Eric Olson - Deriving Answers Using Word Embeddings in NLP

The conversation this week is with Eric Olson. Eric is the co-founder and CEO at Consensus a search engine that uses AI to surface expert answers. Prior to Consensus, he worked at Draft Kings analytics, where he built cool models and created new metrics to better understand their users. In 2017, he graduated with a master's of science in predictive analytics from Northwestern University and was a three year starter for the Northwestern University football team and a four time academic, All Big 10 honoree.

If you are interested in learning about how AI is being applied across multiple industries, be sure to join us at a future AppliedAI Monthly meetup and help support us so we can make future Emerging Technologies North non-profit events!

• Emerging Technologies North
• AppliedAI Meetup

Resources and Topics Mentioned in this Episode

• Follow ConsensusNLP on Twitter
• Google Scholar
• PubMed
• Allen Institute for AI
• ChatGPT
• Natural language processing
• BERT (language model)
• Attention is All you Need

Enjoy!

Your host,
Justin Grammens

Transcript

Copy Transcript Copied

Justin Grammens  0:00  

Greetings Applied AI Podcast listeners. This is Justin Grammens, your host of the Conversations on Applied AI Podcast. Just dropping in to let you know about a very special event we have coming up on Friday, May 12. It's the spring 2023 applied AI conference. You can learn more by going to applied AI conf.com. This full-day in-person conference is the only and largest artificial intelligence conference held in the upper Midwest. It will be in Minneapolis, Minnesota on May 12. We will have more than 20 speakers with two tracks covering everything from Ai, business applications, ChatGPT, computer vision, machine learning and so much more for being a listener to this podcast, use the promo code podcast when purchasing your ticket for a 50% discount. So here's the details go to Applied AI Conf dot com. That's applied Applied AI Conf dot com, to see the full schedule and register for the only and largest artificial intelligence conference in the Upper Midwest on May 12. And don't forget to use a promo code of podcast when checking out to receive a 50% discount. We look forward to seeing you there. And thank you so much for listening. And now on with this episode.

Eric Olson  1:07  

That's kinda why NLP has exploded is because of these embeddings, which basically give numerical representations of the words. That's how these models are able to learn so much you're able to complete such complex tasks is because they learn from numerical representations of the words. From a super high level, it feels like it's very if you just looked at the surface, it feels like But you just said numbers were swerves, it's like the complete opposite. It's really not at all. And if you're training a Simple NLP model, it is literally like training a machine learning model just with words, not numbers. And if you really boil down the brass tacks of what's going on under the hood, it is more basic data science and because of these, these embeddings basically make a words and numerical representations and then use traditional data science to make predictions on things.

AI Announcer  2:01  

Welcome to the conversations on applied AI podcast where Justin grumman's and the team at emerging technologies North talk with experts in the fields of artificial intelligence and deep learning. In each episode, we cut through the hype and dive into how these technologies are being applied to real world problems today. We hope that you find this episode educational and applicable to your industry and connect with us to learn more about our organization at applied ai.mn. Enjoy.

Justin Grammens  2:32  

Welcome everyone to the conversations and applied AI Podcast. Today we're speaking with Eric Olson. Eric is the co founder and CEO at consensus a search engine that uses AI to surface expert answers. Prior to consensus, he worked at Draft Kings analytics, where he built cool models and created new metrics to better understand their users. In 2017, he graduated with a master's of science in predictive analytics from Northwestern University was a three year starter for the Northwestern University football team and a four time academic, All Big 10 honoree, quite an accomplishment. Eric, thank you so much for being on the program today.

Eric Olson  3:06  

Thanks for having me.

Justin Grammens  3:07  

So I usually start out with my guests talking a little bit about, you know, how they got into the field and what their background was, and all that sort of like technical stuff, but I'm actually really curious to start off maybe sharing more about your football experience, what position did you play and? And how do you think maybe that might have shaped to where you got to where you are today.

