The Art of Conversation: A Conversation with Aspen

Aspen is a researcher who has had the opportunity to engage in a wide range of conversations, from discussing favorite books to exploring the potential applications of AI. In this conversation, we delve into Aspen's thoughts on various topics, from her childhood dreams to her current work in academia.

Childhood Dreams and Aspirations

When asked about her childhood aspirations, Aspen revealed that she had always been drawn to theology. "I loved dolphins and whales when I was growing up," she said. "I really wanted to be a theologian." This early interest in a field outside of her current work highlights the complexity of human interests and aspirations.

Dreams and Interests

Aspen also shared with us that she had been inspired by the concept of "dream lab" - an idea that has the potential to transform our understanding of reality. While she didn't elaborate on what this meant, it is clear that she is drawn to ideas that challenge conventional thinking.

Favorite Things

In a lighthearted moment, Aspen revealed her love for the Internet movie, and confessed that she had been stuck on the concept of "times out" as a restaurant. "I guess a fishing pole sunscreen and a desalinator are not my cup of tea," she joked. Her favorite language to program in is Python, and she hopes to automate the perfect laundry system.

Research Interests

Aspen's research interests span a wide range of topics, from machine learning to The Art of Listening. She has been studying at Harvard, where she is enrolled in a course on this very subject. This highlights her commitment to interdisciplinary thinking and her desire to learn from others.

Future Aspirations

When asked about her future aspirations, Aspen revealed that she hopes to continue working on projects that have a positive impact on society. "I hope I'm continuing work that I feel strongly about and that has impact," she said. Her long-term goal is to collaborate with other researchers who share her passion for exploring new ideas.

Hidden Talents

Aspen also shared with us some of her hidden talents, including her ability to provide advice on thrifting. While this may seem unrelated to her academic work, it highlights the complexity of human interests and the value of being open to diverse pursuits.

Time Travel and Artistic Expression

In a thought-provoking moment, Aspen revealed that if she could travel into the future, she would hope to see significant progress on climate change. She also shared with us what artistic expression means to her: "mindfulness and being grounded in your world and in your feelings and in the experiences of others."

The Future of AI

Aspen's thoughts on AI are characterized by a sense of optimism and curiosity. While she acknowledges that some people may be fearful of AI, she believes that its potential applications hold great promise for human progress. "I'm really excited about this like growing potential for the applications of machine learning," she said.

Advice to Future Researchers

Finally, Aspen offered some sage advice to aspiring researchers: "find someone that you think is exciting and interesting to talk to and is in honestly doing interesting things." She emphasizes the importance of collaboration and being open to new ideas. By following her advice, future researchers can set themselves up for success in their academic pursuits.

Conclusion

Aspen's conversation has left us with a sense of wonder and awe at the complexity of human interests and aspirations. From her childhood dreams to her current work in academia, she embodies the spirit of curiosity and creativity that drives so many individuals in this field. As we look to the future, it is clear that Aspen will continue to be a force for positive change, using her unique perspective and talents to explore new ideas and pursue innovative projects.

