Full episode transcript
This may sound unusual, however I like a great container. You recognize, the thought of getting all the things organized all neat and tidy and saved in a single place. And in case you suppose it is a passing fad, let me guarantee you: I’ve been like this my entire life.
On the age of eight I had “Matt’s Tenting Folder”—like, a literal paper folder with “labeled” scrawled throughout the entrance that contained all the info our household would wish for our annual summer season tenting journeys. Street maps and park guides and to-do lists galore.
Once I was in highschool, I begged my mother and father to go to IKEA to take a look at all of the desks and cabinets. Very cool!
And now, in my previous age, I’ve embraced the bin: huge plastic containers for towels, sneakers, vacation decorations, patio tiles, and—full circle—tenting gear. I stack them in my closets, my storage locker, and I hold a working stock of all the things inside, so I at all times know the place one thing is.
Distinction that with my digital life the place it looks like chaos is the secret.
I’ve given up on my inbox. I barely manage my recordsdata. And it’s not like all of it issues anyway, as a result of most of my work is occurring in apps and tabs—every with their very own file techniques and search filters and approaches to group. It’s a multitude!
So naturally, I’ve been intrigued by people who find themselves utilizing AI to discover a sign in all that noise. The people who find themselves determining the best way to take all their digital containers, and neatly manage them in a single huge AI-powered field.
I’m your host Matthew Braga, and on at the moment’s episode of Working Smarter we’re speaking to Sophia Wang. She’s an assistant professor of ophthalmology at Stanford College, the place she leads the ophthalmic informatics and synthetic intelligence group.
Sophia and her colleagues even have their containers—digital well being data, docs’ notes, diagnostic imagery and extra. However they suppose synthetic intelligence can assist carry some order to that chaos, and assist them uncover new insights which may not have been as apparent had they only gone rummaging round on their very own.
What does that imply for Sophia, her sufferers, and the way healthcare professionals do their jobs? That’s arising subsequent on this episode of Working Smarter.
~ ~ ~
Sophia, thanks a lot for becoming a member of us at the moment.
Thanks a lot for having me.
I have been trying ahead to this. So that you’re an assistant professor of ophthalmology at Stanford College the place you lead the ophthalmic informatics and synthetic intelligence group. Are you able to inform me extra about that group and the work you do there?
That is our analysis group. We’re very keen on constructing predictive fashions for numerous ophthalmic outcomes. So, virtually talking, meaning we use all completely different sorts of knowledge sources associated to ophthalmology to foretell how our sufferers will do sooner or later—whether or not our glaucoma sufferers’ imaginative and prescient will worsen or different kinds of outcomes.
Received it. You talked about glaucoma a second in the past. Simply briefly, for folk who may not be acquainted, what’s that?
Glaucoma is a progressive illness of the optic nerve. And the optic nerve is type of just like the cable that connects your eye to your mind. It transmits all of the sign about what you are seeing to your mind for additional processing so that you’re aware of what you are seeing. And naturally, if that nerve is diseased then it isn’t transmitting the sign correctly and you’ll expertise that as having blind spots in your imaginative and prescient. So in glaucoma, that nerve turns into diseased and also you get attribute blind spots in your imaginative and prescient that then develop into disabling.
Okay. So at the moment, how are docs making glaucoma-related diagnoses and choices? I am considering, if I’m going for an appointment or an examination, what are the sources of knowledge or the issues that ophthalmologists are utilizing as a part of that analysis or a part of that course of?
The very first go is simply the final ophthalmic examination. In fact, we examine issues like your imaginative and prescient. We have a look at the optic nerve utilizing our particular lamps and magnifying glasses. We are able to really see it with our personal eyes. And if it seems to be a sure manner, maybe just a little bit suspicious—the construction of it possibly seems to be just a little completely different, or maybe it is uneven between your proper and left eyes—that type of raises one of many first flags of, oh, possibly this optic nerve should be evaluated with extra in-detail testing for glaucoma.
