Subscribe
Share
Share
Embed
Between Two Joels is your go-to source for making sense of AI. Hosted by Dig Insights AI experts, Joel Anderson our Chief Data Science Officer and Joel Armstrong, VP of AI, each episode explores the latest AI news and deep dives into specific topics, tools, trends, and ideas transforming business and insights.
Joel Armstrong (00:09.368)
Welcome to Between two Joels, where even though we both married women who were born in May nineteen eighty three, we're definitely two different people. I'm Joel Armstrong. And we're gonna get straight into some AI news.
Joel Anderson (00:16.109)
And I'm Joel Emerson.
Emma Sabry (00:22.842)
Microsoft is reportedly planning another round of layoffs, its third major workforce reduction in just over a year. The interesting part is why, while it's cutting jobs in areas like sales, consulting, and Xbox, it's simultaneously committing enormous amounts of money to AI infrastructure. It raises the question, are we seeing normal corporate belt tightening, or is this what an AI first company starts to look like? Do you think that AI is genuinely replacing roles yet?
Joel Armstrong (00:48.61)
So I think a lot of this is sort of corporate belt tightening under the guise of AI as an excuse to make these cuts. I will say I don't think it's just normal corporate belt tightening. I think there's two sort of historically exceptional things that are happening right now. One is the recovery post-COVID, where, you know, for a couple of years everyone was feverishly competing to hire every single person that they possibly could. And then as the reality has sort of set in, they realized that like
they didn't know how to use all of those people effectively and had overhired and had to make some changes while at the same time AI was suddenly becoming a thing. And so there was this thing everyone could point to if they wanted to to say, like, we don't need as many people as we used to because of AI. And there was this under the hood sort of thing happening where particularly at tech giants like Microsoft and the other fang stocks and all that, where they were suddenly spending hundreds of billions of dollars building out data centers as well. And so part of it is just sort of a, you know, a basic
capital rotation where all of a sudden instead of being software, which is you know super high margin and all you're paying for is talent to fund the future of your business, you need a bunch of super expensive hardware and then also some amount of jobs changing because of AI and then a bunch of overhiring to go with that. So in my head, that's what I think are the two major contributing factors. So
Joel Anderson (01:58.24)
I would say I'm gonna take the position, a slightly different position, where I think there's a lot of, you know, the companies that are doing this, like Microsoft as an example, right? They're they're a software company, surely they they also build a data center, but they also do a lot of software. Yeah. Software is the area that AI has demonstrated the biggest utility and the most economic applicability to the specific role the roles that they do, like writing code, like you know, Anthropic, they're all saying like they don't write any code. It's all written by Claude. And so
I do think that there is an effect. I mean, it's it's just all the things, right? It's it's all of them coming always everything. So there's always gonna be the macro effect, macroeconomic effects. It's gonna be sort of a good time to make these excuses because everyone's doing it. So it makes it easier to, you know, you don't get as many questions, easier to defend in a quarterly earnings call. I don't think it's broadly affecting every type of job and wiping off jobs all over the place. But I do think that with software developers, you can be more eff effective. Yeah, for especially for these big stops.
Joel Armstrong (02:33.048)
It's always everything.
Joel Anderson (02:57.398)
Software tech.
Joel Armstrong (02:58.27)
Yeah. And I mean we have a large software engineering team here at Dig and we work with them very closely. And we know that they are changing the way that they work. Like we can see up close that software engineering is changing pretty radically, especially in the last six months. So yeah, it does make sense. So it's kind of funny because like you say, like software engineering is particularly well suited to this because it's a linguistic equivalent in that
coding languages also have syntax and order and all the things that LLMs are really, really good at. And it has a bunch of thumbs up, thumbs down, did this work, did this not, in terms of error handling and error codes and all the way. Yeah. You get direct and objective feedback on did the code you write work or not. Yeah. And so as these larger language models have transitioned into being largely reinforcement learning based, then that matters a lot.
Joel Anderson (03:43.83)
Yeah. And the companies that are like Anthropic and OpenAI and and Google, they're AI companies, but it's all code. Like they're they have to write code to create their products and to create their yeah, to create the large language models in the first place. And so they can obviously the ta the best test ground is with themselves.
