Srujana Bobba

🚀 WHY MASTERING LLMs IN 2025 IS NON-NEGOTIABLE! 💡 What are LLMs? Large Language Models like ChatGPT, Claude, or Gemini are reshaping industries. These AI systems understand, generate, and analyze text, code, and more, unlocking unprecedented productivity and innovation. Key Features of Modern LLMs 🧠 Parameters: Billions (even trillions!) powering human-like intelligence. 📚 Training Data: Trillions of tokens processed for knowledge. ⚡ Compute Power: Thousands of GPUs enabling advanced performance. 💸 Costs: Millions invested to build state-of-the-art capabilities. How LLMs Work: Must-Know Basics Context Window: Defines how much AI “remembers” in one go. Temperature: Adjusts creativity (from factual to imaginative). System Prompts: Fine-tune AI’s behavior for specific tasks. Function Calling: Connects AI to real-world actions and APIs. Why You Should Care Learning LLMs will: ✅ Future-proof your career ✅ Empower smarter solutions ✅ Position you as a leader in AI adoption Top Skills to Master in AI - Prompt Engineering: Craft impactful instructions for better outputs. - Retrieval-Augmented Generation (RAG): Combine databases with AI. - Fine-Tuning: Customize AI for specialized use cases. - Quantization: Optimize performance and reduce costs. - Caching: Deliver faster responses to users. 💡 Stay Ahead, Stay Responsible Always validate AI outputs. Filter for ethical and accurate usage. Detect and address AI “hallucinations.” 👉 Be the professional who leads with LLMs in the AI era. #FutureOfWork #GenerativeAI #LLMs #AILeadership #CareerDevelopment

19

4 Comments

👉🏼 How effective are large language models in product risk assessment? 🤓 Zachary A Collier 👇🏻 https://lnkd.in/eZy9nVrh 🔍 Focus on data insights: - Large language models, like ChatGPT, show promise in brainstorming potential failure modes and risk mitigations. - However, there were errors and inconsistencies in some results. - The guidance provided by the models was perceived as overly generic and occasionally outlandish. 💡 Main outcomes and implications: - ChatGPT performed better at divergent thinking tasks but lacked depth of knowledge compared to subject matter experts. - Similar patterns in strengths and weaknesses were observed when comparing ChatGPT to other large language models. - Despite challenges, there may be a role for large language models in assisting with ideation in product risk assessment. 📚 Field significance: - Product safety professionals can leverage large language models for brainstorming and idea generation in risk assessment processes. - Experts should focus on critically reviewing AI-generated content to ensure accuracy and relevance. 🗄️: [#productrisks #AI #riskassessment #ChatGPT]

1

1 Comment

Amidst a range of commercial black box LLMs from large enterprises, we are also seeing Open-Source Machine Learning Frameworks for Advancing Large Language Models in Healthcare. Check out Hippocrates, an open-source LLM framework specifically developed for the medical domain https://lnkd.in/gw5xHdUD

1

👉🏼 Reliability of large language models for advanced head and neck malignancies management: a comparison between ChatGPT 4 and Gemini Advanced 🤓 Andrea Lorenzi 👇🏻 https://lnkd.in/eSW4Admy 🔍 Focus on data insights: - ChatGPT 4 generally outperformed Gemini Advanced in adherence to guidelines and comprehensive treatment planning. - Higher AIPI scores were observed with ChatGPT 4, indicating better overall performance. 💡 Main outcomes and implications: - Both large language models showed potential in aiding multidisciplinary management of head and neck oncology. - Discrepancies in critical areas emphasize the need for further refinement and validation against current clinical standards. 📚 Field significance: - Growing role of AI in enhancing clinical decision-making. - Necessity for continuous updates and validation to fully integrate AI into healthcare practices. 🗄️: [#largeLanguageModels #headAndNeckMalignancies #AIclinicalDecisionMaking]