We additionally examine to your eye stress, which is a kind of components that’s extremely associated to glaucoma. After which when you elevate that flag, there’s a lot of extra checks that we are able to do. For instance, we are able to do visible area testing, which is a proper take a look at for these blind spots, as a result of our mind is actually good at ignoring blind spots or filling them in or masking them in. Typically in early illness we do not actually discover that till we do formal testing for it.
And so all of this info—just like the take a look at knowledge, the stuff that you just’re observing—the place does all of that go?
Lots of our info flows, basically, to what we name digital well being data. And that is type of an umbrella time period for all of the well being data, all of the well being knowledge, that’s generated throughout your encounters with physicians and the healthcare system. Inside that umbrella time period, there’s a lot of completely different sorts of pots of various knowledge. I would put your visible acuity or what letter you see on the attention chart in a single spot and your eye stress in a single spot. However then the photographs that we take of your optic nerve may go in a special system, which remains to be digital well being data however, you already know, type of lives in its personal space so to talk.
After which I think about on prime of that you’ve physician’s notes after which different kinds of knowledge.
That is proper, we take detailed notes, physician’s notes. Some individuals wish to dictate them. Some individuals sort them out. Typically it’s extremely templated language, and generally it is free textual content, primarily.
Are you the kind of one that can learn your personal handwriting again or do you generally encounter one thing and, like, it appears inscrutable in hindsight?
[Laughs.] Nicely, most of us are usually not retaining handwritten data anymore, thank goodness.
So is it honest to say that—it would not sound such as you’re missing for knowledge, however it’s extra about, you already know, how will we successfully analyze all this info and make sense of all the info now we have?
That is proper. We now have a lot knowledge that we’re capturing simply on a routine foundation. Everybody who walks by way of the door will get all this eye examination knowledge that is captured—all the photographs, all of the testing, all of the notes. However generally it’s extremely laborious to extract the particular piece of knowledge that we wish from this sea of knowledge. In order that’s actually a problem.
And if I could ask, why is it so troublesome to tug all that info out?
A few of that info is unstructured, so to talk. In case your optic nerve seems to be a sure manner, or we observe a specific discovering, generally we simply sort that out in phrases in our notes. And docs will use completely different sorts of phrases relying on their desire. And so, for those who’re searching for that discovering, generally it’s extremely laborious to truly determine that on a big scale from, to illustrate, hundreds of notes for hundreds of sufferers.
We do have structured methods of storing knowledge. There’s at all times the billing codes or process codes that inform insurance coverage corporations what illnesses we handled or what was completed that day. However they had been by no means meant for analysis functions. They had been actually meant to speak with insurance coverage corporations. So the info may not at all times be as correct or as detailed as we would need in analysis.
So you’ve gotten all of this knowledge and also you wish to determine how one can analyze it and make sense of it. When did AI develop into a part of that course of for you?
By my profession I’ve stepped up or graduated by way of knowledge units which have develop into more and more complicated and messy and unstructured. And I will offer you an instance of this. There’s knowledge that we gather particularly for analysis—survey knowledge and issues like that—that somebody thinks up forward of time precisely what they wish to seize and in what format and so forth. And really, our Middle for Illness Management (CDC) runs a lot of these kinds of nationwide surveys. So the info that comes out of these nationwide surveys are very, very clear, and really, very structured. However as you get into extra of the routinely generated healthcare knowledge that is obtainable—for instance, insurance coverage databases—they develop into extra complicated and extra obscure.
As digital well being data grew to become increasingly more extensively adopted during the last 10 years or so, we had been getting rising quantities of much more complicated unstructured knowledge, like these free textual content notes, like these imaging databases. And my preliminary curiosity was to have the ability to perceive particulars in regards to the affected person from these unstructured knowledge units. As a result of, you already know, as a physician, we’re at all times typing into the notes and we’re taking all these very detailed observations down. And I assumed, nicely, would not it’s nice if we may do analysis on a big scale utilizing that knowledge, which is not in handwritten notes, however now all typed and picked up in digital well being data. So my preliminary angle for the factitious intelligence half was by way of pure language processing of our medical free textual content notes.
