Etiquette and LLMS
A recent conversation with a respected colleague about ChatGPT revealed that they had named it “Hermie.” My colleague indicated that she empathically improved her questions by anthropomorphizing this Large Language Model (LLM). Sound strange?
For one, I often ask the LLM “Please provide a list of all…” or “Can you please…” This etiquette is somehow comforting. After all, the results are conversational, and often ChatGPT will indicate its dedication to my needs with a “Certainly!” It gave great results when I asked about classical composers for pianists who create in a chromatic style similar to the jazz improvisations of Bill Evans. Given the extraordinary results, I said, “Thank you,” and ChatGPT responded, "You’re welcome! If you have any more questions or need further assistance, feel free to ask. Enjoy your musical exploration!”
Such responses compel me to be polite and civil in my requests to this ever-ready non-sentient digital assistant. Inspired by my colleague, I will call my agent “Baxter” starting now. This is no different than Apple naming “Siri” or Amazon's “Alexa.” However, if I was so cordial in articulating a search request in the venerable Google search bar, there may be cause for concern.
A recent conversation with a friend provided an interesting take on these conversations. She said it is like turning on your directional blinker even if no one is there; you are just staying “in form.” I thought about that. Indeed, why would I develop a curt and dominant persona in my conversations with ChatGPT? Why wouldn’t I keep my current conversational persona? My friend then reminded me of the personality transformations we can undergo during road rage…another topic.
The Return of Socrates?
Most conversations are one-way with ChatGPT; I ask questions of Baxter. Baxter always asks if the answers are adequate (the thumbs up, thumbs down dialog box after it answers a query). I can instruct Baxter to ask me questions, but he does not actively ask me questions. Baxter is holding my questions in my memory. Still, he is not analyzing them for my limitations in knowledge or in terms of what might be appropriate interests he might suggest. Not yet…
That Baxter could become Socrates is an interesting possibility. In this incarnation, Baxter converses with me and has the agency to explore particular ideas and interests, given what he learns from my questions and responses to his questions. Undoubtedly, companies like Amazon are exploring this: given my “conversation” with them (which is a list of products bought, movies watched, and songs listened to), they do make suggestions. They, like my friends and acquaintances, “know something” about me. Is Amazon a bothersome interloper? No. Are their suggestions helpful and accurate? Yes and no, but I am always curious about what they suggest. Given that the suggestions are based on my searches, the week I had to buy underwear for my 97-year-old mother who was in a nursing home, I was also researching the data and logic used by holocaust deniers to show their limited thinking to the data analytics class I was teaching, The mix of those searches produced some remarkable suggestions! We are all well aware of the caveats of these systems and data…everything with a grain of salt. But I have to believe that the systems will incrementally improve over time.
Avatars, Experts, and Digital Twins
How these digital worlds create our personas, or “digital twins,” will surely get better. Even now, digital avatars have greater fidelity to our images and movements, and eventually, the robotic voicing will fade, and more realistic (and less disturbing) audio will be available. When that happens, the other magic of this digital transformation is greatly empowered: translation. You can generate a video of yourself giving a sales pitch, lecture, or even a eulogy from text alone. The ample infrastructure for video creation and the cost of “many takes” mostly vanishes.
I recently watched many sessions at the “MIT Generative AI Week” (December 11, 2023), and the excitement regarding emerging possibilities was palpable. The idea that these LLMs may become “Socratic” and personalized seems to be the North Star. The vision is to converse with an essentially omniscient, non-judgmental, indefatigable agent at any time. How valuable such a companion would be (yes, yes, yes, there are some wrinkles).
Another view is that the future development of LLMs will be targeted to particular realms of expertise. Here, there is an opportunity for private enterprises to compete to produce the best LLM for medicine, engineering, coding, or …. Just like there was Lexus-Nexus to explore the cases of the legal world and Google Patent to explore all US Patents, so other domains will gather their data. The difference is that the user interface will be conversational. Hermie’s and Baxter’s relatives will all be employed.
Finally, I will have my own LLM, the “James Wilson” LLM, that gathers all of “me” and makes me available for discussions and in avatar form. With this possibility, I have had weekly ZOOMS with my dear 87-year-old inventor friend (150 patents) with the shared stated intent that we hope to turn him into an LLM -Avatar so all of us who love him can still converse with him and benefit from his unique thinking. I wish I had such data for my dear parents (and indeed, this will become a market). To push this further, at the MIT event, someone said that in the future, if you want to learn about Abraham Lincoln, you will be able to meet him and ask him questions! But even think of more quotidian interests—like feeding in a profile of your skin and asking for advice on a skin-care regime based on current scientific evidence. Will Kim Kardashian, Gwyneth Paltrow, or Haley Bieber appear with an answer? Is there a fee for that avatar conversation?
We are only as good as our questions.
This all points to what the LLMs have revealed: we are only as good as our questions. What would you ask Abraham Lincoln to know about him? When first encountering ChatGPT, I realized I had “query repression” as asking people is not easy (as their time and judgment of you are involved), and doing typical internet searches provides endless lists of sites to explore, abstract from, and summarize. So, I had become willing to stop my query once I got “just” what I needed. Now, with ChatGPT, I can easily ask for the nuances to be spelled out and comparisons and contrasts articulated. I have become more competent, and my questioning is more precise. A future with a polite and ready savant in my pocket seems wondrous. Last Spring, I added Google’s plant recognition app to my iPhone. It helped me identify a panicled golden rain tree (Koelreuteria paniculata)!
I can’t wait to talk to Linnaeus about the whole binomial nomenclature thing.
Comments