1

1 Comment

The Evolution of AI: From Automation to Replacement AI systems typically progress through three key phases. Understanding where a solution stands in this progression can be insightful, especially amidst the buzz around LLMs and Generative AI. Here’s a breakdown: ▪ 1. Automate  The initial phase focuses on automating routine tasks to save time and reduce errors. For example, at Lucid Lane, we’ve automated administrative tasks, aiding our therapists by summarizing sessions and sending reminders to patients and therapists alike. While workflow automation tools exist, they’re just the beginning of the AI journey. ▪ 2. Enhance  As AI matures, it begins to augment human capabilities, allowing for scalability. The sophistication of AI in this phase correlates with fewer errors and greater human efficiency. Examples include Microsoft Copilot, GitHub Copilot, and Google’s auto-fill features. At Lucid Lane, our AI analyzes patient history and suggests medication dosages, aiding therapists and doctors in decision-making. Real-time suggestions during sessions can also enhance therapist-patient interactions, exemplifying AI systems with a human-in-the-loop. ▪ 3. Replace  The final phase is when AI becomes reliable enough to replace certain human functions. While companies may be cautious about claiming their solutions can fully replace humans due to the fear and negativity it may evoke, some are boldly making such claims. For instance, Devin, the touted first AI software engineer, represents this claim. Although I approach these assertions skeptically, it’s undeniable that some AI systems will reach this phase. It will become evident as AI enhancements repeatedly surpass the need for human intervention. These phases signify deeper AI integration into operational processes, aiming to improve efficiency, accuracy, and outcomes. However, it’s crucial to remember that despite AI’s potential to replace some functions, human oversight, ethical considerations, and handling complex situations remain indispensable. This progression not only reflects the future of work but also raises questions about human judgment and the ethical implications of AI, particularly in sensitive sectors like healthcare. What are your thoughts on the future of AI in your industry? Have you witnessed these phases in action? #generativeai #LLMs #EthicalAI #ai

19

1 Comment

No amount of scaling (throwing billion$ for computational resources & trillions of words) would help LLMs achieve the two key cognitive capabilities that make human intelligence so special: 1. Learn - Autonomously, Adaptively, Conceptually, Contextually & Incrementally. 2. Reason - Conceptually, Contextually & Metacognitively We discussed the criticality of this in greater detail here: https://lnkd.in/ggZm_W-w It is for this reason, LLMs can never get us to AGI: https://lnkd.in/d3vEPJ2P So, what will get us to AGI: https://lnkd.in/ggFTxuUz It's no wonder Demis Hassabis, CEO of Google DeepMind said this about LLMs Intelligence on July 9, 2024 at the Future of Britain Conference hosted by Tony Blair: “We’re still not even at cat intelligence yet, as a general system,” https://lnkd.in/gtyQNZyc It is great to see Yann LeCun arriving at the same conclusions about LLMs. https://lnkd.in/gm9Nfe4P

342

68 Comments

Google DeepMind, a leader in AI research, has released a groundbreaking 274-page paper titled "The Ethics of Advanced AI Assistants" (link: https://lnkd.in/eQzS4ebn). This thought-provoking exploration delves into the intricate ethical dilemmas surrounding advanced AI assistants. Designed for a broad audience, the paper caters to developers, policymakers, researchers, and the public. Each chapter can be read independently, allowing readers to focus on specific areas of interest, such as technical safety, privacy, or security. To delve into specific areas, consider starting from page 17. This release is particularly timely in light of the Department of Homeland Security's recent establishment of the Artificial Intelligence Safety and Security Board (link: https://www.dhs.gov/ai). Chapters 8, 12, 13, and 19, which explore privacy, trust, and security, offer valuable insights that can inform discussions and guide the board's crucial mission of ensuring responsible AI development.

61

3 Comments

This weekend I read "The landscape of emerging AI agent architectures for reasoning, planning, and tool calling: a survey" by Masterman et al., 2024. In this paper the authors review AI agents implementations. The AI agent term is used to define LLM systems that tackle more complex tasks that require planning, reflections and actions that use the "reasoning" capabilities of LLMs, in contrast with more zero-shot tasks (eg RAG systems). The paper focuses on comparing the multi-agent vs single-agent architectures to solve different tasks taking into account their architecture, capabilities and tools that they use. To me, this paper is important because it gives a good general overview of the AI agent systems (and I am new to the topic). I find very interesting thinking about tasks and problems that can be tackled using this "agentic approach". For me, that in itself it is a skill (that I need to improve on). Enjoy! Paper: https://lnkd.in/gucNytmF #LLM #AIAgents #genAI

