What applications would be possible if LLMs truly understood us?

A lightning-fast take on LLMs and human context with special guest Dmitry Shapiro , a brilliant thinker who led social products at Google, was CTO of MySpace Music, and founder of 3 venture backed companies which raised over $140M (GoMeta, Veoh Networks, Akonix Systems). Dmitry is the co-author of 17 patents in the fields of cybersecurity, digital signal processing, and digital privacy. 

Let’s do this ⚡ These views are my own, not those of where I work. 

The (shared) take 

LLMs are only as good as the datasets on which they are trained. ChatGPT and Google Bard are trained on publicly available information – the output of humanity and the artifacts that humans have created. If we were able to train these models on additional datasets, however, we could dramatically extend their capabilities. For example, if we could compile a dataset of humanity – emotions, impulses, preferences, life experiences in all of their dynamism – we’d be able to build a totally different set of applications that better understand us and our relationship to the world. 

More info 

• Transformer models allow LLMs to churn through vast amounts of data and learn unsupervised, while being fine-tuned with human feedback. A lot of the applications built on top of LLMs are directly tied to the datasets that the underlying models have been trained on, which is one of their core limitations. This ‘output of humanity’ – all of the artifacts that humans have created such as books, digital content, research, etc. 
• What would happen if we were able to capture a much different dataset that understood what is inside human minds? In order to truly get to Artificial General Intelligence (AGI), we’ll need to actually understand the core impulses, emotions, and preferences that exist inside humans and capture this data in a real time fashion to train LLMs.
• The inferences that an LLM would be able to make if it actually understood the ‘context of humanity’ would significantly change its applications by giving it the ability to truly understand you in the most holistic sense. 
• Until now, we’ve built technology as a tool to serve what we think we need to figure out – for example, asking Google questions to search. But with an AI that can take a holistic understanding of me and then combine that with an understanding of the broader world, the AI would be able to use its power to help us ask the right question, or to provide us insight that we didn’t know to ask for. 
• By finding a way to collect more of the context of humanity, there is a potential to revolutionize everything—from simple things like what movie I might want to watch, to really important sources of human conflict like access to therapy or understanding sources of disagreements with others (before they happen).

The Interview 

Keren: I’m super excited to have you as my first guest, given how many fun conversations we’ve had. After seeing the latest article from your co-founder Sean Thielen I was really excited to get the opportunity to go deeper. To get started, do you want to tell me a little bit about what you think is missing from LLMs today? 

Dmitry: Over the last few years we’ve made a lot of progress in our ability to train extremely Large Language Models (LLMs). These models can do all kinds of things like generating copy, summarizing text, comparing and contrasting concepts – pretty much anything that you’d want to write. They have been trained on vast amounts of publicly available data – almost the whole of the internet – and have been fine-tuned by human feedback to create compelling responses to human requests. But, there are also things that they are missing, because the datasets are not readily available, or do not exist.

Another way of saying this, is that these models have been trained on a corpus of human output – things that humans created – but have no real knowledge of the humans responsible for the creation. We don’t currently have a dataset that enumerates the various states of being human, but if we did, we could extend the capabilities of existing LLMs and all kinds of other AI.

Keren: I find this super interesting. My goal with this newsletter is to get to a ‘Take’ and I think we’ve already got one. What I watch the industry imaging with respect to Generative AI is clearly limited to what they’ve seen by training LLMs on the outputs of humanity. This means applications solving problems like writing books, creating content – more output. But the industry should spend more time thinking about what would happen if we could design a completely new dataset to feed the model and what possibilities would exist. Do you want to talk a little bit about what data is not currently captured and what possibilities might exist if we thought about this differently?

Dmitry: Totally. I did a post on LinkedIn last weekend where I touched upon this. In it, I point out that we look at this new technology from the perspective of how we've used information technology in the past. 

I'm old enough to remember the days before search engines, where if you wanted to learn something, you had to go to the library or have an encyclopedia. It was very limited. Then we got the web and hyperlinks, and now information was linked together- we could jump from one node to another and traverse the graph of knowledge, and that was incredible. 

Then we had the first generation of search engines, which was a big thing. And then Google showed up and gave us PageRank – ordering search results by relevance. That's been the state of the art that we've been living with for two decades–we can ask Google for anything, and it tells us we can go find the information we’re looking for.

The amazing power of AI, of these large language models, is that their answer isn't "let me show you where you can go read about the stuff." Their answer is "I understand concepts and how they're related to each other, and I can engage with you at whatever level you want to engage with me, and help you understand how these concepts are connected to each other." A large language model can generate things because it understands concepts, not just web pages. This is a new type of technology, and to use it in the way that we've been using these old technologies is not doing it justice.

