So as an example… I just created a TL;DR lens and ran my original post through it. It’s possible that you keep the original content intact but use lenses to direct the summarization of the content. That’s another way to think about this. Just re-time summarization that is personalized for each user.
First here’s the lens:
### TL;DR
[Re-write the original post for clarity. Retain all of the original ideas but add analogies if needed.]
### Extensions
[Propose extensions to the original posters ideas]
Now here’s my post run through the lens:
TL;DR
Imagine language as a tool that compresses complex ideas into words, much like zipping a file to send it over the internet. The effectiveness of this tool depends on how much shared knowledge the sender and receiver have. For instance, conversations about programming or medical topics can be like high-speed data transfers if both parties understand the jargon. The author introduces a concept called “lenses,” which are techniques to guide language models in tailoring their responses. These lenses could adjust the way information is presented to align with an individual’s existing knowledge, potentially making communication much clearer and faster. The author is enthusiastic about this concept and is eager to hear others’ thoughts.
Extensions
One possible extension of this idea is to develop a library of lenses tailored to different fields or levels of expertise, allowing users to select the most appropriate one for their needs. Another extension could involve creating adaptive lenses that learn and evolve based on user interactions, further enhancing personalized communication. Additionally, integrating these lenses into educational tools could revolutionize learning by providing students with explanations that match their current understanding, thereby improving comprehension and retention.
Here’s a thread explorer lens I created and ran this thread through. First the lens definition:
### Post
[Re-write the original post for clarity. Retain all of the original ideas but add analogies if needed.]
### Replies
- [Create a tl;dr of each reply]
### Analysis
[Create a detailed analysis of the post and replies]
### Extensions
[Propose extensions to the ideas in the thread]
This thread passed through the lens:
Post
Imagine using GPT as a universal translator for concepts, where language acts like a zip file, compressing complex ideas for easy transmission. Language models (LLMs) can decode these compressed ideas, making intricate topics accessible to everyone. I’ve developed a technique called “lenses” that tailors responses based on what you already know, potentially speeding up and enhancing understanding. I’m excited about this concept and would love to hear your thoughts.
Replies
Reply 1: Suggests creating personalized knowledge files but raises concerns about scalability and risks of over-customization.
tl;dr: Personalized knowledge is promising but could face scalability issues.
Reply 2: References Neal Stephenson’s “The Diamond Age” and its similar concept of personalized educational tools.
tl;dr: Similar ideas exist in literature, highlighting potential for personalized education.
Reply 3: Emphasizes the need for filtering and re-aggregating information rather than just reprojecting it.
tl;dr: Focus on refining and reorganizing information, not just presenting it differently.
Reply 4: Warns about the risk of becoming trapped in one’s own knowledge bubble, which could impede growth.
tl;dr: Beware of knowledge bubbles that limit development.
Reply 5: Discusses the interplay between language, thought, and culture, cautioning against losing meaning through excessive adaptation.
tl;dr: Excessive adaptation might dilute meaning, affecting cultural context.
Analysis
The post introduces an innovative approach to using GPT as a tool for translating complex concepts into understandable language, emphasizing the role of language as a compression mechanism. The “lenses” technique is a promising method for customizing information delivery based on individual knowledge, potentially revolutionizing how we learn and communicate. The replies offer a range of perspectives, from enthusiasm about personalized learning to concerns about scalability, cultural implications, and the risk of knowledge isolation. The discussion highlights the balance needed between personalization and maintaining a broad, meaningful understanding of information.
Extensions
Develop a comprehensive library of lenses tailored to various fields and expertise levels, allowing users to access information in a way that aligns with their background.
Create adaptive lenses that evolve through user interactions, continuously refining their ability to match explanations with the user’s current understanding.
Integrate these lenses into educational platforms to enhance learning experiences, providing students with explanations that align with their existing knowledge and learning pace.
Are you suggesting that language, by virtue of conveying a concept, induces flaws in reasoning or that imprecise language induces flaws in reasoning? Alternatively, the fundamental nature of language, whether spoken or internalized, is itself inescapably flawed as a reasoning tool.
Just curious.
-K
Your idea has merits, and I don’t want to stifle your creative spirit, but your concept does carry with it some major flaws.
AI Ethics: There are considerable ethical implications in something like this. If an AI system knows me so intimately to communicate with my “world knowledge”, then that means that if someone hacked into the system, they would know such intimate knowledge. Or if the company sells my information, then others could use it. Advertising thrives on this kind of information and knowing and understanding people they’d pay top dollar for. Many websites that claim they don’t sell customer information have been caught doing so. It would need very strict guardrails that is not something we’ve been able to enforce today.