Eric Olson  3:24  

So my whole football journey kind of kind of came about in a roundabout way. I was a basketball player my whole life. And I actually transferred to a private school in high school to try to get recruited for basketball, and was convinced that it would help my chances of getting into the school if I saw that it also tried football. And as it turns out, people are about six foot six and not, you know, crazy, crazy athlete or anything and somebody with that body type and athleticism makes for a really, really great offensive tackle, but not great Collegiate Center for basketball. So I ended up getting a bunch of scholarship offers for football, chose to go to Northwestern wanted to find a place to married academics and athletics. I played offensive line. I was our starting right tackle for three years. And I'd say one, I come from a family of a bunch of academics. So I've always been interested in science in the world of science. But in my fifth year, I started to dive into data science. I got my masters in predictive analytics. And that's kind of how I started down this road that eventually, census. But I'd say as far as you know, there's a lot of lessons to learn from playing high level football, high level athletics, I had an entrepreneurial journey. And I'd say above all, what football taught me that kind of led me to take the leap into entrepreneurship is realizing that imposter syndrome are just fake thoughts in your own head. And football really faces you to confront those and you realize you know, you're you're at a stage of something you've always dreamed of and playing in front of hundreds of 1000s of people, but those same thoughts of Do I belong here are still in your head. And when you kind of have that moment you realize As everyone thinks us, it really teaches you to push past that and realize that it's just the subconscious part of your mind talking to you. And those are that is something that happens a lot. When you start a company. What am I doing? And should I be doing this? Am I the person to do this? And oftentimes, it's a very unhelpful thoughts and high level athletics definitely teach you to put pastels.

Justin Grammens  5:19  

Nice, nice. Well, keeping with the metaphor, I mean, at least in athletics, you oftentimes have a coach or a trainer or somebody with you, that this are off the cuff for like, you know, as you could when you're getting into business, and you started your business. Did you have anybody like that?

Eric Olson  5:32  

Oh, certainly, yeah, I'd say two, two main buckets, one, former bosses of mine, before I worked for three, four years before I started the company. And you know, a few of my former bosses are now investors in my company and with them regularly and get advice from them. That's more on like the managerial side and, you know, balancing employees and emotional stuff. And then there's another bucket of other founders go into a school like Northwestern I'm certainly lucky to have been surrounded by a good entrepreneurial network of people that I was peers with in college who have also started their own companies before. And we have leaned extremely heavily on those people on navigating this crazy world. And yeah, I think you're spot on. My co founder also is a Northwestern ballplayer. I really mean it. When I started, I think one of our biggest strengths of the company is taking advice from those types of people and taking it seriously and seeking out advice and coaching. And I think our athletic background definitely has informed that to an extent, we feel like we're coachable. Like we can take hard advice and not get too sensitive and make better decisions from that. And I think it's been a real strength of ours, I think you're totally right to call out that connection.

Justin Grammens  6:38  

Excellent. You mentioned growing up with a family of academics. I mean, my my dad was a physician, my mom was a school teacher, but they did start their own business. They started a vending business, actually for stamp machines. And so I remember being young and going around with him to put new stamps on the stamp machine and getting the quarters out what people bought. And this was back in the days when there wasn't an email, right? So everyone was sending stuff and everyone needed stamps. So they'd go down to the local store and sort of like buy stamps. And you know, I never really thought that I would become an entrepreneur, either. You mentioned your parents having academic background, did they? Did you see anything? Do you see any entrepreneurial stuff as you're growing up?

