Tag: Machine Learning

The Big Difference Between This AI Craze and the Dot Com Bubble in the 90s

Photo Credit: Robynne O

“Tech startups are rushing to build the next game-changing innovation that will ‘change the world;’ while financiers are also rushing to fund each one in the hopes that they will own a stake in whatever world-leading companies emerge.”

This could just as easily describe the AI boom over the last few years and the Dot Com Bubble in the late 1990s. Both have a few major features in common: new technology produced a craze. In the Dot Com Bubble that was the internet, and now it is all the new AI technology. In the 90s, the hype caused people to overvalue tech companies, funding companies that never had a chance in the first place. This created a stock market bubble that eventually burst causing many overinflated companies to go belly up. The same will likely say the same thing about the current explosion in AI investment.

At the same time, the internet did significantly change the world. Among other things, it created new companies, including the corporate giants like Google and Facebook. Thus, investment, even significant investment, did make sense. The overinvestment may have created a bubble with many losers, but there were still many financial winners as well. The same likely could be said about this AI craze.

There is one very important difference, however. A difference that could significantly worsen how new AI technology develops and becomes implemented in our society:

The internet was created by a decade-long series of government-corporate sponsorships with the overall goal to decentralize communications (this book https://www.penguinrandomhouse.com/books/534709/the-code-by-margaret-omara/ is a fascinating detailed discussion of that history), but the recent AI technology has been mostly created by a handful of highly centralized large corporations, most often with the goal of creating a centralized “super-AI” or suite of AI products. This difference may prove crucial to how these technologies develop.

The internet was created to be decentralized, and decentralization became ingrained into the ethos of the technology. No single government or corporation uniquely controls what gets put onto the internet, and even as major companies/websites like Google, Facebook, and Amazon now dominate the internet ecosystem, they do not actively control or prevent people from contributing to the vast world that we call the internet. There are a lot of details about how to best implement this decentralization, as anyone who has sat through net neutrality or internet content debates knows, but the internet overall became a sprawling web that anyone could help create.

The latest AI systems, in contrast, tend to be created by one large company seemingly with the goal of becoming the AI platform everyone uses. Large portions of current funding has gone into creating massive, super-AIs to be able to be the first company to create the supposedly massively intelligent AI so that they can own and thus profit from whatever it can do. This goal encourages consolidation.

Now, AI technologies like large-language models have a certain amount of centralization almost baked into their development. For example, it takes a lot of people and resources to scour the internet and create a large-language model, and more still to run and maintain such a large computational system everyday. Thus this industry may tend towards a certain amount of consolidation, but that doesn’t seem to explain all of the centralization.

The internet also takes a tremendous amount of resources to be able to allow billions of people to be able to communicate on multiple devices simultaneously. Other similar communication systems (such as the system of phone lines connecting everyone by phone) historically tended to create natural monopolies because of how much resources it took to build a phone network system, and inherently, the internet is not that different. Yet the usage of the internet became different. We don’t see a supply side monopoly in the same way. (Overtime, monopsony has arisen, where a few social media companies disproportionately leverage their widespread usage to their advantage, but that is different.) The communications themselves became something everyone had access to.

This centralization seems to have had a tremendous impact on how the current AI technology has developed and will continue to develop. AI seems focused on the interests of big (mostly tech) corporations. One of these companies’ main goals seems to be to automate routine human tasks. This can have an appeal with regular people, but this is really a priority of “business management” interested in saving money by increasing efficiency. Further, many companies are using the hype around this “AI revolution” as a face-saving way to reduce headcount anyways. Finally, the AI interfaces are built to suck users in, to make them want to use it more and more, in ways pretty common on most apps and major websites today. This too tends to be a business management interest.

A new major technology, like say this current generation of AI technologies, usually provides different people with varying amounts of opportunities, helping some and hurting others. Who it helps or hurts is often rooted in who designed it and what their interests/goals lead to it being designed in the first place.

For example, it’s not surprising that agentic coding platforms like Claude Code are really useful for software engineers. It seems to transform their space. Originally, most of the software engineer’s time was spent on busywork instead of the “cool” parts of programming, and Claude Code more or less automates their busywork, allowing them to focus on the cool stuff. Software engineers and other techies were the main people who created Claude Code, so of course, it’s useful for that space.

Many in other professions seem noticeably more cautious about the introduction of AI into their fields. For artists, for example, AI image generation tools seem to allow others to make their art without them. It takes away some of the most interesting parts of making art. Instead, the tools process their past art and allow others to make it without them. Here the process was designed for regular people in mind who do not have the ability to create their own art and noticed how it benefits them over artists.

Thus, we should think more about who is involved in building the AI technology, because that will influence what kinds of problems it gets built to solve. Right now, it seems to favor the interests of business elites and large corporations. I am not necessarily trying to morally condemn these people and corporations (maybe some are good people, maybe some are bad), but I want to point out how slanted their interests involved in building this technology have been and suggest we may need to broaden who “leads the charge” in the development of AI.