You can have insight in the various things I know, but that won’t gurantee that you can communicate in my way of understanding. To do this, it would have to know a vast variety of concepts of me, from my world knowledge, but also my learning style, and these are but two concepts needed. How I learn and what I know are two separate concepts, so thus the AI knowing my “world knowledge” only is ineffectual.
Diminished Critical Thinking: Part of how we as humans adapt and grow is by trying to understand concepts that are foreign to us, especially if it uses words we are unfamiliar with. Moreover, part of idea generation that is the most beneficial to the process is being forced to explain concepts in new ways to try to communicate with others. Your concept would strip humans of this experience. They won’t adapt to new ways people try to communicate ideas to them because they have a system that can trasnlate how they understand, thus they won’t ask questions such as, “can you give me an example of this”, which the creator of the thought won’t get this answer and not be required to think of their thoughts in new ways.
Reduction in Cognitive Diversity: Since there is nothing that requires me to think beyond my means most comfortable to me, then I become reliant on a system to basically do the thinking for me. If it only translates how I understand things, then to understand anything in this world, I am forever limited and dependent on the AI.
Lost in Translation: When you translate something from one language to the next, not all ideas are so easily adapted. And when you translate something that is highly abstract into something simplistic, there is a lot you are removing to achieve simplicity. The idea of taking a thought someone has and translating it into an explanation that is specific to the user, opens the opportunity for content lost in the process.
Not Knowing What You Don’t Know: Now you talk about multiple filters to understand concepts in different ways to comprehend a thought, which could resolve #5. The challenge with that, is what if the user who receives something that translates into their understanding believe they understand it, and in truth they don’t. They don’t know there is a component missing likely because of the translation process. Not knowing there is information missing, they won’t necessarily know to use other lenses to try to better grasp concepts if they think they understand it just fine. It has the potential to create problems, or rather, not solve problems that is experienced in everyday communication.
As humans, for us to grow and thrive, we need to be challenged. That challenge has to encourage us to do things that make us uncomfortable. Your Universal Translator idea again has merits, but unless you could resolve some of these problems, I see it would do more to dissolve what it means to be human and have everyone reliant on a machine to do the thinking for them, reliant on the machine deciding what is important and what is not.
Super interesting. Any thoughts on how you apply or implement? I am primarily using A.I. Assistants for projects I am building out. Could this be something that almost lives in a “personal” profile
Lenses are the approach I’m using @curt.kennedy calls them projections but basically the same concept. The idea is to provide a structure as part of the prompt that defines the shape of the response you want back.
There’s no real rocket science here as you’re already doing this when you ask the model to return a response that’s shaped like JSON. The observation with lenses/projections is that you can control more than just the shape of the output. You can control the concepts that get returned.
Understood. I guess I was thinking more at scale. Essentially how you could deploy this in a use case where it’s not assumed the user would add via prompt manually, but instead perhaps drawn dynamically. Kind of like “personalized settings” which can be pulled in to create the same results.
Most of the systems I am building are minimizing the dependency of the end user inputting a majority of the prompt.
Automatically inferring the right lens is tricky. OpenAI is trying to do this with their chat gpt memories and they don’t work (at least not in my opinion.)
I would pre-build a bunch of lenses and then you could have the user install them like plugins. You could then have the LLM select the best lens to use based on the query. So let the user either manually pick the lens to apply from a list or when in “auto” mode let the LLM choose.
To have the LLM choose, include a use statement for each lens as part of the plugin definition. At query time send the LLM the query and the list of lenses (name+lens) and let it pick. Always include a default lens the LLM can select as a fallback
I don’t have the response handy but I had asked GPT to identify a bunch of personas for potential lens types and it came back with 15 - 20. I then gave it an example lens definition and asked it to define a new lens for one of the personas and it did a good job.
Another approach you can take for automating the assignment of lenses is to do what all of the major news sites do to personalize the home page they show a user. I’ll use MSN as an example since I know the basics of how that site works.
MSN clusters users into one of 15 personas based on the stories they click on. Each persona gets a hand crafted list of stories mixed with location specific stories that are aligned with the persona.
You could do something similar based on the user’s queries. Assign a category to each query the user performs and slowly move them into buckets that put them near other users with similar likes. You can then tune a lens for each persona
Reminds me prompts like “explain relativity theory to my 6 yo son”, but more personalized. Many years ago I liked idea of “teacher world cup”, where the best teacher will be able to explain complex topics even to the children. but if get it right, it is not about explaining new things, but rather interpersonal communication “bandwidth” increase, which is great idea! How to learn user “language”-from all past chats? Also may be it can lead to lock people in his wor(l)d(s) even more, same as social networks keeps people in their social bubbles.