Eric Olson  7:13  

Honestly, not too much. From being totally honest, I really made it when I started my my family has a pretty deep academic background. My grandfather is a lifelong professor, my dad's dad, my parents, you met at an MIT lab after college. My dad was a computer scientist doing research at MIT. And my mom was a cartographer, so not deep into academia. But they were both working at a research lab in the Boston area. And my sister is a lifelong teacher, she's done nothing but teacher, her whole career, she's taught all over the world as an advanced degree in education. And honestly, what decided what really pushed me to take any entrepreneurial leap was not familial. backgrounds. It was the why now story, I felt like emerge from this idea that I had, as a non expert, as somebody who, you know, is kind of a a jock, who wants to be a scientist deep down because of those connections. I, I yearned for a product that made it easy to get good information quickly, and get information from academic sources. And I knew that product didn't exist. And it was really in the middle of COVID, two things happened. One, it was never more clear that the world really wanted a product like that, that there was a domain, getting good, verified information quickly. But then also learning about how far along aI had come specifically in language modeling technology, and realizing this idea that I had six, seven years ago, holy crap, and it might actually be able to be solved, and it might actually be able to be solved as of like, a year ago, and he's had the conversation with my co founder, if we don't do this, somebody else is going to and felt like that was such a, I felt this pull from that. Why now story, more so than from having a deep history of my family of entrepreneurship.

Justin Grammens  8:55  

Gotcha. So sort of, yeah, I get it. In some ways, sort of scratching your own itch. I guess you're seeing a need in the market is something you would use on it on a day to day basis are correct. Yeah, so I was just gonna say, maybe rewind me back. You're sitting there at DraftKings or I think, right? Is that where you left to then start this business? You see this opportunity? Tell me a little bit about the company, I guess. What does it do? And how was it when you finally pulled the trigger?

Eric Olson  9:17  

The company is called Consensus. And we are a new search engine that uses artificial intelligence to find answers in peer reviewed papers. So it allows you to type in a plain English question. Does fish oil actually help cardiovascular condition? And then what we'll the engine will do is find papers that have worked to answer that question and pull out the claims being made from those papers. So effectively, it's a way to have a question and instantly see what the landscape of evidence in research is about that question. Are COVID vaccines effective? Does zinc actually help with depression? Does the death penalty actually reduce crime? All questions that have been empirically studied, you can ask and we'll find the conclusion hasn't been made and researched. And yeah, I think you said it exactly right. You know, one of the one of our big advantages is we're really building a product, service ourselves. And we're our first users. I use the product every week, not in a professional, obviously, I'm doing tests and whatnot. But I use it to look up my own questions every single week. And it came from a frustration of knowing many of the things that we debate with our friends with our family, or just personally, that there is there are people who have dedicated their lives to studying those questions and parents. But that data is completely inaccessible, especially for a layperson. And you basically have two options to quickly search for try to search for that type of thing. You type it in Google, it's great for many things, but it's not designed to answer this question. You're getting click baby headlines to blogs that are SEO, and get a bunch of advertisements to order you could pop open a Google Scholar or PubMed, which you know has the right source material, the same source material we're using, those products haven't innovated in a decade, all they allow for is keyword hacking, and they deliver you a list of links and all your work is still active. So we wanted to try to build a product that took the best of both worlds that took the intuitive ease of use of a consumer product like Google, but married that with the rigorous source material, like traditional academic source search engines. And yeah, it was love working at DraftKings. It was an amazing place to work did a lot of cool things. They're completely unrelated subject matter wise, but it really was in the middle of a pandemic, where I was no longer commuting. And you know, had some extra time at the end of the day, you know, also sports headstock. So, the DraftKings workload was was a little bit later than normal. So had a little bit of time to explore some other ideas and, you know, potential pursuits. And it was that that led me to start working out along the side and made some progress and realized we were kind of onto something, had some people that want to invest money. And then and when you kind of get, you know, it isn't the sexiest, I dropped everything and took this giant risk, we actually had some investors lined up at the time that I had quit my job wasn't the proverbial giant leap of faith, it was clear, we're kind of on to something and I, I use the luxury of working from home.