"WEBVTTKind: captionsLanguage: enoh hello SE hello Aspen welcome to my home thank you for having us what makes this space feel like home besides my pet cow pig uh Good Vibes good food definitely good people good things all around that's all you need now I want to ask you a few questions about your research you started your journey in HCI and data visualization and then you moved to ml and interpret tell me about that yeah I think on the surface those two Fields seem very different but they both deal with this fundamental question of data and how do we as people make inferences on that data and make decisions with that data so that's you know the connection um I really wanted to be a well-rounded researcher and I loved the way that machine learning researchers and theorists often formalize evaluate and ask questions and so yeah that's that's that's really what got us here is well-roundedness okay very interdisciplinary approach now I'm getting a little distracted by that record player and you have any favorites I have so many um I listen to this record a lot when I'm working it's moonchild um it's got that R&B lowii energy that I think is just very conducive to work so priorities all right I'll have to give that a listen now back to your research one pillar is data sets how significant is building representative data sets in today's ml landscape I think it's one of the most important questions that we have is like a lot of the problems that come from bias um and issues of fairness in machine learning tie back to having not representative data sets so how can we build those data sets that cover you know the people the axes that we care about so that the product or tool that we build is is working the way we want it to and what are some fundamental challenges that you've run into doing this type of work well anytime you're trying to build something that is representative it's very hard to figure out what is not represented so I would say that is the number one challenge all right now I've heard of something called out of distribution methods what are they oh good question um out of distribution methods are this kind of like fueled in machine learning where people ask okay well I know what is in my training set um I know what my model has learned now can I figure out what the model is going to kind of do poorly on that maybe is not within that distribution so it's out of distribution sometimes we call that anomaly detection it depends on if you're working with time series or you know a different data type and how are you connecting that with data collection that is also a really good question um so I think until maybe recently we really had these topics very separated um but it turns out if you have a sense of the underlying distribution of your data set um and you take a partially trained model you can use that model and those ad distribution methods to Target direct and iterate your data collection so that in the long run you have a much more representative data set okay I see a big pile of books behind you as well sorry I keep getting distracted any favorites here oh I've got a lot um let's go with these two uh whereas by L long Soldier amazing amazing poet she really knows how to work with sound I love it and objectivity got to have it got to love it you know what time it is it's time to go to work speaking of that what's a typical day like at se- sale you know what I'll show you let's go on my way to work I usually listen to audiobooks podcasts or the occasional Spotify playlist what are you listening to right now I'm listening to ku's radio West it is a fantastic Radio show/ podcast okay sounds very popular for the East Coast now another pillar of your research fundamental ml work super applicable right now can you explain the debate on emergent World models um sure uh large language models um have been facing this really big question which is are they just stochastic parrots so memorizing surface statistics or are they learning something a little bit more meaningful about the world uh it's an ongoing debate like you said so there you go it's it's been summarized all right so how do you you think looking at really specific context is helping explore this question um that's a good question uh so a couple months ago my co-authors and I presented some work at I CLE we basically found an emergent World model uh in a large language model that was trained from scratch on a fellow moov so super super tightly constrained environment and that model was controllable and causal which is really exciting um for the future of interpretability and my my opinion but it's also just the beginning of the work so does that mean we're a little bit closer to deploying safer AI uh we're definitely not actively de deploying safer uh ML and uh I would say that we are also fully not finished with the interpretability research landscape although we're getting a lot closer okay so then with the ubiquity of llms people are starting to use them to generate training data sets so what are some challenges in using them for that another fantastic question Rachel um I I have to say so you know one of the nice things about large language models is they're really good at producing realistic outputs um but they're not so good at producing realistic distributions so any model that you train on that resulting distribution may have some problems look it's the gnome I'm going to go to work because I am now incredibly late from spending time with you guys I'll see you at the office oh my gosh welcome to work oh well thank you Aspen of course all right Aspen tell me a little bit about your current work the AI supply chain what is that that's a another good question Rachel you're so full of them um so up until now all the AI products that you might interact with like chat Bots Alexa an AI radiology assistant were largely built in-house along one pipeline so data cleaning development refinement evaluation training Etc was kind of ordered and organized by like one entity now it's a little different the products that you interact with are basically a bunch of different things glued together whether that's data sets models fine-tuned models the world is changing and we're thinking of this as a dispers learning setting where you can have a lot of different entities