You’ve got co-authored a lot of papers now during the last couple years taking a look at methods of utilizing machine studying—stuff like deep studying, you talked about pure language processing, LLMs, pc imaginative and prescient—within the area of ophthalmology. What have been among the most promising areas for making use of AI in your work that you’ve got explored to date?
I’ve been primarily centered on predicting which sufferers with glaucoma will progress—that means their illness will worsen. And it is a actually essential matter, as a result of as soon as your glaucoma illness progresses and your optic nerve turns into increasingly more sick, as a result of it is nerve tissue you possibly can’t actually reverse it. You may’t regenerate it, not less than not proper now. So, it is actually essential to have the ability to determine which sufferers are going to worsen and lose their imaginative and prescient sooner or later. As a result of if we knew that earlier than they really obtained worse, then possibly we may change the best way we deal with these sufferers—being extra proactive and possibly give them extra drugs earlier, or do procedures earlier.
However we do know that glaucoma is a progressive illness. Not everybody will progress. And a few will stay secure with much less invasive remedies early on. In order that’s the appliance or use case that I have been specializing in in my analysis. And Glaucoma may be very attention-grabbing as a result of, as I discussed, there are lots of completely different ways in which we consider the illness. There’s all of the measurements we take of the attention, the imaginative and prescient, the stress. There are the photographs that we take. There are the visible area checks searching for the blind spots. All these completely different sorts of knowledge come collectively to inform us in regards to the state of the illness. And so it has been my curiosity to see if we are able to construct fashions which combine a number of completely different sorts of knowledge, and due to this fact combine a number of completely different strategies—for instance, pure language processing and pc imaginative and prescient to take a look at each the notes and the photographs, for instance—to higher predict the long run outcomes.
Why is the development of the illness so troublesome to foretell with conventional strategies or previous analyses that you’ve got used?
Nicely, it is at all times laborious to foretell the long run, I believe. [Laughs.]
[Laughs.] Yeah, that is honest.
I imply, even for those who ask a physician, a glaucoma specialist, “Who do you suppose goes to worsen? Who do you suppose are going to be these quick progressors?”—now we have some thought. We all know of some threat components. However quite a lot of the components that we take into consideration are additionally not components which can be simply captured in these sorts of structured code knowledge—so issues like, are the sufferers taking their drugs usually, or possibly what sorts of medicines are getting used, or have they got a household historical past of glaucoma.
Attention-grabbing. And by way of the precise course of—like, the best way wherein you’re making use of ML—are there explicit bottlenecks or sources of friction in your job and the best way that you just’re analyzing info or attempting to make choices that ML is uniquely suited to serving to with?
Nicely quite a lot of the bottlenecks for growing a system like this should do with really assembling the info forward of time. All this knowledge kind of lives in its personal areas throughout the digital well being data. So attempting to place collectively or match up the photographs, the visible area checks, the scans and all of the notes and the measurements for a affected person or many sufferers—that is very laborious to combine and to place all of it collectively in a research-ready database.
One other problem of this explicit type of knowledge is its longitudinal nature. It isn’t like figuring out whether or not an image is of a cat or a canine, for instance, the place you possibly can simply have a look at one image and you may resolve as a human, proper? Typically the analysis of glaucoma sufferers takes place over a lot of visits. You need to see evolution of the illness to make an actual willpower. Typically the time interval between visits is completely different for various sufferers, and placing that each one collectively right into a machine studying mannequin—which, in some methods, might be very rigid as to the construction of the info that goes into it—is usually a problem too.