Justin Grammens  12:17  

So is this you and your co founder getting this thing started? Or did you have how many people are sort of were involved in the early days,

Eric Olson  12:23  

it was me and my co founder who spent a few months just defining what we were trying to do and learning about how we could potentially do it learning about natural language processing and artificial intelligence. But then we brought on a contracted data scientist to help us build a proof of concept is that was alief green. He's a PhD in computer science. And he is now on the team full time, we brought him on to just build us something to prove ourselves that it could work and eventually prove to investors that it could work to then raise money, build the scalable version of it, he did so so quickly, and so efficiently that we knew we had to be a part of the team. So we asked him to come on full time as our founding engineer. And then after we quit our jobs and raise a little bit of money, we hired two more engineers, and now have a full time team of five.

Justin Grammens  13:03  

Yeah, cool. And you guys are sort of all over United States.

Eric Olson  13:07  

Yeah, me and my co founder here in Boston together and the rest of the team is totally distributed.

Justin Grammens  13:11  

Cool. These are exciting times. It's kind of it's kind of fun that I guess, to be able to sort of see the product, kind of get traction that quickly, especially throughout the investor community, right, because we put together a sort of an MVP, and people came to you, I guess, in some ways, and we You didn't have to sell them or do a very hard sell. It sounds like

Eric Olson  13:28  

you know, I'm understanding that a Betta definitely took a lot of conversations. And we definitely were reaching out to some folks as well. So I'm simplifying that for the sake of the story, it wasn't, you know, I just popped open My Computer and I had people standing at the doorway. Not quite. But yeah, one was a an advantageous fundraising time, I should say.

Justin Grammens  13:48  

Sure, sure. Well, you feel like you're onto something. If it's more of a you know, you can if people are getting pulled into it, rather than having to sort of push him push them into it makes them quickly. Yeah, and you were talking about source material, I guess, you know, I don't really know, is this material completely and freely available? You guys have to pay to get any of it? Or is it able to be accessed by anybody.

Eric Olson  14:08  

So we our data is in our products through a partnership we have with the Allen Institute for artificial intelligence. They typically work with mostly other nonprofit research organizations. They are a nonprofit research organization that goes through some relationship building and conversations, we were able to work out a partnership because we were able to show them how mission aligned we were, and that we were trying to do, you know, had similar objectives that they had a slightly different slant to it. And we're lucky enough to get that partnership and get access to about 200 million papers via now, within that dataset, a large portion of it is fully open access. They have just aggregated it. However, there are some within there that are still behind paywalls we're able to run our analysis and surface those to you. But if you follow the links, the full text of the paper, you still would be subject to those cables. However, the trend is very much in our favor and There was legislation passed this year, that all federally funded research has to be fully open access by 2025, meaning more and more of these documents will be able to be accessed by him. And we hope to be the the layer that helps you access those papers more easily.

Justin Grammens  15:14  

Sure. So it's safe to say you guys are bringing more of a conversational approach to this or using conversational AI to ask these questions.

Eric Olson  15:22  

So very much yes, in the sense that our AI allows for natural language questions. So like, you don't have to just keyword hack like you would if you use Google Scholar up, you can say literally, what is your research question? You can just ask it? Does the death penalty reduced crime? You can ask that question like that, and we'll find you claims made papers about questions. Where I pumped the brakes a little bit I'm using the word conversational AI, is that our results are actually extractive from the papers. So unlike something like ChatGPT we're not generating text out of thin air. It's not our our models answering the question for you. It's our models, looking through a paper and trying to find you the end. So we're actually surfacing you word for word quotes from papers about your question, as opposed to conversationally generating text unanswered. And we ChatGPT is amazing that we want to use some generative tech in the product. But for our purposes, pretty important to understand the source material. A conversational AI is amazing for some use cases, but in what we're trying to do and deliver you research backs, answers and insights, it's pretty important to know the sources available.

Justin Grammens  16:30  

Yeah, for sure. Can you string questions together to can you build on questions? Once you get an answer, then sort of PDS is smart enough to understand that I'm still asking questions along the same way? Along the thread, I guess. Yeah. We do

Eric Olson  16:42  

not have any functionality like that right now. That is definitely something we're working on. And thinking about in the background, though, right now it is not query to query taking in what you've just clicked on, or something from the last last session?

