Top 5 Insights from my Deep-Dive into Large Language Models
An exploration journey into the world of Large Language Models (LLMs)
I have embarked on an exploration journey into the world of Large Language Models (LLMs) and I wanted to share my key learnings with you. The AI landscape is evolving at a breakneck pace making it hard to keep up !
🎯 The Art of Contextual Prompting 🎯 - LLMs, especially ones trained on 100+ billion parameters, are often capable of answering based on simple prompt (such as how should I analyze a startup pitch deck) however they provide better answers if they don’t have to guess what you might be looking for. Coming from world of google search, this is a big change. If you tell them within the question who you are, who are you analyzing/searching, who is your target audience or what are you intended outcomes and how should it go about analyzing etc. its results would be far superior.
- Example: I am a physician Enterpreneur interested in healthcare innovation and would like to analyze pitch decks of healthcare startups with innovative solutions. I would like to screen the pitch decks to find companies with the most chance of commercial success to invest in or advise. What questions should I ask and what metrics I should be looking for.
- A well-structured prompt, including your identity, goals, and target audience, can make a significant difference in the quality of responses.
- (If you are not familiar with training parameters and size of model then look at my previous posts but essentially bigger the better)
🎩 ChatGPT Customization: The Magic Hat 🎩 - ChatGPT has introduced customization, making it far easier to add contextual prompts in all your chats. It allows setting up a default context (about you and your preferences) for all future queries, thus tailoring answers more accurately to your needs. (Works in paid version, 4.0, turn on as beta feature in settings and then add details).
- It allows you 2 customization. 1 - who you are 2 - how should it answer you. But lot of space to provide information about both questions.
💼 LLM Landscape: Navigating through the Maze 💼 - Its very confusing how many LLMs are out there and how to compare. There are 3 main general purpose LLMs with an interface and lot of ‘AI products’ that basically call on these 3 through APIs. ChatGPT, Claude and Google Bard (PaLM). Though there are some small publicly available LLMs but most are not deployed for easy use (and have to be built/deployed on a server). There is Bing AI (likely based on ChatGPT so I am not counting) and there is Llama 2 (from Meta but no public facing dashboard to query it and you need to have someone running it in servers and products should start coming out soon using it).
- When it comes to reasoning tasks, I find ChatGPT 4.0 to be the go-to tool, surpassing its predecessor (3.5 version) in training and reasoning, making the older version practically obsolete. One downside, however, is the lack of internet access, a feature briefly trialed as a Bing functionality within ChatGPT, but later replaced with code interpreter (See my post on it).
- For internet search-based tasks such as finding resources for startup pitch deck analysis, profiling startups, or researching competitors, I turn to Claude 2. Although its interface isn't as intuitive as ChatGPT's, its performance in finding and summarizing online content is impressive - and it's still free (for now).
- My trials with other tools such as Bing Chat, Google Bard, and those based on smaller LLMs haven't yielded any serious contenders to the Big 3 yet. The caveat: I haven't tested anything on LLaMa 2 yet, a smaller parameter LLM that might need additional, purpose-specific training to match up to ChatGPT.
🧩 ChatGPT Plugins: Your New Superpower 🧩 - I'm genuinely astounded by the added functionalities plugins bring to the table. They are fantastic tools I relish using, with favorites like LinkReader for summarizing anything and 'There’s an AI for that', which recommends AI tools for any task.
- The brilliance of ChatGPT is that it recognizes when to use a plugin and seamlessly integrates it into the conversation. There's a flip side, though. In each chat session, you can only select three plugins. Once chosen, they are locked in for that conversation, and you can't switch them later. To utilize a new plugin, you need to initiate a new chat, thereby losing the context and continuity of your previous discussion.
- Moreover, the plugin store leaves much to be desired in terms of navigation and description. However, I'm optimistic about its evolution, potentially into a more intuitive, App Store-like platform. Despite the hiccups, plugins undoubtedly enhance our interaction with ChatGPT.
- (Available for Paid subscribers, works with GPT 4, turn it on in settings as a beta feature, find plugin store in the drop down menu on a new chat in GPT 4, they are constantly updated so look often or search for one when needed for a task).
- I think these plugins will likely make separate websites for every LLM feature redundant and they integrate in your overall workflow with ChatGPT much better.
🔍 The AI Solution Revolution 🔍 - ‘There’s an AI for that’ is one of my favorite plugin in ChatGPT 4 but literally true. There are 100s if not 1000s of ‘products’ providing a specific solution out there such as summarize YouTube videos, write social media posts, chat with your pdf, build forms for you based on prompt and my recent favorites - develop presentations from concepts with built in image generation and AI text editing (I have tried Tome.app but there are 3-4 more I discovered and likely more) and Writing Research papers. Here are a few commonalities among these AI innovations:
- They are primarily web interfaces that trigger an API (usually ChatGPT) or run on dedicated servers hosting open-source LLM code.
- They've layered specialized features on top of General Purpose LLMs, often resulting in robust tools.
- Some utilize proprietary LLMs or tools not publicly available, even when the core LLM is open source.
- Many offer free or 'freemium' versions, providing swift solutions without a price tag.
- Most provide premium features under paid subscriptions, which are often affordable if used frequently.
So, the next time you're pondering a task or trying to crack a problem, consider searching for an AI solution. There's a good chance you'll discover something helpful. Most of these tools are from startups, and while there isn't a dedicated search engine for them, some newsletters compile a handy list of these products on their blogs, making discovery easier.
🎁 Bonus Tip: Unleashing Power Within 🎁 - Do you know your favorite subscription services might already be AI-integrated? Services like Canva are embracing AI to enhance their offerings, sometimes outpacing startups. Take Canva AI, for instance, it seamlessly integrates a text-to-image generative engine into its existing workflow, making it more powerful. So, before venturing out for a new solution, delve deeper into what you already have. It might surprise you!
In this era of AI-driven innovation, the key is to adapt, learn, and leverage these tools effectively. So, are you ready to dive into the AI realm and transform your work processes? Let's discuss!