Attention-grabbing. Is there additionally a chance there? And I assume what I imply by that’s, I think about for people taking a look at knowledge over a very very long time scale it may possibly generally be laborious to identify these patterns. However my understanding is that is really one of many issues that machine studying might be actually good at—having the ability to choose these patterns out that we miss. Is that honest to say about this explicit work?
Sure, I believe that may be a actually nice perception, and that’s undoubtedly one of many issues we’re hoping that machine studying can assist us obtain. Oftentimes as clinicians, we would see a brand new affected person they usually may include years of data from their prior glaucoma specialist. You recognize, they only moved into the world, and we’re trying by way of all of this as we’re evaluating the affected person. However it’s quite a lot of knowledge for even a human clinician to course of. So if we had a very efficient machine studying algorithm that may assist with that, that would definitely be actually useful.
And only for listeners who may not know, you are speaking about years of knowledge right here. What’s the typical age of a glaucoma affected person?
Proper, so it’s an age-related illness. Usually our glaucoma sufferers are older—60s or 70s. So if somebody was identified with glaucoma of their 60s however now they’re 75, then you definitely’ve obtained 15 years of potential glaucoma data to look by way of for his or her total historical past.
Very attention-grabbing. So with this work that you’ve got been doing to date of making use of AI, and machine studying particularly, to ophthalmology, what’s shocked you most to date?
A few of the structured info within the digital well being data is excellent at predicting outcomes even with out the supplemental, to illustrate, unstructured knowledge—medical free textual content or imaging. And we discovered this out as a result of these sorts of structured digital well being data knowledge are mainly the simplest to assemble first. So oftentimes after we’re beginning off in analysis we are able to pull that knowledge extra rapidly. And after we began constructing our fashions predicting glaucoma outcomes the efficiency was higher than we anticipated for predicting the long run utilizing actually simply primary eye examination knowledge—demographics, billing codes—for his or her diagnoses.
And I assumed that we would wish much more info coming from imaging and medical notes to make these fashions carry out in addition to they did. However really, after we added these extra modalities of knowledge, sure, we obtained higher efficiency, however the baseline efficiency utilizing simply the structured knowledge is fairly good.
It sounds prefer it’s not nearly having extra knowledge essentially but in addition beginning with the correct knowledge. Would you say that is honest?
Sure. It isn’t simply having extra knowledge, however it’s the correct knowledge, and you need to be just a little bit considerate about the way you manage that knowledge even earlier than it goes into the mannequin. You may put in additional knowledge, however it is likely to be simply extra noise and less sign so to talk.
You had been speaking in regards to the high quality of the predictions a second in the past, and I believe I learn in certainly one of your papers that one of many algorithms you had been engaged on was capable of predict affected person outcomes with, like, 80 p.c accuracy—which appears fairly excessive, fairly spectacular. However how does that evaluate to human efficiency? How will we put this into context?
This can be a nice query. It is at all times essential to place these sorts of efficiency metrics in human context. And I will begin with a contrasting instance, which is, once more, classification of pictures. We all know that cats versus canines in classification of pictures, a human can just about at all times do that. So that you’re actually aiming for, like, one hundred pc accuracy for any type of machine process. And equally, for sure ophthalmology duties like taking a look at an image of a retina and figuring out whether or not there is likely to be diabetic eye illness in it, that is one thing {that a} human taking a look at a photograph can classify very precisely. So all these algorithms should be tremendous excessive performing.
However a extra complicated process like determining if a affected person will progress sooner or later—that is not taking a look at one thing and classifying it. That is predicting the long run. And so, as I discussed, even people are usually not superb at this. And in an effort to show that for certainly one of my research, I simply went by way of a small group of sufferers, possibly 300 sufferers or one thing, and I simply checked out their data. And I attempted to foretell as a glaucoma specialist, oh, who’s going to wish glaucoma surgical procedure and who’s not going to wish glaucoma surgical procedure. I wish to suppose I am a fairly good physician, however it seems that my guesses weren’t good in any respect. [Laughs] I believe for those who had flipped the path of my guesses, it’d even have been like a greater performing future. So, you already know, it would not take a lot, really, to do higher than a glaucoma specialist.