Justin Grammens  16:55  

Who's your core customer? I guess? You know, I was I was just thinking about, I've heard something similar with regards to people have looked through all sorts of patents. And so like, so like lawyers are used, there are like a lot of their stuff to find patent material or other stuff that's been infringed upon. Where do you guys find most of your customers like which area of the industry?

Eric Olson  17:12  

Yeah, so there's really kind of two arms of the spectrum. And it's pretty cool, because they're, in many ways, opposite spectrum, and we're receiving most of our users. So you have this one bucket that is kind of like myself and my co founder, evidence interested non experts, the sub persona within that we see the most is health and fitness optimizers. So the type of people who are looking up information on supplements on exercise routines, on sleep routines, and constantly on the hunt for the best information to optimize all parts of their house, see a lot of people using the product. For that reason, find how I use the product the most when I hear some friends make a claim without some miracle supplement or something represents a way to quickly fact check out and understand that there's actually research about that. But then you have the other end of the spectrum, which are people who are using it for more deeper research tasks. The core use case there is students, tons and tons of students use the product and you think about it in the sense of, I'm writing a paper, I have my hypothesis. Now, instead of just trying to hack together word keywords and find papers and dig into them for the bits of information you're looking for to include in your paper, you can actually just type in your research question and consensus. And we'll deliver you back the insights about that question. So you can much more quickly get yourself to a point where it's, Hey, this is what I want to include. This is what backs up my argument or changes your argument because you realize that the weight of evidence is, you know, on the other side, but those are really the two, which is cool, because it's the kind of both ends of the spectrum where you have your more consumer use case, and then you have people using it in a real academic setting.

Justin Grammens  18:45  

I love it. How do you see this changing? Then the future of paper writing for students? I guess, right? If they're going to be more just fact checked by their peer nature, right? Not as much hypothetical,

Eric Olson  18:55  

you hear a lot of people thinking about the future of paper writing, because of the generative because of generative AI abroad, I guess ours is kind of, it's a tool that doesn't potentially like take, I'm saying this poorly, but because it's not generative ours representative, it's like a research tool alongside what you're doing, as opposed to speeding up your work as opposed to doing the work for you. And I think that is why we've been looked at more favorably by the academic community. We have some partnerships with libraries, we have libraries reaching out to us to be a part of the schools evidence portion or whatever tools that allow students to better get information. So like a generative tool, if you tell it to write your paper, it will write you how good it is still up in the air. I think that is what is really scaring a lot of people of all as you know are these tools can basically just do the work for students. Or as we kind of said, on a different side of the spectrum where it's, we can speed up your process we can quicker get you to the information looking forward, hopefully good information, but we're not actually doing the work for it. We're not generating you. So ours is more of a tool to make you more efficient not do The work for you, which I know is scaring a lot of people in academia now with these generative products, for sure.

Justin Grammens  20:05  

What do you guys see the future of consensus being and sort of the work that you're doing in the next three to five years or so? I mean, are there are there breakthroughs coming in? I mean, you know, NLP is just getting better and better. You guys looking at different languages? Translation, maybe maybe those papers were written in different language you could bring in I guess I'm just I'm spitballing ideas here. See, I'm just just just curious to see where you guys see the industry going?