influencing the downstream inferences okay and how does dispersed learning impact the mitigation of bias and fairness in AI that's a really good question so when you have you know more than one entity contributing to a downstream product um and by entity it could be an organization it could be a person um but you know you have kind of like this idea of the downstream product doesn't have control over up and the Upstream products that are influencing it um as a result it's really hard to tell if something changes upstream and there's now a failure here who's responsible and this is a question of like liability and accountability and it's one of the big questions that we are trying to explore with the disperse learning setting and how could these challenges alter future developments in ml yeah I mean that's a really you know big problem um in disperse learning in the AI supply chain uh we're now trying to figure out how do you deal with bias and fairness when there are many actors that are playing a role in learning um it's hard you can have two perfectly Fair models and you maybe combine them um through um voting there's a lot of different ways to combine models but um you can average their weights for example and the result you know has this weird sense of like well who did what and now how do we know what's wrong um so it challenges stereotypical debugging strategies um debiasing um strategies Etc okay big open problem so I want to shift a little bit how have all of your non-academic experiences shaped your research and policy interests oh my gosh uh well I mean I think that's a really big question and one that is hard to answer uh I would say I I'm just very passionate about making people's lives better um and kind of getting to the heart of Truth and there are many ways to approach that from a policy perspective and there are many ways to approach that from a research perspective okay now for all of this work how do you balance the interdisciplinary nature of projects especially with Fields like sociology and anthropology I think finding good collaborators is an excellent way to do that um also reading a lot how did you stay on top of what's happening in your research field I think my lab is a really good resource uh and academic Twitter although I guess it's called X now um yeah talking to people it's a good way to do it all right moving on from the more technical stuff you moved from Utah to Cambridge for MIT that must have been quite the transition big culture shock for sure all right so did you venture outside of MIT for any collaborations now that you've been here you know what I did I got really lucky with um collaborations at Harvard and with people at Caltech and art center and with people at Apple from doing internships so very happy about those experiences that's wonderful now research can obviously be a long and arduous process so talk to me about some of the setbacks and what insights you've gleaned from that yeah I mean research is an arduous process and you often think something will work and then it doesn't and then you try another thing and it doesn't and uh I think one of the biggest things that you can develop as a researcher is a sense of resiliency and stubbornness um and a thick skin as someone once told me and throughout that process have you had any amazing mentorships or mentors you know what I really I definitely have um I think my current adviser Alexander majer is an amazing Mentor I got really lucky in undergrad with having amazing women mentors that that were professors that otherwise I probably would not have even considered Academia an option so uh yeah how has your approach to research matured over time you get better at asking good questions you have a better sense of the existing literature of what questions are interesting to the research community and to yourself and also how to start answering them so uh maturity I guess all right so are there any fun might Traditions or hidden spots on campus that Outsiders might not know about um maybe the muddy it's a pub uh for MIT Affiliates and it is quite the space to hear the occasional witty banter it's quite a scene is there anything about MIT that you wish you knew before you started here yeah um you know what they say about the hose being a water hose when you come here uh it's definitely true um you feel like you're drowning for a little bit and then you get your feet under you and you recalibrate and figure out your people and takes a minute but it's worth it I can only imagine so after a long day's work then what helps you recharge um yoga mindfulness sleeping all good things oh hi hi asan I was wondering through you Journey do you have any piece of advice that is the most impactful find people you like and work with them although we haven't worked together yet this isia she's awesome all right we're going now byee Aspen favorite CS resource the Internet movie oh God this is going to be hard happy as Lazaro all right favorite restaurant in Boston I didn't say this was going to be easy ah um times out isn't a restaurant but it's got a bunch of options and I like it okay artist um my mom class at MIT bills and billions oh you're doing great aren't you H this is pretty tough all right if you're stranded on a desert island three things you could take with you what's it going to be ooh uh I guess a fishing pole sunscreen and a desalinator is that is that how you say that not what I would choose but okay what's one thing that recently inspired you besides you oh you're too kind um I'm taking a Harvard class called The Art of listening and I think it's been really inspirational it sounds incredible all right Dream lab anyone from history who's coming you know what I don't know those people so I'll probably stick to the people I'm working with now because they're wonderful kind and super smart favorite language to program in Python all right if you could have any piece of AI technology automate something what would it be I want the perfect laundr mat I would like that as well let me know if you figure it out okay if your research project was a dish what would it be Sushi I love raw fish all right if you hadn't pursued a career in Academia in research what other career path would you have taken I mean this is still in research but I