So what? What can we do about this? The interests with which a technology is first developed sets a course for how that technology is conceived and what kinds of things people imagine it doing, but nothing is ever completely set in stone. Technologies often grow and evolve over history, becoming used for purposes very different from their original conception. But obviously, the sooner one does this the better. Given that the technology is still young, this is the best time to incorporate other forms of thinking. How it develops now will have cascading impacts on what AI looks like in the future.

How can we build AI in different ways and move beyond the specific interests that hog the AI field now? Exactly how, I don’t know. It’s worth thinking about. I wrote this article to help you think about the problem and get your juices going. Maybe together we can develop a way to build AI in a way that serves the interests of regular people.

Is data science still the sexiest job?

Photo Credit: Mahdis Mousavi

In 2012, this Harvard Business Review article argued that data science will be the sexiest job in the 21st century. At the time, data science was new and unheard of, with companies eager to use data scientists to revolutionize their practices. Is it still the sexiest job now? Well sort of, but not really. The field has gone through some significant transformations since these “wild west” early days. Now, data science as a discipline has become more streamlined and specialized.

Often key data scientists have slightly different titles like machine learning specialist, data engineers, etc. Machine learning and AI technology have changed the way data is processed and analyzed. This has automated parts of the tasks that data scientists have spent a long time working on, such as data cleaning (which still can take a long time) and initial data exploration, shifting the work necessary for humans to perform in the field to more specialized and fringe tasks. For example, many data scientists have become machine learning specialists focusing on fine-tuning these models or communication specialists focusing on how to use their business expertise to communicate complex findings with stakeholders and help decide what they should do about the results.

I think more than the technology, what has driven the specialization is the routinization of data science processes within an organization. Gone are the days of a lone data scientist at a company doing cutting edge work by themselves just figuring out what is possible. Data science as a field has fallen within the discipline and expectations of corporate bureaucracies. In its early stages, most data scientists worked alone or in small teams doing pioneering, experimental work figuring out how to apply the tools of the field to their organization in ways that people did not know were possible. That can still be the case. In every job I have had as a data scientist, for example, I have been the first data scientist in the entire organization or specific department I work in. But this is increasingly rare. Data science is now mostly one department at an organization, doing important but predictable routine work. All white collar professions get grafted in the “corporate machine” like this overtime.

Recent AI technology has contributed to this too by automating many of the low-level data processing and analyzing tasks so that non-specialists can perform them on their own. This is great, increasing the accessibility of tasks once considered obscure or even “magical” by regular people. Back in the day, to do much of any data modeling, you had to code it yourself, requiring a level of programming knowledge that was beyond a typical office worker or manager. That’s why they needed to hire a data scientist to analyze the data themselves. I hope in the long run using AI tools to tinker with data themselves and try out different theories will increase the data literacy and skillsets of regular professionals. It also means that data scientists are increasingly spending less time on these tasks and have moved to more complex, specialized work that still require quite a bit of technical human thinking.

Another factor that has driven this routinization is the increase in the number of people studying and doing data science. As demand for data science increase, more people have tried to become a data scientist, whether by receiving a degree in it or transitioning their careers into the field. This has led to more data scientists in the market. If this trend continues, eventually the field will become oversaturated, but the demand still seems to be higher than the supply, with more open jobs than people able to fill them.

This has still redefined what data science is. When many people join a field, it becomes difficult to maintain the same level of pioneering eclecticism. Instead, the types of tasks people do become routinized and standardized to provide consistency for a larger number of people, paralleling the transformation Max Weber describes religious movements undergoing from a charismatic leader to a routine social institution.

All of this leads to the current state of data science. This is not necessarily bad, but it is different. So, is data science still the sexiest job? Yes and no. Some of its specialist roles like machine learning specialist, I think, better maintain the excitement and cutting edge of that moniker. It’s still in high-demand, however, a fine field to work in.

The Question-Driven Data Scientist: Why Social Science is Key in the AI Era (Conversation with Eesha Iyer)

In my conversation, Eesha Iyer, an economist-data scientist, discusses how machine learning and artificial intelligence have changed what is possible. We are seeing a transition both from static inferential models common in economics for decades to dynamic, interactive systems that adjust in real-time.

We are also seeing a revamping of the workflow with AI systems clearing up time to do rudimentary programming tasks. Trivial programming tasks that once took quite a bit of a data scientist’s time are easier than ever, so now the key issue is becoming, What kinds of questions should we ask of the data? Qualitative and social science thinking are crucial for this new space. For Eesha, gone are the days when data scientists were technical workers spending hours writing code. In the current era, the question becomes how to formulate relevant research avenues to explore. For this, social scientists are more useful than ever.

