Tips for Improving Robot Language Understanding

As Generative AI (GenAI) becomes more common place, a new Human superpower will emerge. There will be those with expert ability at getting quality information from LLMs (large language models), and those without. This post provides simple tips and tricks to help you gain that superpower. TL; DR: To better interact with specific #GenAI tools, bring focused problems, provide sufficient context, engage in interactive and iterative conversations, and utilize spoken audio for a more natural interaction. Couple background notes. I'm an applied linguist by education; historically, a communicator by trade (human-to-human communication); and passionate about responsibly guiding the future of AI at Honeywell. When we announced a pilot program last year to trial use of LLMs in our daily work, I jumped on the opportunity. The potential for increased productivity and creativity was of course a large draw, but the opportunity to explore an area of linguistics I haven't touched in over a decade: human-computer interaction and communication (computational linguistics) was as well. Words are essential elements of effective communication, shaping how messages are perceived, understood, and acted upon. Similar to H2H communication, words we use in conversation with LLMs largely impact the output of the interaction, from both user experience and quality. A drawback is that we often approach an LLM like a search engine, just looking for answers. Instead, we must approach like a conversation partner. This will feel like more work for a human, which is often discouraging. ChatGPT has a reputation of being a "magical" tool or solution. When we find out it's not an easy button but actually requires work and input, we're demotivated. But in reality, the AI tool is pulling your best thinking from you. How to have an effective conversation with AI: 1. Bring a focused problem. Instead of asking, "What recommendations would you make for using ChatGPT?" Start with, "I'm writing a blog post and I'd like to give concrete, tangible suggestions to professionals who haven't had much exposure to ChatGPT." 2. Provide good and enough context. Hot Tip: Ask #ChatGPT to ask you for the context. "I'm writing a LinkedIn post on human-computer interaction. Ask me 3 questions to would help me provide you with sufficient context to assist me with writing this post." 3. Make your conversation interactive and iterative, just as you would with a human. Never accept the first response. (Imagine if we did this in H2H conversation.) 4. Interact via an app versus web. Some web browsers mimic a search box, which influences *how we interact with the tool. Try to use spoken audio. Talk naturally. And try using different models, just as you would speak with different friends for advice. What tips can you share? A special shout out to Stanford Graduate School of Business' Think Fast, Talk Smart podcast for some of the input exchanged here. Sapan Shah Laura Kelleher Tena Mueller Adam O'Neill

Recently, I got my hands on this research paper (check it here: https://lnkd.in/dTUTE3gh) that's flipping the script on how we chat with AI, especially with #LLMs. The researchers have come up with 26 guiding principles for querying and prompting LLMs. They cover areas like how to structure prompts, how specific they should be, user interaction, and handling complex tasks. So, what's the big deal? The impact of these principles is quantifiable. For example, in the ATLAS benchmark that includes multiple questions per principle, specialized prompts increased the quality and accuracy of GPT-4's responses by an average of 57.7% and 36.4%, respectively. Interestingly, these improvements are even more significant with larger models. Transitioning from a smaller model like LLaMA-2-7B to GPT-4 showed performance gains exceeding 20%. This paper is a reminder of how we're only scratching the surface of AI's potential. As we get better at communicating with these models, their ability to help us solve complex problems just keeps growing.

Improving response quality in generative AI is a hot topic right now. At Botco.ai we've been experimenting with a new technique that takes inspiration from the scientific peer review process. In a recent experimental project to enhance the reliability of our language models, we've been delving into innovative ways to improve the quality of their output. Here's a glimpse into the process we tested: 1. A user types in an input/question. 2. This input is directly fed into Botco.ai’s InstaStack (our retrieval augmented generation - RAG - product). 3. A language model (LLM) then processes the output from InstaStack, carefully extracting relevant information from the knowledge base pertaining to the user's question. 4. Then, the LLM crafts a response, drawing from the insights it gathered and the original input. 5. Experimental feature: Another LLM critically reviews the proposed answer, cross-examining it against the user's input and the information to ensure accuracy and coherence. 6. If it detects a low quality output (we’ve tested many thresholds), the process is iteratively refined, modifying the response and reassessing to avoid potential infinite loops. 7. Once the response is verified as accurate, it is then delivered to the user. Overall, this method yielded good results but the user experience did take a hit as the messages took longer to deliver. From a business perspective, we can say that this experiment was successful today in terms of quality control: much higher accuracy in responses and nearly zero hallucinations. Now the challenge is generating this result with less latency. On to the next experiment! I'd love your feedback on what we should try next. CC: Crystal Taggart, MBA Vincent Serpico Ana Tomboulian Jacob Molina #genai #aiexplained #llms

Fellow AI enthusiasts, are you looking to better understand and optimize how you interact with large language models (LLMs) like Anthropic's Claude, GPT, Falcon40b and more? One key bit of advice is developing your prompt engineering skills. Prompt engineering is emerging as a critical discipline for interfacing with LLMs more efficiently. It goes beyond just writing prompts - it encompasses techniques like: Few-shot learning - priming the LLM with just a few examples Chaining - breaking down prompts into logical steps Demonstrations - providing the LLM with demonstrations of the desired behavior Safety prompts - guiding the LLM to avoid harmful responses Knowledge prompts - integrating external knowledge into the prompt With prompt engineering, you can guide LLMs to new capabilities like reasoning and knowledge integration without extra training data. You can make models more safe, robust, and aligned to your needs through careful prompting. I'd love to hear about your experiences applying prompt engineering. What challenges have you encountered? What capabilities have you unlocked in LLMs through prompting? Let's collaborate to further push the boundaries of what's possible. Drop your insights below!