Joel Armstrong (03:59.942)
Yeah. And they know the industry super well. So they know how to get it used. They know what people want. It's like how there's a ton of different TV shows about making TV shows, right? Like is it people? Yeah, you know, 30 Rock and Sunset Strip and the Studio recently. That's their world. That's what they know. That's what they think about all the time. It happens to be a topic that fits their medium well as well as being the medium itself. and that's kind of what's happening with AI coding.
Emma Sabry (04:23.318)
So the next story we have is about cybersecurity. Researchers have analyzed what's being called the first AI-assisted ransomware attack. Although the headline is slightly misleading because AI didn't fully carry out the attack on its own, a human still had to step in multiple times. What do you think this means for the future? Are we headed towards AI-driven cyber attacks that will have little to no human involvement? Or what do you think?
Joel Armstrong (04:49.43)
I mean, I think we're definitely headed towards AI driven cyber attacks. As we were just talking about, AI is particularly well suited to software. So there will be people orchestrating AI driven software attacks and there will be cybersecurity companies using AI to get more effective at cybersecurity. The, you know, white hat, black hat arms race will continue as it has for the history of all software.
Joel Anderson (05:08.59)
Yeah, and we've talked a lot about how AI sort of amplifies your abilities. And so if you already are a hacker or r already trying to take ransom of different people's technical property, software property, then it then you're gonna use the tools that you have available, like AI, to to be better at what you're trying to do. I think it is particularly well suited to anything software. Like a like AI is particularly well suited to coding and all the stuff that we were just talking about with the other news article. And so yeah, it's it's gonna it's gonna make it
even more possible too. I mean on the on the flip side, that's a bit of a gloomer sort of perspective. But on the flip side, AI is also making it more effective to find security holes and patch them and and identify areas that we can make things more secure. So, you know, there's some some hope too. Yeah.
Joel Armstrong (05:55.742)
And we were talking a bunch actually for unrelated reasons. We were just talking about the fact that like we both feel like it's starting to be confirmed for us, what we've kind of suspected for a while, which is that like you have to know what you're trying to do to get AI to do it well. Like it doesn't offer you proactive no corrective suggestions. It will always try and do what you ask it to do. And so your vocabulary and your knowledge and your specificity when it comes to the thing that you're trying to do is fundamentally deterministic of how good a response you're going to get from AI and how useful it will be for the thing you're trying to do.
Joel Anderson (06:24.898)
Yeah, exactly. Like neither of us could go out, take advantage of someone's computer and and and ransom off their hard drive or something like that. If you have the thinking but your limiting factor is like mechanics of e of executing it, then AI is a huge amplifier.
Joel Armstrong (06:34.604)
The
Joel Armstrong (06:39.458)
Yeah. That's interesting too. It reminds me about a year ago someone said as like vibe coding was starting to enter the the vocabulary, the lexicon. The more the coding is no longer the blocker to the ideas that we can execute, the more important it becomes to have good ideas. We've been sort of saved from ourselves historically because it was so hard to execute on ideas, it protected us from demonstrating how hard it is to come up with good ideas.
So let's get right into our topic of the day, which is idea generation in the age of AI. We do a lot of innovation testing here at Dig, and so we know that idea generation is a fundamental component of innovation testing. If you're going to try and test new ideas and you have to be able to come up with new ideas. And the way that's done used to be a lot of, you know, people in rooms brainstorming, ideation sessions, all that kind of stuff. And that's changed a lot since AI, in particularly since large language models, which actually predates Chat GPT.
from when at least we started playing around with large language models. So why don't you tell us about some of the early experiments you were running in idea generation?
Joel Anderson (07:39.618)
Yes. Yeah. So we've been doing we started our AI masters a year before Chat GPD came out. And so we were playing around with a bunch of different ways, including I idea generation with with early large language models. And I remember testing, you know, like come up with ten different new ideas for donuts. And and back with GPD three, it would write down like chocolate and banana and strawberry, you know, very basic ideas.