Nice discussion. The concept is about a hundred years old, I studied it a bit when was on faculty of linguistics in Belarus.
We called it “message framing” at the time, meaning the messengers need to “align terms” and delimit the subjects perimeter to boost communication effectiveness.
Please mind the purpose of language does not limit to message transmission (while it is it’s primary role), so for the sake of simplicity I’ll consider only this aspect of the language.
Why this thing boosts effectiveness:
implied context removes the necessity of detailed explanations
clear and pre-defined terms (often in the implied context) reduce ambiguity in messages
common associations (or higher percentage of their match to be more precise) help to expand the implied context and get more control over the non-verbal part of the message (through emotional and visual associations evoked by the received message)
The above helps a lot in social communication, knowledge transmission and …SALES!
Sales are particularly affected because of the 3rd point above.
Practical use: tune brand messaging to potential customers and better conversions.
Common name today: addressing the target audience / rewriting for target audience.
Parts to consider in the prompts/fine-tuning:
common associations (describe associated situations and emotions in the input targeting, and ask the model to match them to emotionally similar situations in the output targeting)
sometimes the logic used by one audience is different from the logic of another, mostly because the underlying motivation. So the idea flow may need adjustment because of that.
I’ve been playing with this thing for a while, kind of hard to approach properly, so currently, I simply include a step or two in my workflows to adjust for audience targeting and message conversions.
Got me thinking about how typical “explain evolution to a 4 year old” statements work (or don’t), and how these shortcuts actually are not the way we humans create new knowledge. Hence, I think as a general/universal translator, your idea may not work, and the bottleneck may be the excessive length of the “translations”. Let me explain: since the gap in knowledge between an actual recipient of some information, and the knowledge the issuer presupposed of its hypothetical recipient/audience, can only be bridged with more bridging knowledge (and btw, not only about the information itself (inner message) but about why it matters (outer message)), the farther away a recipient is from the originally intended recipient of a message, the longer the “translation” will be, reaching impractical lengths at some point.
However, it your idea gives me good ideas a developer could use in products and apps: create a few different key “user personas” of your product according to some major relevant aspect and tailor the content to them (e.g. videogame tutorial content according to self declared gaming experience of the player)
We have been using around 5 personas surrounding sales personas for a few years now (ever since the DaVinci days).
One surprising one that is used frequently is the “sarcastic” persona. You have to use it at the right time, but when used properly, it’s extremely effective
Oh, beside sounding like a hidden schizophrenia pool tests (kidding, did it myself), that sounds really interesting. Can you elaborate how you guys used those personas in sales? That is definitely a must know for me.
More concretely, this is how I used Selfet to translate an chapter in a story to appeal to Terrance Tao (a mathematician) through AI.
Specifically in context of below, the “bridging knowledge” is coming from non-tuned LLMs; because Terrance is a famous personality.
TRANSFORMED VERSION
Consider the topological space of Shantivana, a forest that remains untouched by the flow of time, much like the fixed points of a dynamical system. The forest’s canopy forms a fractal-like structure, with its branches intertwining to create a self-similar pattern that repeats at multiple scales. This intricate web of connections between the trees can be seen as a geometric manifestation of the village’s collective memory, where stories and legends are woven together like threads in a tapestry.
As the morning light casts its rays upon the forest, the villagers begin to emerge, each carrying their own unique storyline, much like the curves of a Riemann surface. Arjun, the young charioteer, moves through the village with a quiet confidence, his presence like the fixed point theorem in a mathematical equation. However, he is drawn to the edge of the woods, where the whispers of his past linger, much like the lingering tails of a Gaussian distribution.
In this forest, the whispers of old are not just the wind rustling through the leaves, but the echoes of mathematical principles, guiding him towards the truth. The sacred tree, with its gnarled and twisted branches, serves as a metaphor for the intricate networks of mathematical relationships that weave together to form the fabric of reality.
In the heart of the modern world, where concrete jungles thrummed with life, there lay a hidden gem, wrapped in the delicate embrace of nature. Shantivana, a forest untouched by time, whispered tales of legends long forgotten among the rustling leaves. The intertwining branches overhead formed a verdant cathedral, revealing patches of cerulean sky while casting soft shadows upon the forest floor. Here, serenity resonated, yet a subtle undercurrent of mystery lingered, like the faint echo of a beckoning call.
As the first rays of dawn painted the sky in hues of gold and rose, villagers began their daily routines, weaving in and out of existence like threads in a tapestry, each carrying stories, memories, and ancient fables. Among them was Arjun, a young charioteer, whose humble presence masked the legacy that haunted him. He moved through the village with a quiet confidence, casting a glance toward the edge of the woods where he felt a magnetic pull, a whisper of his past clinging to him like an unwelcome shadow.
