Brian T. O'Neill’s Post

Every company will tell you, “We do a really good job in checking out our AI systems.” That’s great. We want every company to do a really good job. But we also want independent oversight of somebody who’s outside the company — someone who knows the field, who’s looked at systems at other companies, and who can bring ideas and bring understanding of the dangers as well. These systems operate in an adversarial environment — there are malicious actors out there who are causing trouble. You need to understand what the dangers and threats are to the use of your system. You need to understand where the biases come from, what dangers are there, and where the software has failed in other places. You may know what happens in your company, but you can benefit by learning what happens outside your company, and that’s where independent oversight from accounting companies, from governmental regulators, and from other independent groups is so valuable. - Ben Shneiderman ⬆️ Join 2k readers: tap “Visit my website” to get my latest episodes and show summaries in your inbox 🎧: https://bit.ly/4brqXUR

If you are someone who has just started out exploring LLMs for some of your conventional ML projects and come across terminology like Fine-tuning LLMs and RAG, you might wonder which one would best suit your use case. This article explains Fine tuning and RAG, when to use what, and how they could complement each other. A very interesting insight I got from this article is how the author explains that it is a false dichotomy when we try to compare RAG vs FIne tuning because they are not opposing techniques but rather complement each other for improved performance. A few high-level bullet points if you want to quickly skim through: 1. Fine-tuning is preferred when you need a deep alignment with domain-specific vocabulary and conventions 2. RAG is likely the better option if your application heavily relies on external data and data is dynamic 3. RAG systems provide better mechanisms to minimize hallucinations 4. RAG is preferred if transparency and interpretability are priorities 5. When cost considerations are paramount, training and deploying smaller models can yield substantial savings, particularly at scale—fine-tuning is a better choice. Thanks, Pratik Bhavsar for this wonderful article!

Very interesting read. Definitely agree that software development in general will be so much faster with AI. I'm already using GPT quite often when coding, especially for 'standard code' (e.g. edit-distance function) or areas where I'm not an expert. The 'wow' moment for me was when GPT successfully rewrote my O(n^2) function to O(n). https://lnkd.in/eePm-kGG

It makes sense that a talk-to-my-data or finetuning approach, as opposed to prompting would work. In finetuning you apply better data, more data, relevant data, together with initial prompting (submitted through the code), to give the model a more complete understanding of the task. But what if data makes the model better, yes, but not adaptively intelligent in the way really good prompting in the moment does through the chat interface? In skillful prompting not only are you giving it good data examples, too, but also context (act like, be like, refrain from), *and* - this is likely the key - you’re responding back to its responses again and again. You’re in dialogue with it and your creativity is making its creativity better and vice versa. So is prompting defined as “data plus context plus dialogue” more effective than finetuning? I think the answer is yes. Not that finetuning doesn’t have its value, even a lot of value when done right, but I don’t think it entirely gets at what’s happening in skillful prompting. How could it if the biggest value is in the creative back and forth with the model? Soon, finetuning and prompting will probably merge and be one process and interface that supplies great data and initial prompting to the model, followed directly by in-the-moment prompting by a person or machine. https://lnkd.in/gCXyRzP7

I've been obsessed with a ridiculous task for some months now: getting LLMs to rhyme in ABBA schema. Turns out, it's not as easy as it seems, even for the most advanced models. In my latest blogpost, I show how I evaluate LLMs on this specific task. I discuss why it is an interesting problem, and share the lessons I learned along the way. Key takeaways: 1. Evaluate LLMs on tasks you care about 2. There's a lot to learn from labeling the data 3. Good instructions can go a long way 4. Use tools that make it easy for you to get started Thanks to Michiel van Rijthoven for introducing me to the task! https://lnkd.in/edeueQsD

I wanted to share something I've been working on for a little while... Introducing AlphaCharter! (alphacharter.xyz) Have you ever felt frustrated by the steep learning curve and hefty price tag of financial analysis tools? What if there was a way you could perform sophisticated financial analysis and create engaging charts without knowing any Python or financial jargon? Welcome to AlphaCharter.xyz, where complex financial data becomes accessible and actionable, all through a virtual financial assistant, AlphaAI. AlphaCharter is an AI Assistant and product I built over the past few months. You might have been jealous of your coworker who makes cool charts with Python to visualize their data and wished you had those powers. Even if you're an all-star Excel user you could have a number of issues sourcing data for your own analysis and visualizations. Now you don't have to know how to do either of those boring things. All you need is an idea and the ability to describe what you want in natural language! Take a look at my blog post (https://lnkd.in/g58__5qA) to see more, or email me at max@alphacharter.xyz for more questions. I'm happy to give anyone free access in exchange for feedback!

Towards Data Science August edition lists the popular articles published in August about data science and AI (including the one I wrote about ChatGPT Code Interpreter). #datascience #artificialintelligence #chatgpt https://lnkd.in/dFrGu4Xe

🚀 Help Shape My Practical Fine-Tuning Blog Series! 🚀 Dear LinkedIn friends, I'm excited to announce the launch of my new blog series, "Practical Fine-Tuning," with the first post titled "How to Determine If Your Task Will Benefit from Fine-Tuning an LLM." This series aims to provide actionable insights and guidance on fine-tuning LLMs. I'd love your feedback on this first post, as well as your input on which topic to tackle next! Please vote or comment with your preference: 1. Continual pre-training or supervised fine-tuning? 2. LoRA fine-tuning or full-parameter fine-tuning? 3. How to predict that my data is useful for fine-tuning? 4. How much data do I need to fine-tune a model? 5. Which base model to fine-tune? 6. Should I fine-tune a larger model or a smaller one? 7. Is hyperparameter tuning necessary? 8. How to avoid forgetting? #finetuning

"Migrating prompts across models is a pain in the ass" - a choice quote and one of many practical insights shared in the second part of an excellent series on building production applications with large language models (LLMs). Written by Eugene Yan, Bryan Bischof, Charles Frye, Hamel Husain, Jason Liu and Shreya Shankar, this article dives deep into the operational aspects of developing LLM applications, covering critical topics such as data management, working with models, product design, and team building. https://lnkd.in/ert5qfjr