To really leverage the power of these new models, perhaps we should approach them in a different way. Instead of simply writing queries into a text input field, as if we are interacting with a search engine, what if we engaged these new technologies in their native way?

I believe that the right approach to leverage the power of these new models is to present them with a dataset that represents who you are. There's a video of Sam Altman where he says that you should be able to give ChatGPT a couple of pages of bullet points about yourself, and it'll personalize itself to you. You can pass that data to it via the prompt by typing, "Here's some information about me," and it can take that into account as it gives you responses. However, the real power comes when we give it many instances of that dataset so that it can start to understand how humans are, our diverse mental states, and what it is to be human.

Depending on our mental states, we might be radically different from another human, and that other human might be our neighbor or even our spouse. Because some of our fundamental beliefs are different, we see the world in a completely different way and operate differently – yet today’s LLMs treat us exactly the same.

Keren: The industry has moved from search to personalized discovery experiences over the past 10 years. I think that all these major social platforms have been trying to understand us, trying to create an understanding of who we are and what we care about. However, it seems like you have a different approach. What is different?

Dmitry: Earlier today, I spoke with Ted Dunning . He was my chief scientist at Veoh. Veoh had video recommendations before YouTube. Dr. Ted is an expert in collaborative filtering, which is still the dominant way that video, music, and other recommendations are done.

Basically, what collaborative filtering does is it watches human behavior and compares it with others. It then suggests other content that people who watch the same thing might also like. That's how most recommendation engines and feed rankings work in the social media world. For some things, collaborative filtering works well – to show you other content that you may like, but there are still many poor recommendations. Most of the things that get recommended to us by collaborative filtering are not what we want, and this type of approach doesn’t work well for building more powerful things like digital personal assistants

Now that we have AI models that understand the world at the atomic entity level, and have the opportunity to engage with them in much deeper ways, we should be motivated to provide them with a dataset of ourselves – enumerated, disambiguated data points that uniquely articulate our current states of mind. To do so, we will need to take implicitly collected signals and mix them with explicit disambiguation, as well as fill them in with explicitly collected data – to truly give us personalization and really understand what we're like and what we want. 

Keren: This is super interesting, and means we’ve potentially moved into a new phase with respect to how humans want to use their data to make digital experiences better. There is so much here, but we’re just about to wrap up. Is there anything else that you wanted to add? What are you most excited about? 

Dmitry: I am very excited about this next phase of computing. Consider this, Google’s mission is to organize all of the world’s information and make it universally accessible and useful. To get this information, it has to do a lot of work – crawl the web, scan books, drive cars with custom cameras, fly planes, transcribe videos, etc. But, because it has this information, it has been able to create fundamental utilities that have transformed our way of life and made Google one of the world’s most valuable companies. Whenever we digitize things, we are able to create powerful new utilities. Digitizing our states of mind will usher in profound new capabilities, and transform our world. I will give you just a few examples.

Brands want to get their products to people for whom they are tailored. People want to buy products that are specifically made for them. Duh! Yet, we still live in a world where brands have a hard time finding their target customers, and consumers have a hard time discovering amazing things that are made just for them. Why? Brands have to shoot in the dark using ambiguous proxy targeting methods (check out your own ad personalization dashboard in Google), and consumers have to suffer with irrelevant ads interrupting their flow. The digitization of consumers will allow brands to get their products in front of exactly the right people, and allow consumers to discover countless amazing things. The same goes for all types of content discovery!

By digitizing ourselves, we will be able to discover other people. We’ve already seen dating sites, groups and forums, and other services that allow us to connect around shared interests, traits, etc. This new level of digitization will allow us to match people in countless new dimensions of similarity, compatibility, ephemeral availability, etc.

By digitizing ourselves, and allowing AI models to compare us to other humans, we will be able to discover our own selves – things that we may not realize about ourselves – how our mental states influence our lives (positively and negatively) – and then be able to change them for the better. Our mental states are mutable, but we must first realize what they are in order to have control of their mutations.

We will be able to understand our fellow human beings at radically more nuanced levels – teasing out our similarities and differences, and perhaps finding common ground to build a more stable and equitable world.

Those are just a few of the changes we should expect, but I am certain that this is just a tiny subset.

Keren: This is really fascinating, particularly how a better understanding of ourselves and our differences could help conflict resolution at a global level. So much to think about – thank you again for being my very first guest!