For new donuts, the kind of idea that no one's gonna actually suggest in an innovation workshop. Not very helpful, right? But it's a starting point. And then at the time it was like, you know, prompt engineers was sort of starting to become a a thing. really, really kicked off after Chat GPD came out and AI went mainstream, but
Joel Armstrong (08:27.096)
But yeah, we should talk a little bit about how it used to work to prompt GPT three, because most people don't know because they didn't really start using LLMs until Chat GPT came out.
Joel Anderson (08:34.486)
Yeah, yeah. ChatGPD was the first large language model that was designed to be an assistant. So did all the pre-training, all the next word token prediction. So like complete the sentence kind of thing, like advanced autocomplete. That's kind of the fundamental architecture beh underpinning all of the largest language models. But the first three generations of GP GPT one, GPT two, GPT three, they they were just that. They were pre-trained models or trained models. There was just nothing after the pre-training, so just training. And then
One of the big unlocks, well, there's kind of two big unlocks with ChatGPT. One was it being like a convenient user interface that anyone could use. That was kind of a big thing that they didn't really expect to be big. But the other part of it, they engineered really specifically to be helpful. And they did expect that to be a huge boon to AI researchers. And that was changing it from unstructured predict the next token to question and answering type type of assistant. So before it was pre-trained, what's the next token it's gonna say?
And now at the time of GPT three point five, which was Chat GPT, it now did kind of question answer pairs. So they did post training on top of the pre training to get it to kind of act as a human exactly.
Joel Armstrong (09:41.742)
Assistant. Yeah, to make it conversational. Exactly. Right. The natural rhythm that people were used to where you say something and you expect an answer in return from someone who's kind of participating in a two person conversation.
Joel Anderson (09:51.522)
Yeah, exactly. Like you'd never walk up to a person and be like, What's the meaning of life? And then that person goes, Why am I here? Yeah. If that's how AI would act. Exactly. It treated it like it was a continuation of what it was saying, not like a conversation between two people.
Joel Armstrong (10:04.588)
Yeah, going back just a little bit, like the way that LLMs emerged, I think, kind of is informative of what the of the way that they used to work, right? Because it seems like they're supposed to be this kind of like super functional, like business AI sort of personal assistant sort of thing. And where they actually came from was actually just language models. Like we've been trying forever and we studied in our masters, like some of the historical models that were an attempt to make sense out of language, to use the natural structure inherent to language.
to predict what comes next and to make rules about what fits together and what makes sense and what doesn't. And for basically the entire history of machine learning and artificial intelligence, people had been trying to make some sort of accurate model of language. And that wasn't really possible or hadn't been achieved yet in a meaningful sense until Transformers, until Google discovered Transformers as like a new architecture of AI, large deep neural nets, and applied that to language, which happens to have a very convenient property.
which is that it's what's called self-supervised data, meaning that it tells you what the correct answer is when you're trying to generate the next word because the text contains the next word. So if you have a sequence of thirty words in a row and you're trying to guess what the next word is, you have all of text as a representation of the correct next thing to say. Yeah.
Joel Anderson (11:17.208)
Yeah, exactly. And so getting back to the idea generation part of things, back with GBD three, you could then like the early days pre-CHAT GPD, you could have a very sophisticated prompt to give it all the information, kind of like you know, a client might give us a sophisticated request for a proposal, like a project brief kind of thing. So then you have a lot of information up ahead so you can give them back an informed answer in the form of a proposal. Yeah. Right. So if you just said, I want to do a research study.
then we go, well, we have a lot of questions for you. Like we you know, we can't just give you an answer for that. We need to know more more specifics of what are you trying to research, what's your business objectives and what are your research objectives and what's some some of the background that you need to know. And that's that was the age of prompt engineers. It was the kind of idea that you needed to know how to have that conversation with AI because you you 'cause it's not as simple as just saying, give me five good donut ideas.