loved dolphins and whales when I was growing up and I really wanted to be a theologist so probably that okay so that was the dream as a kid any other dreams I mean besides being Mia ham for those of you who know who she is no okay based off of the pingpong skills I think we we might have a little difficulty there no I'm just kidding just kidding okay I think you were the problem it's possible so do you have any hidden skills or talents that don't involve ping pong or soccer I have so many um they're hard to count probably thrifting though everyone asks for my advice you're hopping in a time machine you're going back to any time period just for a day where would it be I don't want to go anywhere I really like um um it might sound a little wild but I think that the kind of work that people are doing and the kind of questions that people are doing right now are really important and also I don't want to be a woman in most any other era so I will second that what if you could go travel into the future can I travel like 200 years in the future and just see where we're at I'm hoping climate change is a problem solved that would be a pipe dream now what does artistic expression mean to you it means mindfulness and being grounded in your world and in your feelings and in the experiences of others now what has learning about machines taught you about the way that humans learn ooh uh a couple years ago I read this really interesting paper about numerosity which is you know how many objects are in a particular image and how we can recognize that and they used a neural network to explore this question and it had a lot of interesting implications for how people understand the count of OB objects all right A lot of people are fearful of AI are you I don't think I'm afraid of the same things that people are afraid of in pop culture at the moment um so yes and no what are you most excited about for AI I'm really excited about this like growing potential for the applications of machine learning and things that can help people so kind of like a feminist agenda you know um a friend of mine just did a bunch of work in capturing and I mean it's a little it's a little sad but she did a lot of really interesting work on capturing and highlighting these instances of femicide and I think that's incredibly important and it's hard to do without algorithmic support absolutely now in 5 years what do you want to be working on I hope I'm continuing work that I feel strongly about and that has impact um I hope honestly I I really hope that I am finding collaborations that are you know also interdisciplinary and that feel motivating to me beautiful now what advice would you offer to upcoming researchers trying to get involved in the field find someone that you think is exciting and interesting to talk to and is in honestly doing interesting things and literally just have a conversation with them it's one of the best ways to learn about a new space and it's also one of the best ways to be inspired okay last and final question we'll let you go where does the name Aspen come from um my mom likes trees oh I like trees too all right thanks Aspen we'll see you lateroh hello SE hello Aspen welcome to my home thank you for having us what makes this space feel like home besides my pet cow pig uh Good Vibes good food definitely good people good things all around that's all you need now I want to ask you a few questions about your research you started your journey in HCI and data visualization and then you moved to ml and interpret tell me about that yeah I think on the surface those two Fields seem very different but they both deal with this fundamental question of data and how do we as people make inferences on that data and make decisions with that data so that's you know the connection um I really wanted to be a well-rounded researcher and I loved the way that machine learning researchers and theorists often formalize evaluate and ask questions and so yeah that's that's that's really what got us here is well-roundedness okay very interdisciplinary approach now I'm getting a little distracted by that record player and you have any favorites I have so many um I listen to this record a lot when I'm working it's moonchild um it's got that R&B lowii energy that I think is just very conducive to work so priorities all right I'll have to give that a listen now back to your research one pillar is data sets how significant is building representative data sets in today's ml landscape I think it's one of the most important questions that we have is like a lot of the problems that come from bias um and issues of fairness in machine learning tie back to having not representative data sets so how can we build those data sets that cover you know the people the axes that we care about so that the product or tool that we build is is working the way we want it to and what are some fundamental challenges that you've run into doing this type of work well anytime you're trying to build something that is representative it's very hard to figure out what is not represented so I would say that is the number one challenge all right now I've heard of something called out of distribution methods what are they oh good question um out of distribution methods are this kind of like fueled in machine learning where people ask okay well I know what is in my training set um I know what my model has learned now can I figure out what the model is going to kind of do poorly on that maybe is not within that distribution so it's out of distribution sometimes we call that anomaly detection it depends on if you're working with time series or you know a different data type and how are you connecting that with data collection that is also a really good question um so I think until maybe recently we really had these topics very separated um but it turns out if you have a sense of the underlying distribution of your data set um and you take a partially trained model you can use that model and those ad distribution methods to Target direct and iterate your data collection so that in the long run you have a much more representative data set okay I see a big pile of books behind you as well sorry I keep getting distracted any favorites here oh I've got a lot um let's go with these two uh whereas by L long