Rob Murtha’s Post

A strong threat model can create massive positive ripples across your organization!⁠ ⁠ Did you know? This September, we're again hosting our "Rapid Threat modeling with GenAI & LLMs" bootcamp. It's your chance to master threat modeling with the speed and precision of AI.⁠ ⁠ Use coupon 'MONTHSTOMINUTES' and save $300: https://lnkd.in/dqqdAnDg ⁠ #RapidThreatModeling #GenAI #LLMs #Bootcamp #MonthsToMinutes⁠

A few weeks back, we delved into the fascinating world of Support Vector Machines (SVMs) and the intriguing concept of lifted functions. The idea that struck me the most was how we can simplify a complex, non-linear function by transforming it into a linear function in a higher-dimensional space. This technique, known as "lifting," is like finding a secret passage in a maze, allowing us to solve intricate classification problems that seemed impossible in their original form. We also explored the importance of learning linear separators in SVMs. Imagine trying to separate two groups of people standing in a field, using only a long rope. SVMs are like expert rope handlers, finding the best way to lay the rope that creates the maximum distance between the groups. This ability to find the optimal hyperplane makes SVMs a robust tool for classification, even when the data is a bit messy. But the real magic happens when we combine lifting with SVMs. By transforming our data into a higher-dimensional space, we can use the kernel trick to efficiently compute dot products without having to calculate the transformed features explicitly. It's like being able to solve a complex jigsaw puzzle without having to look at every single piece. This not only saves computational resources but also allows us to handle infinite-dimensional feature spaces, enabling the model to learn complex decision boundaries. One area where this concept could be a game-changer is Fraud Detection in Financial Transactions. Fraudulent transactions often exhibit complex, non-linear patterns. For example, fraudsters may vary transaction amounts and timing to avoid detection. By lifting these features into a higher-dimensional space, we might be able to make fraudulent and legitimate transactions more linearly separable, enabling us to find a separator hyperplane that effectively distinguishes between the two. #MachineLearning #DataScience #SVM #LiftedFunctions #LinearSeparators #TechSkills #AI #LinkedInLearning

If you are interested in learning about AI Agents, a technology that allows LLMs to make Plans and use Tools (i.e. searching the web), watch this video where I cover the basic concepts and most common challenges of this technology. An example of how a Travel Assistant Agent looks is also shown and discussed.

🏆 Thrilled to share our project from AI ATL 2024 Hackathon: "Are You Sure About That?" - a game-like platform that puts humans against Claude 3.5 in detecting flawed AI reasoning. Built natively in Python using Streamlit. Deployed with DigitalOcean, Cloudflare, and Docker, Inc. With the rise of LLMs, we wanted to tackle a crucial question: Can AI models spot unfaithful reasoning in other AI outputs? And how do humans compare? Our submission to the Alignment track explores this critical aspect of AI safety. The platform tests both humans and AI on their ability to identify flawed logic in AI responses, creating a unique benchmark for reasoning capabilities. 🎯 Key Features: • Interactive Chain-of-Thought evaluation • Real-time AI vs Human performance tracking • Human feedback collection for model improvement • Comprehensive scoring system Video: https://lnkd.in/ex_tKhRy Github: https://lnkd.in/ej3WM62z Devpost: https://lnkd.in/e3NV3gYF 🔍 Try it yourself: https://lnkd.in/eMCgp6Dq 📚 Read more about our methodology and findings in our alignment track paper: https://lnkd.in/e6PaiCSp Thanks to AI ATL for organizing this amazing event! Looking forward to more opportunities to contribute to AI safety research.

🌟 Anthropic Unveils Claude 3's Family. Competition is nice. Anthropic just rolled out the Claude 3 model family, and the spotlight's on 'Opus' A model that is probably leading the race, sprinting past GPT-4 and Gemini Ultra... Lineup 🔍 ➜ The family: Opus, Sonnet, Haiku. Each designed with unique strengths, balancing capability, speed, and cost. ➜ Opus comes with unparalleled expertise in knowledge, reasoning, coding, and even math. ➜ Compared to GPT-4, Claude 3 brings to the table vision capabilities that rival the best. ➜ With a staggering 99% recall on long context prompts. (this is needed!) ➜ 200k context window, scalable to 1M. (no detail gets left behind) ➜ API access, inviting creators to weave Claude 3's capabilities into their digital worlds. Claude just stealthily clinched the AI throne. #Claude3 #AIRevolution #TechTrends #OpusOverlords #SultanSpeaks #CuriousHabibis

The most important talks in AI are happening at AI_dev: Open Source GenAI & ML Summit Europe! Check out "Code Security Reinvented: Navigating the Era of AI" with Joseph Katsioloudes of GitHub & register to join us in Paris from 19-20 June. Reserve your spot by 2 June & save US$150! https://hubs.la/Q02vzLHK0 #AIDev #GenAI #GenerativeAI #ML #AI #MachineLearning #OpenSource #OSS #GenAI #LinuxFoundation #events #OpenSourceDevelopment #OpenSourceCommunity #OpenSourceSoftware

Last weekend I participated in my very first hackathon, the Inception Studio AI Buildathon with four awesome engineers: Mario, Eoin, Przemek, Vitou. We had fun building an AI assistant called Joshu (助手, assistant in Japanese), using simple AI Agents. John Whaley's feedback was the cherry on the cake. We had built a technically great solution, fast. But we lacked communication and product vision: how to deploy it in a real-world scenario while ensuring safe data separation. Data separation is an obvious specification from an engineering point of view, but for a listener, it is not. Communication! Pitch + demo https://lnkd.in/gvnzwK3U Source code https://lnkd.in/gWqwrpCQ

TWO WEEKS LEFT to compete for $10,000 in prizes while improving the quality of prompt datasets useful for identifying algorithmic bias in LLMS. Our first bias bounty program builds on the 100k+ prompts collected at the Generative AI Red-Teaming challenge. Our unique dataset collected adversarial attacks on Factuality, Bias, and Misdirection. Our challenges are built for all levels of experience, and we are hosting an in-person hackathon July 9th with All Tech Is Human for you to learn more. https://lnkd.in/gKUn533Y

Short but an impactful course packed with ideas on red teaming #LLM Applications. Focus on testing at #scale, automated vulnerability/bias testing. If you're like me, concerned about security, bias and fairness in the new world of #GenAI applications, this is a great place to start. #course #AI #deeplearning #course #certification

Excited to have participated in our department's HackMaster Hackathon! 💻 We created a project using Machine Learning/AI for loan approval and insurance recommendation systems. It was an intense and rewarding experience working on real-world applications of AI. 🚀 Grateful for the teamwork and looking forward to applying these skills in future projects! #Hackathon #HackMaster #MachineLearning #AI #LoanApproval #InsuranceRecommendation #TechInnovation