Joel Armstrong (12:11.884)
Okay. So then with Chat GPT, we had this new form of training that we talked about, right? Which was GPT three point five. And that really started to shape the model into being like a conversational agent that would reply to things that you said. But prompt engineering was still a very new idea at the time. And so a lot of the work at the time, a lot of people, including us who are, you know, pretty deep into AI, when we were trying to figure out how you actually use it, a lot of it was, you know,
looking into basic research around different techniques for eliciting desired behavior from it. So a big one at the time, a big paper in prompt engineering, arguably the biggest paper in prompt engineering for what it's led to was chain of thought reasoning, right? That if you tell a model to think step by step or give it examples of thinking step by step, that elicits from it a better behavior because we were learning how to get the models to behave the ways we wanted to and how you had to communicate with them and ask them for things to do that.
And so when we were doing that with idea generation, we found that if you gave it stylistically the type of ideas you wanted from it and gave it like a sort of sample structure and a sample number of adjectives and a style of adjectives and like you were saying, information about a business context and the purpose of the study and everything that you're trying to do, that we were like learning how to elicit desirable responses from the models.
Joel Anderson (13:29.346)
Yeah, exactly. That would work effectively if you gave it really creative ideas of the sort that you wanted to come up with. So going back to the the donut example, you you know, instead of having chocolate and strawberry, those are you know, what it might have given back to you. But then if you if your prompt included really specific and creative ideas like, you know, unicorn rainbow flakes, mm-hmm, then it it has an idea that you're okay, you're going for something that's a lot more creative than than just chocolate. Yeah. And then it would then it would
learn from that and take kind of that cue from the prompt and then come up with something that was more in line with what your expectations were. Right.
Joel Armstrong (14:04.63)
After GPT three point five, after Chat GPT kind of, you know, started this revolution in terms of what people expected from AI and how accessible and usable it was for different people, we had a generation or two where it was just sort of steady improvements in usability, right? GPT four and into four four. we're still in a pretty similar space where it was just it was easier to get it to do what you wanted to do as these methods of what's called reinforcement learning with human feedback, but as our ability to train it to act in desirable w ways.
increased and we got bigger models and more complex and more data, then the models just got kind of better in general, right? They just started being more what we wanted and it was easier to get them to do the things that we wanted. But all these prompt engineering techniques still largely applied.
Joel Anderson (14:49.516)
Yeah, exactly. And and at every step of the way we built different tools that our clients could use to generate ideas. A a little bit of information or a more detailed structured prompt where we would walk a client through all the steps behind it and and and even host client workshops where they work through this type of framework where they would go through, you know, giving all the background information and then giving the sort of your target audience that you wanna appeal that you're designing this new idea to market towards and
all that relevant information that you want to give the AI to to generate them. And then it can generate 50 for you that you can then use that. And then one of the tools that you built at one point was, you know, generate a bunch of ideas and then pick your top five or ten. And then it could use that to iterate to come up with another set of ideas that then, you know, it's using that as an a sort of a recursive prompt. Yeah. That it then gets finer and finer and closer and closer towards what you're what you're going for.
Joel Armstrong (15:46.134)
Yeah, for sure. And one of the elements of that tool too, which I think is relevant, is that you and I, in developing that tool, had control over the prompt and gave as few sort of degrees of freedom to the user as possible. Because at the time, most people weren't really comfortable with AI yet and didn't really know how to elicit desirable behaviors from it. And so we were able to design the prompts behind the scenes with like a few basically like Mad Lib style fill in the blanks. Yeah. And that made sure that we elicited the best possible response while giving the maximum freedom on the relevant
you know, dimensions of interest to the user. And another like just sort of funny detail, 'cause we haven't talked about this in two years now, is we also set the temperature of the model when we were doing that, which was sort of like, I mean, under the hood it's like a random basically a statistical randomness setting. But in terms of idea generation, it's essentially like a creativity and weirdness setting. So yeah, we used to offer the opportunity to kind of like have a slider for like how weird and unhinged of an answer do you want, compared to how kind of down the middle and representative of, you know
the most obvious answers the model would give you. What do want? So yeah. So that was kind of our first generation use of models when it came to idea generation was basically like prompt engineering and that sort of stuff, with a little like dabbling in what would later become context engineering, meaning we would kind of like steer the conversation in a way that made it more desirable. And then I think it's fair to say that things have kind of changed with, let's call it GPT-5, where the models just started getting good enough that you could start to just more
Joel Anderson (16:48.181)
Exactly.