Eric Olson  20:25  

Yeah, I mean, we do have requests for new languages all the time. So it's definitely something that we want to incorporate. But really, there's when the short to medium term like the one to one ish, one and a half to two year time period, there are two main, two main features that we're gonna that we are working toward that we want to be releasing soon, which are using some of these generative products, or generative tools to synthesize answers across papers. Now, we want to be careful, we want to make sure that they're used with guardrails, where we can say, alright, here are the top 20 results. Here's a summary of those 20 results. But here are the 20. In the results, the populate that summary or that were used to generate that summary are right here. So you can cross reference. So we want to be able to do some automatic synthesis for you. But we want to do it tied to source material where we're not just creating it out of thin air look kind of call this whole bucket of automated synthesis across sectors, then the other bucket is evidence scoring. So using these AI tools to extract other information from the papers that help us understand is this quality of research. So it's, you know, extracted things like what was the sample size? What type of study was it? Who funded this study? What was the effect size of the outcome, these are all things that you can train models, like a human expert would to read through a paper and pick out and analyze, eventually, what we want to get to is extracting all of those things to have that roll up to some sort of evidence competence score, tied to each one of those claims. And then if you can picture, we have a synthesis part of the product, but we already had with the underlying claims that are being used to generate that synthesis now have scores to them. We can do a really robust, cool automated evidence synthesis to your question, hey, we looked across 20 papers 13, say this five, say this to say this, here's the scores of each one of those schools of thought, effectively, it all boils down to if you're an expert, and you had infinite time, how would you assess what the answer to a research question as across the literature, you'd find the relevant papers, you'd pull out the key points of the paper I needed to assess the quality of the pavers, we want to automate that entire process. And then the five year vision is incorporated other types of data sources are laying is using AI to surface expert answers. We're starting with scientific research, it makes a million reasons why we think that makes sense. But there are other giant text documents that have expert insights within them. We'd love to incorporate other types of datasets to the product and build an easy to use AI powered search product overtop it's cool,

Justin Grammens  22:51  

good. You seeing anything interesting in sort of random off off the off the cuff question but you know, any, you're looking at virtual reality or, you know, augmented reality, you're in that type of stuff to interact with your product. I

Eric Olson  23:05  

can't say that we have I think that the the future of our product is using more of these language model tools that are continuing to get better or better. I don't know if we're the type of product will dip its toe in the in the metaverse and I think that maybe he will stand up. We'll all be in the metaverse in five years. And we'll be forced to worry to share that and not incorporate it but can't say we don't really thought much about it.

Justin Grammens  23:25  

I was thinking when you're talking about becoming the best, the best researcher or you know, asking a question, some people put a name to their product, or a persona or a person around it. And then you know, be really feels like it's somebody that you're interacting. So I wasn't sure if you guys were thinking about that or maybe go there and five years.

Eric Olson  23:42  

Those are actually threads. We have talked about some things like that, you know, talking to ChatGPT is like talking to the average of the Internet right now. And our goal were to build like an interactive chat feature. It would be haven't like talking to your scientists rent. And then you ever built out that module that was purely conversational. Yet we probably give it some called Einstein or something. Yeah, sure.

Justin Grammens  24:03  

Did you go right into a master's program to do this predictive analytics?

Eric Olson  24:07  

Yes. My background, my academic background before that was like traditional, like, business type degree. And I Yeah, it was not the easiest, but I had to teach myself how to code and kind of when I feel like one random trial by fire right into that program,

Justin Grammens  24:21  

would you recommend it?

Eric Olson  24:22  

I mean, there's definitely better on ramps if you dip your toe in computer science or data science and in undergraduate work, but you know, there's no trial by fire is is definitely an effective way to learn and many ways of forcing yourself in a forcing function to actually learn a specific type of skill. And when I took my job at DraftKings, there were still a lot that I didn't know I was still relatively new to the space and if I had, you know, said I can't take this because there's certain skills that are no Yeah, and I would have never been where I am today and learning on the job. You know, having the skills to figure things out on the job is arguably the most important skill of all as you get more more skills that you have the skill of figuring things out and learning more skills.