Soldier amazing amazing poet she really knows how to work with sound I love it and objectivity got to have it got to love it you know what time it is it's time to go to work speaking of that what's a typical day like at se- sale you know what I'll show you let's go on my way to work I usually listen to audiobooks podcasts or the occasional Spotify playlist what are you listening to right now I'm listening to ku's radio West it is a fantastic Radio show/ podcast okay sounds very popular for the East Coast now another pillar of your research fundamental ml work super applicable right now can you explain the debate on emergent World models um sure uh large language models um have been facing this really big question which is are they just stochastic parrots so memorizing surface statistics or are they learning something a little bit more meaningful about the world uh it's an ongoing debate like you said so there you go it's it's been summarized all right so how do you you think looking at really specific context is helping explore this question um that's a good question uh so a couple months ago my co-authors and I presented some work at I CLE we basically found an emergent World model uh in a large language model that was trained from scratch on a fellow moov so super super tightly constrained environment and that model was controllable and causal which is really exciting um for the future of interpretability and my my opinion but it's also just the beginning of the work so does that mean we're a little bit closer to deploying safer AI uh we're definitely not actively de deploying safer uh ML and uh I would say that we are also fully not finished with the interpretability research landscape although we're getting a lot closer okay so then with the ubiquity of llms people are starting to use them to generate training data sets so what are some challenges in using them for that another fantastic question Rachel um I I have to say so you know one of the nice things about large language models is they're really good at producing realistic outputs um but they're not so good at producing realistic distributions so any model that you train on that resulting distribution may have some problems look it's the gnome I'm going to go to work because I am now incredibly late from spending time with you guys I'll see you at the office oh my gosh welcome to work oh well thank you Aspen of course all right Aspen tell me a little bit about your current work the AI supply chain what is that that's a another good question Rachel you're so full of them um so up until now all the AI products that you might interact with like chat Bots Alexa an AI radiology assistant were largely built in-house along one pipeline so data cleaning development refinement evaluation training Etc was kind of ordered and organized by like one entity now it's a little different the products that you interact with are basically a bunch of different things glued together whether that's data sets models fine-tuned models the world is changing and we're thinking of this as a dispers learning setting where you can have a lot of different entities influencing the downstream inferences okay and how does dispersed learning impact the mitigation of bias and fairness in AI that's a really good question so when you have you know more than one entity contributing to a downstream product um and by entity it could be an organization it could be a person um but you know you have kind of like this idea of the downstream product doesn't have control over up and the Upstream products that are influencing it um as a result it's really hard to tell if something changes upstream and there's now a failure here who's responsible and this is a question of like liability and accountability and it's one of the big questions that we are trying to explore with the disperse learning setting and how could these challenges alter future developments in ml yeah I mean that's a really you know big problem um in disperse learning in the AI supply chain uh we're now trying to figure out how do you deal with bias and fairness when there are many actors that are playing a role in learning um it's hard you can have two perfectly Fair models and you maybe combine them um through um voting there's a lot of different ways to combine models but um you can average their weights for example and the result you know has this weird sense of like well who did what and now how do we know what's wrong um so it challenges stereotypical debugging strategies um debiasing um strategies Etc okay big open problem so I want to shift a little bit how have all of your non-academic experiences shaped your research and policy interests oh my gosh uh well I mean I think that's a really big question and one that is hard to answer uh I would say I I'm just very passionate about making people's lives better um and kind of getting to the heart of Truth and there are many ways to approach that from a policy perspective and there are many ways to approach that from a research perspective okay now for all of this work how do you balance the interdisciplinary nature of projects especially with Fields like sociology and anthropology I think finding good collaborators is an excellent way to do that um also reading a lot how did you stay on top of what's happening in your research field I think my lab is a really good resource uh and academic Twitter although I guess it's called X now um yeah talking to people it's a good way to do it all right moving on from the more technical stuff you moved from Utah to Cambridge for MIT that must have been quite the transition big culture shock for sure all right so did you venture outside of MIT for any collaborations now that you've been here you know what I did I got really lucky with um collaborations at Harvard and with people at Caltech and art center and with people at Apple from doing internships so very happy about those experiences that's wonderful now research can obviously be a long and arduous process so talk to me about some of the setbacks and what insights you've gleaned from that yeah I mean research is an arduous process and you often think something will work and then it doesn't and then you try another thing and it doesn't and uh I think one of the biggest things that you can