Joel Armstrong (17:12.77)
directly and in more natural language, just kind of ask for what you want it, still including relevant details to guide it towards the sorts of things that you want. Yeah.
Joel Anderson (17:20.334)
Exactly. As you were talking before, with the impr incremental improvements for each model, even before they were chain of thought reasoning within themselves, they would still got to a point where they would come up with pretty good ideas. Like it had a better sense based on all the the human feedback and reinforcement learning that they did with the models. They would still come up with more creative ideas because it just knew. Just like if if someone asked me, come up with a donut idea, and if I said chocolate, and if I really thought of it, I I would know that that's not really a good donut idea. Right. In the same way that the model
as they got better and better, they just know that that's not a good that's not what you're looking for.
Joel Armstrong (17:54.956)
Yeah, there's like more inferred context, right? And it's the same as like if you had a kid versus an adult and you said to a kid, like, give me a good new idea for a donut flavor, then they're probably just gonna list off donut flavors they can think of. Whereas with an adult, they'll understand that good means like you'll they'll take more context into a kid. It would be a fun experiment. Yeah. Does anybody know the new donut? Does anybody know any kids?
Joel Anderson (18:11.532)
My kids tonight.
Joel Anderson (18:15.47)
Yeah, so the reasoning models enabled us to get even better. Yeah.
Joel Armstrong (18:20.194)
Yeah, that's a big thing that we didn't mention about GPT five yet, right? GPT five was really when the reasoning models, that generation became sort of the mainstream basic way. Yeah, exactly. They're baked into it, deep seek, all that kind of stuff. And so the reasoning models again had a big impact on like just the quality of responses that you would get based on pretty much anything you're putting in.
Joel Anderson (18:37.686)
Yeah, kinda like my example that if I said chocolate and then I immediately go, Well, that's not really the best idea. Like th it would go through that same chain of thought automatically under the hood.
Joel Armstrong (18:45.654)
Hundred percent. And interestingly, the thing that was actually baked into those models, the reasoning process, was directly inspired by those chain of thought papers from the very early days of GPT three and GPT three point five. They just trained the models to do the chain of thought thinking on their own. That was sort of the inspiration for at least the early like O one, O three generation reasoning models. Yeah. So that brings us to today. We're in the, you know, GPT five generation, and equivalent.
Joel Anderson (19:05.186)
Yeah, exactly.
Joel Armstrong (19:11.572)
So now we have a new method of idea generation. And it's pretty exciting because it's not just taking advantage of the best characteristics of modern, you know, GPT five generation foundation models, but it's also using some classic machine learning as well to get the best of both worlds and combine those. So do you want to say a little bit about what we're working on?
Joel Anderson (19:29.442)
Yeah. So the way we do idea generation now is kind of like we were describing earlier with this iterative loop that was kind of like our our V2 that you built a couple of years ago. but what we do now is instead of having having this iterative loop where the feedback is the the user saying, I like this, or this is more close to what I'm going for, we can actually use a machine learning model on our our database of all the ideas we've tested, which is over 300,000 ideas. So we've already learned all these different patterns.
Of what does well and what doesn't do well across all these different categories in our and in all these different countries in our database. And we can give that feedback to the AI that generates some of these ideas or to a human. We can seed it with a single idea to kind of optimize a single idea, or you could seed it with the whole category or all the corpus of all the ideas that you have together. So one is just optimizing an idea and one is coming up with ideas for the whole category, more open, you know, white space idea.
Joel Armstrong (20:25.718)
Yeah. So a while ago, when people started using AI to help with idea generation, one of the sort of early business school findings was that AI is really good at coming up with lots of ideas fast. And humans struggle with coming up with lots of ideas, but they're very good at screening them as being good or bad. And so by kind of playing those strengths off each other, previously the best way to do it was to generate as many ideas as possible and then have humans screen them as quickly as possible using the relative strengths of AI and humans in tandem.