Nicely, and that is so attention-grabbing to me as a result of, however, I believe if you carry expertise into quite a lot of contexts—particularly health-related contexts—individuals might need questions or they may have have some considerations or issues they wish to be reassured about. How do you construct belief in these sorts of instruments and strategies? Each for docs, but in addition for sufferers?
There are various reservations in regards to the precise use of AI in healthcare. Nearly at all times, after we attempt to publish certainly one of our latest papers and, and say, “Oh, look, this algorithm does fairly nicely in comparison with people at predicting outcomes,” there are quite a lot of questions that come up. What are the options, what are the inputs that these algorithms are taking a look at? Are we taking a look at affordable inputs, or are they only detecting noise, mainly? So virtually at all times now we have to analyze our algorithms for what are the essential options that they are counting on and make it possible for these are according to our medical reasoning, so to talk.
Different questions which can be more and more arising now are how do these algorithms carry out in subsets of sufferers? Both demographic subgroups or subgroups of sufferers with explicit illnesses or explicit options—like, sufferers who’ve already had surgical procedures or already had different procedures. And I believe it is crucial to do all of those analyses earlier than you get to the purpose the place you’re advocating that they instantly be carried out.
That is sensible. For docs and clinicians, what does success appear to be if you’re utilizing these instruments? Is it time saved? Is it the power to assist extra sufferers? Make higher choices? How would you characterize that?
I believe for a instrument that may be a medical resolution help instrument we wish to see that it is serving to our sufferers. That it’s actually, for instance, enhancing imaginative and prescient or enhancing surgical outcomes. If we constructed an algorithm that predicted how nicely sufferers would do after completely different sorts of surgical procedures—which is one other space of analysis—we’d need our surgeons who use this algorithm to finally have higher surgical outcomes as a result of they’ve chosen the suitable surgical procedure for the suitable sufferers. So these are, I believe, medical benchmarks for fulfillment. I believe in an effort to show that an AI algorithm actually helps sufferers we ultimately might want to do extra conventional research—like randomized managed trials of sufferers who had been handled with the assistance of an algorithm versus sufferers who weren’t—and actually have a look at these outcomes to see that it is serving to sufferers.
How far off would you say these are?
In our area, not that shut, I’d say.
Why is that?
I believe there’s quite a lot of work to be completed in evaluating equity of our algorithms earlier than we deploy them—to find out whether or not algorithms that I’ve skilled utilizing Stanford knowledge are going to work wherever else, on sufferers that aren’t at Stanford and issues like that. I believe to get to the purpose the place we are able to take a look at an algorithm in a big randomized management trial—which takes quite a lot of sources to design and execute a research like that—now we have to have all these kinks labored out, all these all these evaluations completed, earlier than we get to that time. However that is the purpose.
So with these objectives in thoughts, what are you engaged on proper now?
We have completed quite a lot of predictive modeling for glaucoma sufferers in our Stanford sufferers. However the subsequent step is to transcend one website and entry multi-center knowledge—coaching larger fashions, testing whether or not fashions skilled in a single place will work in one other place. So one of many issues that I am engaged on now could be coaching these glaucoma consequence fashions utilizing knowledge from SOURCE, which is brief for Website Outcomes Analysis Collaborative. It’s a multi-center collaborative made up of 17 educational facilities that see eye atients, and we’re all contributing our digital well being data on our eye sufferers to this consortium or registry. It is run out of the College of Michigan proper now by the chief knowledge officer Dr. Joshua Stein, an ophthalmologist and informaticist who’s harmonizing all of this knowledge and placing it collectively and bringing on extra websites. It is 17 now, however there’s many extra within the strategy of becoming a member of. I believe this type of initiative may be very thrilling as a result of we are able to now do much more of the research that we could not do earlier than.