3

2 Comments

Today the most important article in chatgpt and "Detection of ChatGPT fake science with the xFakeSci learning algorithm" - The study introduces the xFakeSci learning algorithm designed to distinguish ChatGPT-generated articles from those produced by scientists. - It demonstrates the algorithm's effectiveness through testing against various diseases like cancer, depression, and Alzheimer's. - xFakeSci achieved F1 scores between 80 to 94%, surpassing traditional data mining algorithms significantly. - The algorithm's success is attributed to the incorporation of calibration and proximity distance heuristics, enhancing its performance in identifying fake science generated by ChatGPT. #AI #FakeScience #xFakeSciAlgorithm #ChatGPT #DataMiningAlgorithms

1

It’s weird how LLMs hold information‘directionally’. Inside an LLM, there are millions of layers and connections that process language. When an AI decides to decline a request (like refusing to generate inappropriate content), or can answer a question in one direction but not in another, this decision isn’t random—it follows a specific “path” within its activations. This path is essentially a direction in the vast space of the model’s neural activations. Researchers have identified that by altering this single directional subspace. For example, adjusting this subspace could fine-tune how strictly the model follows ethical guidelines. This discovery is groundbreaking because it isolates a specific, controllable factor that governs refusal behavior, rather than it being scattered across the entire model. Why does this matter? It gives developers more precise control over AI behavior. Imagine an AI tuned to always enforce strong ethical boundaries while remaining highly useful in professional settings—this research could make that a reality. Understanding this directional mechanism also opens the door to safer, more customizable AI systems.

5

Stanford published the AI Index Report for 2024. The main points are listed below. Link to the report in the first comment. 1. AI beats humans on some tasks, but not on all. 2. Industry continues to dominate frontier AI research. 3. Frontier models get way more expensive. 4. The United States leads China, the EU, and the U.K. as the leading source of top AI models. 5. Robust and standardized evaluations for LLM responsibility are seriously lacking. 6. Generative AI investment skyrockets. 7. The data is in: AI makes workers more productive and leads to higher quality work. 8. Scientific progress accelerates even further, thanks to AI. 9. The number of AI regulations in the United States sharply increases. 10. People across the globe are more cognizant of AI’s potential impact—and more nervous. #artificialinteligence #ai

4

1 Comment

We’re getting to that point in a technology’s journey where we open or keep closed source. Meta just released Llama 3.1, the largest-ever open-source AI model, which apparently outperforms GPT-4o and Claude 3.5 Sonnet on several benchmarks. MZ says that Meta AI will be the most widely used assistant by the end of this year! Interesting times ahead for what is set to be the biggest tech advance since the internet. Where do you sit in terms of open and closed source and does the same rule apply for AI? #ai #genai #chatgpt #metaAI

11

1 Comment

3 levels of AI voice agents 1️⃣ Put together speech to text (STT), an LLM and a text to speech (TTS) Simply pipe the voice of a human through the STT to get the transcription. Feed it to an LLM to get the answer and send it to TTS to get the response in sound wave You now have a bare bone intelligent agent that can carry a conversation with human intelligently However this doesn't handle many nuances like ensuring AI agent stop talking when the human interrupts with "No" but keep talking when it's "uh huh," or "exactly" 2️⃣ Use an AI voice agent platform like Retell, Vapi & Daily Bot These platform handles lots of things that are challenging to do including integrating with API, knowledge base, multi-step flow You can get incredibly far leveraging these platforms. But because they are a generic purpose platform, to really go into production you need level 3 3️⃣ Production-ready AI voice agent Our use case is make outbound call to follow up with leads or engage prior customers. Here are just a few of the things we have to implement for our agents to make it production ready. - Auto redial: Most people won't pick up the first time but if you dial back 5 or 10 minutes later they will. Implementing this increasing our connect rate by 26% - Reschedule: If the lead is busy, they'd likely request a callback. Saying it'll call back, confirming the time and actually call when the time come are not trivial - Don't call at 2am: We once call a customer at 2am. In our defense, they just filled out the lead form Jokes aside, we had to implement a schedule when we can make call. It needs to be local to the timezone of the customers Did I tell you how much I love date time math? - Keep calling: Turn out calling 1 or 2 times doesn't cut it. You have to call up to 6 times. We had to build a campaign functionality. Doing this increase our connect rate to almost 76% There are a many more things we had to implement to make even the best AI voice agent production ready for our use case. If your use case is simpler, using Retell or Vapi will suffice most of the time #AI #sales #marketing

22

4 Comments