develop as a researcher is a sense of resiliency and stubbornness um and a thick skin as someone once told me and throughout that process have you had any amazing mentorships or mentors you know what I really I definitely have um I think my current adviser Alexander majer is an amazing Mentor I got really lucky in undergrad with having amazing women mentors that that were professors that otherwise I probably would not have even considered Academia an option so uh yeah how has your approach to research matured over time you get better at asking good questions you have a better sense of the existing literature of what questions are interesting to the research community and to yourself and also how to start answering them so uh maturity I guess all right so are there any fun might Traditions or hidden spots on campus that Outsiders might not know about um maybe the muddy it's a pub uh for MIT Affiliates and it is quite the space to hear the occasional witty banter it's quite a scene is there anything about MIT that you wish you knew before you started here yeah um you know what they say about the hose being a water hose when you come here uh it's definitely true um you feel like you're drowning for a little bit and then you get your feet under you and you recalibrate and figure out your people and takes a minute but it's worth it I can only imagine so after a long day's work then what helps you recharge um yoga mindfulness sleeping all good things oh hi hi asan I was wondering through you Journey do you have any piece of advice that is the most impactful find people you like and work with them although we haven't worked together yet this isia she's awesome all right we're going now byee Aspen favorite CS resource the Internet movie oh God this is going to be hard happy as Lazaro all right favorite restaurant in Boston I didn't say this was going to be easy ah um times out isn't a restaurant but it's got a bunch of options and I like it okay artist um my mom class at MIT bills and billions oh you're doing great aren't you H this is pretty tough all right if you're stranded on a desert island three things you could take with you what's it going to be ooh uh I guess a fishing pole sunscreen and a desalinator is that is that how you say that not what I would choose but okay what's one thing that recently inspired you besides you oh you're too kind um I'm taking a Harvard class called The Art of listening and I think it's been really inspirational it sounds incredible all right Dream lab anyone from history who's coming you know what I don't know those people so I'll probably stick to the people I'm working with now because they're wonderful kind and super smart favorite language to program in Python all right if you could have any piece of AI technology automate something what would it be I want the perfect laundr mat I would like that as well let me know if you figure it out okay if your research project was a dish what would it be Sushi I love raw fish all right if you hadn't pursued a career in Academia in research what other career path would you have taken I mean this is still in research but I loved dolphins and whales when I was growing up and I really wanted to be a theologist so probably that okay so that was the dream as a kid any other dreams I mean besides being Mia ham for those of you who know who she is no okay based off of the pingpong skills I think we we might have a little difficulty there no I'm just kidding just kidding okay I think you were the problem it's possible so do you have any hidden skills or talents that don't involve ping pong or soccer I have so many um they're hard to count probably thrifting though everyone asks for my advice you're hopping in a time machine you're going back to any time period just for a day where would it be I don't want to go anywhere I really like um um it might sound a little wild but I think that the kind of work that people are doing and the kind of questions that people are doing right now are really important and also I don't want to be a woman in most any other era so I will second that what if you could go travel into the future can I travel like 200 years in the future and just see where we're at I'm hoping climate change is a problem solved that would be a pipe dream now what does artistic expression mean to you it means mindfulness and being grounded in your world and in your feelings and in the experiences of others now what has learning about machines taught you about the way that humans learn ooh uh a couple years ago I read this really interesting paper about numerosity which is you know how many objects are in a particular image and how we can recognize that and they used a neural network to explore this question and it had a lot of interesting implications for how people understand the count of OB objects all right A lot of people are fearful of AI are you I don't think I'm afraid of the same things that people are afraid of in pop culture at the moment um so yes and no what are you most excited about for AI I'm really excited about this like growing potential for the applications of machine learning and things that can help people so kind of like a feminist agenda you know um a friend of mine just did a bunch of work in capturing and I mean it's a little it's a little sad but she did a lot of really interesting work on capturing and highlighting these instances of femicide and I think that's incredibly important and it's hard to do without algorithmic support absolutely now in 5 years what do you want to be working on I hope I'm continuing work that I feel strongly about and that has impact um I hope honestly I I really hope that I am finding collaborations that are you know also interdisciplinary and that feel motivating to me beautiful now what advice would you offer to upcoming researchers trying to get involved in the field find someone that you think is exciting and interesting to talk to and is in honestly doing interesting things and literally just have a conversation with them it's one of the best ways to learn about a new space and it's also one of the best ways to be inspired okay last and final question we'll let you go where does the name Aspen come from um my mom likes trees oh I like trees too all right thanks Aspen we'll see you later\n"