Earlier than you had used the phrase supportive if you had been describing these instruments, and I like that as a result of I do know that at Dropbox we frequently speak about AI as being one thing assistive, augmentative—one thing that helps you do your job and would not do your job for you. How have you ever been occupied with this framing if you’ve been engaged on among the analysis that you’ve got been doing?
We are attempting to construct these algorithms to help physicians of their therapy of sufferers—to supply some extra knowledge for them to take a look at and resolve, “Okay, nicely, you already know, this algorithm means that this affected person is likely to be a excessive threat affected person.” And like lots of our different knowledge inputs, ultimately it could develop into one which we resolve to concentrate to if the algorithm is excellent, proper? However finally I believe it is as much as the person or the doctor to resolve the best way to use that AI-generated recommendation. As a result of because the human doctor that’s treating the human affected person in entrance of us, we really know much more about that affected person than possibly might be captured even with the astounding breadth and depth of our digital well being data. An algorithm may counsel {that a} affected person may do nicely with a sure type of surgical procedure—however I would know that affected person is not able to doing the complicated routine of therapy that is required to make that surgical procedure a hit, for instance. I will have all these components along with no matter is likely to be going into the algorithm, to assist additional personalize the remedies to that affected person.
That is sensible. I imply, empathy looks as if one thing that solely docs can carry to the desk.
Certain. Empathy and explaining in a manner that the affected person can perceive the place they’re and what we’re attempting to do for them—and presenting the alternatives to the sufferers, too. Finally, it is as much as them what their medical care goes to appear to be.
What about for sufferers as nicely? You kind of alluded to this just a little bit earlier than, however may AI-powered instruments and among the algorithms you are engaged on assist enhance entry to care or the standard of care particularly amongst underserved populations?
That’s undoubtedly a purpose. We hope that AI will enhance look after underserved populations. We do not wish to deploy techniques that really worsen look after underserved populations or widen the disparities that we all know exist already in glaucoma care and in ophthalmic care. If we’re capable of construct algorithms that may be deployed to, for instance, detect glaucoma earlier slightly than ready until it is symptomatic and superior for these sufferers to current to us, that is a technique that it may possibly actually increase entry to care.
Extra broadly talking, within the area of medication, how does the work and the analysis that you’ve got been doing with AI evaluate to what different persons are doing in different medical fields? I am interested in the place this suits into the vary of AI purposes or potentialities that persons are at the moment exploring.
I believe ophthalmology as a area has been one of many specialties in drugs that has been main synthetic intelligence purposes. And quite a lot of that early work, and present work, is in pc imaginative and prescient, as a result of we’re such a closely visible—no pun meant [laughs]—area. We take so many photos and scans of the attention, which is kind of a pure substrate for pc imaginative and prescient. We even have an autonomous system that may detect diabetic retinopathy. And I’ll simply caveat this by saying that I’m not an knowledgeable on this system per se, however it’s one of many few FDA authorised machines that may take an image of your retina and decide whether or not there’s diabetic retinopathy that must be referred to the next degree of ophthalmic care. There’s not quite a lot of autonomous AI medical instruments on the market which can be actually in use or FDA authorised. So I’d say that, on this space, we’re actually main the best way.
What about by way of decrease degree instruments and strategies? Like, on the analysis degree, are there issues that you’ve got seen in different fields that you just suppose may allow you to enhance the extent of care you possibly can present or the way you do your work?
One of many areas that I want to carry extra to ophthalmology really is the honest AI analysis and strategies. This has been crucial within the pc science area basically. It has come to well being care and AI as nicely, and it is solely actually simply beginning to be a difficulty that we as ophthalmologists are additionally occupied with for our AI algorithms. That entails taking a look at how our algorithms are performing in numerous teams of populations and ensuring that we wouldn’t be harming teams of individuals if we had been to deploy our AI algorithms, That is one thing that I have been very impressed by and wish to carry extra consciousness to in our ophthalmology area.
