tf.cast() # Convert data types 10. tf.keras.layers.Dropout() # Reduces overfitting 11. tf.where() # Filter data 12. tf.argmax() / tf.argmin() # Common math operations 13. tf.reshape() # Reshape dimension of data 14. tf.keras.metrics # Model evaluation 15. tf.keras.layers.Embedding() # Creates embedding vectors Know these functions for your day-to-day jobs in ML engineering and data science (specialized in deep learning), and, for ML coding interview questions! 👉 What are your must-know TensorFlow (or PyTorch) operations? Tell me in the comments! 👇 Get an edge in your interviews with these resources👇 🚀 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗣𝗿𝗲𝗽 𝗖𝗼𝘂𝗿𝘀𝗲𝘀: https://lnkd.in/gzgB-dHT 📊 𝗝𝗼𝗶𝗻 𝗗𝗦 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗕𝗼𝗼𝘁𝗰𝗮𝗺𝗽: https://lnkd.in/eiA5Ntdp 🧠 𝗝𝗼𝗶𝗻 𝗠𝗟𝗘 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗕𝗼𝗼𝘁𝗰𝗮𝗺𝗽: https://lnkd.in/e6HbN6dy 👉 Smash 👍 and follow Daniel Lee to land dream data & AI jobs 🚀

277

11 kommentarer

Via Jürgen Geuter... LLMs are great at summarizing long documents, if, by summarizing, you mean making up grammatically correct sentences that probably don't reflect the general content and important details of the original document. Here's a shortcut to using LLMs correctly. Anything you want to do with LLMs, especially if it involves shortcutting human expertise you would otherwise have to pay for... Any activity type you want to do... Do ten real world tests in your real world, where you have the LLM do it, you have a competent human do it, and you have a human competent in the task judge the results. After 10 real world tests, you might have enough data to honestly pick a winner. That's what the Australian government did here. Side note... This is how I figured out that something wasn't as it seemed with this technology. This is how I encourage my clients to discover these truths for themselves. When someone tells you, you have to take their word for it. When you can experience it yourself, it sticks quickly. #WrittenByMe

18

As promised, sharing news for Integrail platform :) -- this time again about working with Images. At launch on September, 11th we will have at least ten (10) text-2-image models that should cover all of your needs! Updating our 45-prompt comparison to all of them is too much, but here are all ten generating image for the following prompt: "A beautiful female avatar for the AI agent in the post-apocalyptic environment, face portrait and in the background - post apocalyptic New York environment, dramatic lighting, HD render" On top of simple image gen, stay tuned for a couple of tutorials showing how to work with supplementary "nodes" such as upscaling, sketch improvement, inpainting, etc, etc. With this, you should be able to build AMAZING image manipulation AI agents in no time. Subscribe to Integrail updates and create with us! #ai #genai #text2image

28

3 kommentarer