What we've done now to speed up that loop and automate it and use more data than any one human would ever do is to train a model that's good at judging whether ideas are good or bad and then feed that back into the AI to again use an L L for its strengths, which is generating ideas quickly, and using a traditional machine learning model to predict specific scores, which it it can do better than humans would be able to any be able to do anyway.
Joel Anderson (21:13.196)
Yeah. And and that that AI model is built off of our proprietary data, like database of all these different ideas that we have. It knows the all these patterns of what does well and what doesn't do well. And then it can rank and and give give a numeric score for each of the ideas and then send that back to the large language model that can then learn and be like, okay, I'm I see that whenever it came up with chocolate ideas, they did well. And then it might learn when it comes up with other combinations of flavors they they didn't do well. And so it can take that feedback and then
come up with new ideas. And we can run this on many iterations of these two a different AI models. One's a large language model, one's more of a traditional machine learning model. And we can iterate them talking back and forth to each other. And what we find is you you actually you get a gradual improvement in your ideas. We've seen them go from, you know, increase on average by up to like two points out of a hundred per iteration. And you can continue to run that. And this is like very quick. Like these are two AIs talking to each other. So it happens very quickly.
And and you can see the pattern of of improvement. so it's a very exciting offering now because it's data driven. It's it's anchored in our own data that we have and that we've learned the patterns from. as opposed to the previous ver version was a human saying, I I like this, this is more of what I'm going for. That the human didn't know if those patterns were, you know, objectively do well with consumers.
Joel Armstrong (22:33.844)
Right. And that's what makes it kind of an interesting approach too, I think, is that you have this sort of mix of math and language. And it's obviously not fully divorced because L LMs are mathematical repr representations of language. But it's very interesting to me that the way that we search for alternatives is not in the math because there's another approach that we could be taking.
Because we do use the embeddings, right? Which is the mathematical representations of words and sentences and ideas. So we do use the embeddings in the predictive algorithm to say whether the score will do well or not. So it's not like it's beyond us to be using the embeddings to make these predictions. But when you just sort of search in the dimensional space using the math, you don't really know what direction you're going because the dimensions aren't very interpretable just by the nature of how these models work. So you'd be kind of randomly searching.
Whereas with this, because we're using a linguistic model, because we're using large language models, it is able to make educated guesses about what elements of the language should be preserved and which ones you can kind of wiggle around to do a targeted search in the area of the ideas that you think are interesting, particularly as you build up a bit of a body of like four or five or seven that are doing better and better. That helps you like zero in on the like meaning space where the good ideas are.
Joel Anderson (23:45.932)
Yeah, exactly. Yeah. Now the power of it is having the AI use the the large language model as the AI, talking to the predictive model that uses the embeddings. And so by doing it that way, we can extract that interpretable information out of it. Yeah. And what we haven't said so far is the large language model as part of this process used all these learnings to say, okay, this, you know, this chocolate, for example, did well. And then these combinations of different, you know, maybe fruity flavors or whatever didn't do as well. And then we can actually
The the AI can then write a report back to the to the user who kicked this whole process off and said, This is what I learned. Like I learned all this information as part of this ideation process. And this is all really useful diagnostic information that our clients find useful as they're exploring this white space and this ideation. it's no longer just about which idea does well and which idea doesn't do well. It's about giving that specific diagnostic and prescriptive feedback to a user about what.
has potential more than others and and what doesn't.
Joel Armstrong (24:44.258)
Yeah, no, I think it's really exciting. Like to me it's very sad like personally gratifying that we are reaching a point where we know well enough what the strengths and weaknesses of large language models are at this sort of like emergent level that they operate at, rather than just in terms of like the mathematical characteristics of them. We're learning like how they work with language from enough sort of like qualitative experience that we can make useful and educated guesses around like, well, it doesn't do this well.