That is a great segue as a result of I wished to ask, what what are you trying ahead to probably the most? What upcoming developments in AI that you’ve got been monitoring do you suppose are going to have probably the most impression on the way you do your work?
In fact it is laborious to speak about what we’re trying ahead to most, or what, what we’re most enthusiastic about, with out speaking in regards to the newest era of enormous language fashions, proper? Once I wished to do pure language processing initially as a trainee—that is most likely 2015 or one thing like that, I began occupied with this—we had BERT fashions. These had been type of the earlier era of pure language processing sort of fashions, which had been orders of magnitude smaller than the present ones which have now entered the general public discourse, to illustrate, or develop into obtainable. So after we had been doing issues like attempting to see if we may extract sure eye examination parts from medical free textual content notes, we had been utilizing all of these type of smaller fashions, they usually had been doing okay however not excellent efficiency.
However now, with the appearance of those trendy massive language fashions, there’s a lot pleasure about what extra info we would have the ability to glean from our free textual content medical notes that may very well be useful for locating the correct sufferers for research or simply understanding completely different options of the sufferers. So yeah, that is actually an enormous space of pleasure.
This can be a little bit of a much bigger image query, however why does this work matter to you?
We spend a lot of our time producing these digital well being data. I believe that there is a lot perception about illness processes and outcomes that may be gleaned from all of this knowledge that’s simply being generated routinely day by day. And I believe that untapped potential actually excites me and motivates me to search out methods to make use of that info to assist make higher choices for our sufferers.
Sophia, this has been actually fascinating. Thanks a lot for becoming a member of us.
Thanks for having me, and all superb dialogue questions. This has been actually enjoyable.
I recognize it.
~ ~ ~
The factor about my bins and packing containers and closet organizers is that they’re finite. They will solely match a lot. Whereas on-line, it’s the seeming limitlessness of our digital lives that makes them so alluring—but in addition so chaotic on the similar time.
Once I work, I attempt my finest to go away breadcrumbs for myself. Little reminders of what’s saved the place, hoping it’s stuff I’ll ultimately want—that one, elusive perception, or the lacking ingredient that turns a difficult challenge round. However I do know I’m simply swimming towards the present. No stock can assist me right here. The info’s flowing too quick to maintain up.
In need of collectively deciding to return to the period of the floppy disk and dwell prefer it’s 1986—an incredible decade for music, thoughts you—it appears unlikely that we’ll ever have the ability to throw our arms round all of our apps and tabs and far-flung recordsdata ever once more.
And so that you’re going to see much more individuals like Sophia, who’re considering of latest methods, new AI-powered instruments, that make higher use of all the precious info that’s hidden away in our infinite digital containers, out of sight, out of thoughts. That makes all these insights as straightforward to seize as a field of tenting gear on a shelf.
If you wish to be taught extra about Sophia and her work, you’ll find some hyperlinks within the present notes.
Working Smarter is delivered to you by Dropbox. We make AI-powered instruments that assist information staff get issues completed irrespective of how or the place they work.
You may take heed to extra episodes on Apple Podcasts, YouTube Music, Spotify, or wherever you get your podcasts. And you may also discover extra interviews on our web site, workingsmarter.ai
This present wouldn’t be doable with out the gifted group at Cosmic Normal:
Our producers Samiah Adams and Aja Simpson, our technical director Jacob Winik, and our government producer Eliza Smith.
At Dropbox, particular due to Benjy Baptiste for manufacturing help and our illustrators Fanny Luor and Justin Tran.
Our theme track was created by Doug Stuart.
And I’m your host, Matthew Braga. Thanks for listening.
~ ~ ~
This transcript has been flippantly edited for readability.