“Morning, Arjun!” called an elderly villager, his voice warm with familiarity. “Have you heard the legends? They say the spirits of old still wander amid the trees, guarding Shantivana’s secrets.”
Arjun smiled, nodding, but inside he cringed at the tales that seemed to mock him, reminding him of his once-celebrated prowess on the battlefield. “Just the wind, old friend,” he replied, choosing not to reveal the tremor that coursed through him at the mention of spirits. “Just the wind.”
As he guided his horses past the last house on the outskirts, the forest ensconced him like an old lover wrapping in the embrace of familiarity — whispers of nostalgia mingled with the rustle of leaves, and the soft murmur of the stream nearby filled the air with life. Swallowing hard, Arjun could not shake the feeling that the woods held a mirror to his soul. Here, his past was not a distant reverie; it was woven into the fabric of the trees, the very essence of Shantivana breathing life into memories that had been long buried.
Arriving at a small clearing, Arjun tethered his horses and sank into thoughtful contemplation. As he gazed up at a sacred tree, old and gnarled, a glint caught his eye. Something lay nestled among the roots. Curiosity ignited, he knelt down to retrieve it — an arrow, pristine in its sharpness, yet strangely out of place, as if a time traveler had left it behind.
“What brings you here?” a voice interrupted his reverie.
Startled, Arjun glanced up to meet the piercing gaze of a figure emerging from the shadows, his presence resonating with an indelible chill.
ORIGINAL VERSION
In the heart of the modern world, where concrete jungles thrummed with life, there lay a hidden gem, wrapped in the delicate embrace of nature. Shantivana, a forest untouched by time, whispered tales of legends long forgotten among the rustling leaves. The intertwining branches overhead formed a verdant cathedral, revealing patches of cerulean sky while casting soft shadows upon the forest floor. Here, serenity resonated, yet a subtle undercurrent of mystery lingered, like the faint echo of a beckoning call.
As the first rays of dawn painted the sky in hues of gold and rose, villagers began their daily routines, weaving in and out of existence like threads in a tapestry, each carrying stories, memories, and ancient fables. Among them was Arjun, a young charioteer, whose humble presence masked the legacy that haunted him. He moved through the village with a quiet confidence, casting a glance toward the edge of the woods where he felt a magnetic pull, a whisper of his past clinging to him like an unwelcome shadow.
“Morning, Arjun!” called an elderly villager, his voice warm with familiarity. “Have you heard the legends? They say the spirits of old still wander amid the trees, guarding Shantivana’s secrets.”
Arjun smiled, nodding, but inside he cringed at the tales that seemed to mock him, reminding him of his once-celebrated prowess on the battlefield. “Just the wind, old friend,” he replied, choosing not to reveal the tremor that coursed through him at the mention of spirits. “Just the wind.”
As he guided his horses past the last house on the outskirts, the forest ensconced him like an old lover wrapping in the embrace of familiarity — whispers of nostalgia mingled with the rustle of leaves, and the soft murmur of the stream nearby filled the air with life. Swallowing hard, Arjun could not shake the feeling that the woods held a mirror to his soul. Here, his past was not a distant reverie; it was woven into the fabric of the trees, the very essence of Shantivana breathing life into memories that had been long buried.
Arriving at a small clearing, Arjun tethered his horses and sank into thoughtful contemplation. As he gazed up at a sacred tree, old and gnarled, a glint caught his eye. Something lay nestled among the roots. Curiosity ignited, he knelt down to retrieve it — an arrow, pristine in its sharpness, yet strangely out of place, as if a time traveler had left it behind.
“What brings you here?” a voice interrupted his reverie.
Startled, Arjun glanced up to meet the piercing gaze of a figure emerging from the shadows, his presence resonating with an indelible chill.
Wow, interesting post! It really reminds me of some concept I was working on some time ago. I envisioned something similar with the help of a diagram. That’s a very cool idea-that models can serve as universal translators of concepts, with the help of better compression and decompression of information, depending on a person’s prior knowledge. On my diagram, I also wanted to see how “context” is really the linchpin in this translation process and how humans and machines could better interact. I feel that your work with “lenses” to frame communication is a potent leap forward. It would absolutely democratize languages and change profoundly the way people shared and perceived ideas. I just love seeing that we are thinking along the same lines!
The examples you give of a TL;DR lens and an explorer lens don’t appear to be what you first described: a modification of LLM output addressed to a particular audience.
Wouldn’t your lenses always have something like “Keep in mind that your interlocutor is a very experienced Typescript developer who knows a thing or two about mixing cocktails, but not so much about virology.” ?
Are those audience-specific prompts something you did not include?