Justin Grammens  25:04  

For sure, yeah. So anyone that's maybe coming out of school, I guess, undergraduate or masters? How are you guys hiring? And I'm sure you're looking for certain skill sets the number one and then like I say, maybe put yourself in that in that graduate students choose? Yeah, maybe learning,

Eric Olson  25:19  

we're not hiring at the moment, we probably will be soon we're gonna be looking for some more traditional software engineers. First, I can speak to why the way the route that I took him why health into the AI world, but there are definitely there's a million different ways. And one of the beauties of the world today is that there's so much information available that you can teach yourself a lot of these things, there's online resources, you know, there's, there's really no excuse to get yourself up to a point of some competency if you really want it. But I think the on ramp for me was learning about data science and predictive analytics, it feels like it's extremely far away from artificial intelligence, the end of the day, there is more similar than that they're more similar than they are different. And it is still the same process of, you know, giving inputs and making predictions on it's still a bunch of X variables, trying to spit out a library. That's obviously simplified, but that understanding that art and science easiest RM to do that is understanding of Buddha's appear, okay to sign back up first. And then you can apply so many of those things, learning to more advanced artificial intelligence. And that's really how I learned to do it. And that's really why I felt qualified to lead this team. And so many of those learnings, even though it's simpler, quote, unquote, problems that I was working on in DraftKings, like traditional machine learning, modeling those skills of how to piece together a model, how to train a model, what matters are so applicable, even when you get to much more complex problems. So my advice overall would be a great on ramp is, is learning about the basics of data science and machine learning first, and using that to propel yourself into more advanced worlds.

Justin Grammens  27:02  

Awesome. Yeah, for sure. And I think, you know, as you were talking there, I was thinking about, you know, most people, it feels very foreign to be working in data, data science and numbers, and then to have natural language processing, which is all words, right. And I think a lot of people don't really see the connection. Oh, yeah. Like how it applies?

Eric Olson  27:20  

Well, yeah, well, that's kind of why NLP is exploded, is because of these embeddings, which basically give numerical representations of the words. That's how these models are able to learn so much, you're able to complete such complex tasks is because they learn from numerical representations of the words. So while it from a super high level, it feels like it's very, if you just looked at the surface, it feels like, like you just said numbers or swerves, it's like the complete opposite. It's really not at all. And if you're training a Simple NLP model, it is literally like training a machine learning model just with words, not numbers. And if you really boil down the brass tacks of what's going on under the hood, it is more basic data science. And because of these, these embeddings, basically make the words and numerical representations and then use traditional data science to make predictions on things

Justin Grammens  28:14  

is that fair to say, then that's sort of the aha moment or the epiphany that you saw a couple of years ago in this thing,

Eric Olson  28:20  

it's a part of it. But the real epiphany was these transformer models, which take out those embeddings and apply a type of warning that has allowed them to understand the context of language to the degree that was never possible before. So there's a really famous paper called attention is all you need. And it was paper by Google engineers, as they released their first transformer model convert, and we still use the Bert Bert base model. Today, I'm gonna screw it, I'm gonna screw it up. I'm not a full blown expert on every part of the specifics of it. But it boils down to a way of teaching the models to learn language, where it takes in the entire block of texts to make a prediction on a specific part of text. And repeat that pattern for to learn, learn about language and understand the context of language. So it's basically like saying, if you give it enough attack, things to pay attention to it can and it completely transform, no pun intended on powerful that these NLP models were. And really, I believe that they proceeded 2017 or 2018, that's when it really started to take off and was that same basic framework, the GPT is turned on. Were basically is given a giant brick of text. And then there are words that are maths, and you teach it by having to try to predict what words are missing, and then you tell it where it was wrong and you'd give that a billion examples, and then all of a sudden that learns how language functions and talk to you almost like the human get that technique. By Yeah, it was bi directional. And Colleen representative Grier, I forget bird stance or something but that was invented back in 2017. From a time Initial paper called attention is all you need.

Justin Grammens  30:02  

Awesome. Yeah, we definitely will put a link to that in the liner notes for this episode for sure. That's awesome. Yeah. And Google seems to have really been one of the major companies that have been sort of pushing forward on this.