But we have, you know, old methods for knowing how to do this. I like that that we're not throwing the baby out with the bathwater when it comes to classical machine learning techniques because they tell you something different. They use different kinds of data, you know, they're interpreted differently, they're four different things. And so it's nice, even if LLMs are now sort of, you know, the group 800-pound gorilla in the room when it comes to artificial intelligence models. It's fun when there's reasons to use other sorts of artificial intelligence models. It's just satisfying from a, you know, spending years and years the way we both have in terms of
Yeah, everything from li linear regression on up to language models, just trying to figure out the best way of predicting and understanding data.
Emma Sabry (25:45.118)
So now we're gonna play a game, finish the sentence, but AI trivia edition. The first question is the term artificial intelligence was first coined at a conference held at Blank in nineteen fifty six. Is it Dartmouth College, MIT, Bell Labs, or Stanford University?
Joel Anderson (26:02.84)
Bell Labs, right? I think it's Bell Labs.
Joel Armstrong (26:04.546)
That was my first thought.
No. Not not Dartmouth, I bet.
Joel Anderson (26:10.669)
It was Dartmouth.
Emma Sabry (26:11.562)
In nineteen fifty six, Dartmouth Summer Research Project, organized by John McCarthy, Marvin Minsky, and others, is widely credited with coining the term in launching AI as a field. Minsky.
Joel Armstrong (26:21.72)
Yeah, I thought so too.
Joel Anderson (26:22.67)
Totally thought they'll lapse.
Emma Sabry (26:25.934)
Question two. A neural network's ability to improve at a task through repeated exposure to data is called compiling, training, rendering, or indexing.
Joel Anderson (26:37.39)
Training. It's gotta be training.
Joel Armstrong (26:41.04)
yes, I agree. Correct.
Emma Sabry (26:42.478)
Okay, next question. The architecture underlying most modern language models introduced in twenty seventeen paper is called the
Joel Armstrong (26:52.43)
Transformer Transformer. The paper is called Attention Is All You Need. Yeah.
Emma Sabry (26:55.638)
Correct. So far you guys have two out of three. In machine learning, when a model performs very well on training data, but poorly on new data, it's said to be overfitting. Underfitting, overfitting, converging.
Joel Armstrong (27:08.47)
It's overfitting. It's overfitting. Correct. Do you wanna do it? Do you wanna try without looking at the screen? See if we can just fill them in? Just have a read and we'll just fill in the blanks.
Emma Sabry (27:16.054)
This next one seems easy. Hard mode. Okay. Breaking text into smaller chunks the model can process like words or subwords is called nice.
Joel Anderson (27:17.814)
Give us hard mode.
Joel Armstrong (27:25.224)
So we are racing. Let's say let's just say it at the same time in the blank.
Emma Sabry (27:31.214)
Okay, next one. The nineteen ninety-seven chess match that made headlines when a computer defeated the reigning world champion featured IBM's.
Joel Anderson (27:32.999)
Okay, sounds good.
Joel Armstrong (27:42.04)
D and Kasparov. Joel I mean w we both know the rules, but Joel is one of his hobbies and I really played chess in twenty years.
Emma Sabry (27:44.174)
Nein.
Emma Sabry (27:49.74)
Can you guys play?
Emma Sabry (27:55.746)
You like to play chess? By yourself or against someone?
Joel Anderson (27:58.434)
Well, not with people.
Joel Armstrong (27:59.998)
Against the mirror. You run back and forth. You're like I've got you now. You always know what your opponent's doing.
Joel Anderson (28:03.861)
Fool.
Emma Sabry (28:09.422)
Okay, last question. The field concerned with making AI systems decisions understandable to humans is called AI.
Joel Armstrong (28:19.414)
Under making it understandable.
Joel Anderson (28:21.304)
I suppose anthropics interpretability. Yeah, I think it's interpretability. I mean interpretability.
Joel Armstrong (28:24.558)
Interpretability.
Emma Sabry (28:28.608)
Nice. That's it. Great job, guys.
Joel Armstrong (28:31.598)
Everything but dark mouth.
Emma Sabry (28:34.446)
But Dartmouth, yeah, really good job.
Joel Armstrong (28:39.886)
All right, that's everything we got for this week. Thanks for joining us at Between Two Jewels and we'll see you next time.
That's all making it.