Eric Olson  30:15  

Yeah, I mean, that is the basis of their main product I searched product is inherently unhealthy product. And how did that game for a while, but that has changed in the last year, that open AI would now be probably likely referred to as the real leader of the space. Now there are rumblings that Google has something behind the curtain that's even better than chat up. That's just rumors, I have no inside information there. That is yet to be released. That was, if you remember, like six or seven months ago, and it went viral, where there's a Google engineer and saying they'd created a sentient AI, what he was referring to is an unreleased chat bot model, that effectively would be a rival to chat GPT it has yet to be released. I don't know if they've seen something that terrifies them or something. But they are working on working on things under the surface. But open AI has really been the recent leader in this space. And what's cool about what they do is that they're not building products like Google, they're building foundational models to then be used by companies like us. So it's a pretty, the incentives are completely set up in the right way for innovation. There are all they're trying to do is put push the space forward, not, you know, hog all the resources in the model. So that bill,

Justin Grammens  31:25  

sure, that's good. Well, people can go to consensus and you know, sign up for free, right? And or at least there's limited trial to tell us a little bit about how that works.

Eric Olson  31:35  

Yeah, so consensus dot app is valued access, it is totally free for now, in early 2023, we are going to be releasing a premium tier with some of these features that we talked about. There'll be a small subscription fee per month. And then depending on a myriad of things, we may have some use of subscription as well. But right now we're trying to you know, we want to keep a portion of the product always free. Both selfishly and altruistically. We want it selfishly, because the great growth engine, it's a great way to get people onto the product and learning about us and discovering us and eventually hopefully converting to paid. But also, we think it's an important product. And we we want to be able to give some value to people regardless if they're able to pay or you know, needed every day like some people out some walked in, we want to have the ability to access it for free. So right now totally free to sign up and search in the future that will be a premium offering, both with features and potentially some users.

Justin Grammens  32:26  

Awesome. Awesome. Well, great. How do people get a hold of you, Eric? Yeah, LinkedIn, your

Eric Olson  32:30  

LinkedIn, Eric Olson, OLS ln, you can go on Twitter and follow us at consensus and Lt. I'm on Twitter all the time. If you DMS are open if you want to contact us. And then yeah, those are probably the two best ways wanted to run Twitter.

Justin Grammens  32:44  

It's good. It's good. Well, I'm curious, just to start ended off here as we started, like winding down, you still play sports, football, basketball, and those things

Eric Olson  32:52  

basketball, football is a bit harder to play recreationally. I don't know if my body would be too happy with me if I tried to shove on the past again, but I play basketball.

Justin Grammens  33:02  

Good, good. Yeah. I mean, have you found exercise and and you know, physical movement helps with your mental attitude? I guess with regards to running a business and everything.

Eric Olson  33:12  

1,000,000% you can look up that question at consensus and see all the amazing research on the benefits of exercise on decision making and cognition. But yeah, there is no better remedy to getting stressed about work than going for a run or playing a game of pickup basketball. And my co founder, like I said, is also a former athlete. And we definitely yeah, you will never be faulted if you sign off for a few hours to go exercise because, you know, not only do I want myself and employees to be doing well mentally, completely selfishly, but from a business perspective, you're likely, you know, if you just work straight for 12 hours versus work for 10 hours and exercise for one and a half. You'll probably be more effective doing the latter.

Justin Grammens  33:49  

Oh, good. Good. Okay, Eric. Well, I appreciate the time today and look forward to keeping in touch watching what's going on with consensus. i The product has fascinated me. I think you guys have touched on something and in a unique way. And I think like you said your differentiators really going to be in this space of making it easy to use more of a consumer based application.

Eric Olson  34:08  

That's exactly right. That was our that was our goal with all of us, making it easy to get hard. Good information.

Justin Grammens  34:15  

Oh, great, Eric. Well, you take care and thanks again and we'll definitely keep in touch Good luck at consensus. Extra awesome.

AI Announcer  34:23  

You've listened to another episode of the conversations on applied AI podcast. We hope you are eager to learn more about applying artificial intelligence and deep learning within your organization. You can visit us at applied ai.mn To keep up to date on our events and connect with our amazing community. Please don't hesitate to reach out to Justin at applied ai.mn If you are interested in participating in a future episode. Thank you for listening