In our conversation, we explore the implications all this has on the field of data science. She also advises how to learn data science in this shifting landscape. I hope you enjoy.

Conversing with AI: Interview with Chelsea Wang about Communications with Artificial Intelligence Systems (Part 3 of 3)

In the final part of our conversation, Chelsea Wang explains how her background in psychology has influenced her work in artificial intelligence. In particular, she describes how her social science background helped her develop and deploy her own version of the Mutual Theory of Mind as a psychologist within the field of artificial intelligence. When socializing, humans employ a recursive feedback loop of conceptualization of each other, and she explores the application of similar concepts to conversational AI systems.

She concludes by discussing her journey as a PhD student: what led her to seek her dissertation and her plans afterwards to use what she is learning now to conduct innovative and impactful work in the business world.

Click here to learn more about the Interview Series.

More about Chelsea:

Qiaosi Wang (Chelsea) is a fifth-year PhD candidate in Human-Centered Computing at Georgia Institute of Technology. Chelsea is a human-centered AI researcher and her PhD dissertation work focuses on building the Mutual Theory of Mind framework, inspired by the basic human capability to surmise what is happening in others’ minds (also known as “Theory of Mind”), to enhance mutual understanding between humans and AIs during human-AI communication. Her work specifically focuses on the human-AI communication process during AI-mediated social interaction in online learning, where AI agents can connect socially isolated online learners by providing personalized social recommendations to online learners based on information extracted from students’ posts on the online class discussion forums.

Chelsea received her Bachelor of Science degrees in Psychology and Informatics from the University of Washington, Seattle. In her free time, Chelsea loves hiking, playing with her cat, Gouda, and spending time at bouldering gyms. 

To learn more about Chelsea and the sources we referenced in our conversation:

Conversing with AI: Interview with Chelsea Wang about Communications with Artificial Intelligence Systems (Part 2 of 3)

Chelsea Wang has spent many years trying to improve the cognitive process of artificial intelligence systems to better interact with humans. In this second part of our conversation, she explains her theories about metacognition, intelligence, and potential anthropomorphization of AI “thought” processes. Through this, she explicates her vision and approach to the potential social life of AI.

Click here to learn more about the Interview Series.

More about Chelsea:

Qiaosi Wang (Chelsea) is a fifth-year PhD candidate in Human-Centered Computing at Georgia Institute of Technology. Chelsea is a human-centered AI researcher and her PhD dissertation work focuses on building the Mutual Theory of Mind framework, inspired by the basic human capability to surmise what is happening in others’ minds (also known as “Theory of Mind”), to enhance mutual understanding between humans and AIs during human-AI communication. Her work specifically focuses on the human-AI communication process during AI-mediated social interaction in online learning, where AI agents can connect socially isolated online learners by providing personalized social recommendations to online learners based on information extracted from students’ posts on the online class discussion forums.

Chelsea received her Bachelor of Science degrees in Psychology and Informatics from the University of Washington, Seattle. In her free time, Chelsea loves hiking, playing with her cat, Gouda, and spending time at bouldering gyms. 

To learn more about Chelsea and the sources we referenced in our conversation:

Conversing with AI: Interview with Chelsea Wang about Communications with Artificial Intelligence Systems (Part 1 of 3)

Chelsea Wang describes her work developing and refining the communication processes between artificial intelligence and humans, particularly the Mutual Theory of Mind framework she has helped build. As a doctoral student in Human-Computer Interaction, she also discusses her journey from human psychology to the social interactions of AI.

Click here to learn more about the Interview Series.

More about Chelsea:

Qiaosi Wang (Chelsea) is a fifth-year PhD candidate in Human-Centered Computing at Georgia Institute of Technology. Chelsea is a human-centered AI researcher and her PhD dissertation work focuses on building the Mutual Theory of Mind framework, inspired by the basic human capability to surmise what is happening in others’ minds (also known as “Theory of Mind”), to enhance mutual understanding between humans and AIs during human-AI communication. Her work specifically focuses on the human-AI communication process during AI-mediated social interaction in online learning, where AI agents can connect socially isolated online learners by providing personalized social recommendations to online learners based on information extracted from students’ posts on the online class discussion forums.

Chelsea received her Bachelor of Science degrees in Psychology and Informatics from the University of Washington, Seattle. In her free time, Chelsea loves hiking, playing with her cat, Gouda, and spending time at bouldering gyms. 

To learn more about Chelsea and the sources we referenced in our conversation:

Designing Machine Learning Products Anthropologically: Building Relatable Machine Learning

Loader Loading…
EAD Logo Taking too long?

Reload Reload document
| Open Open in new tab

Download [2.38 MB]

How do we build relatable machine learning models that regular people can understand? This is a presentation about how design principles apply to the development of machine learning systems. Too often in data science, machine learning software is not built with regular people who will interact with it in mind.

I argue that in order to make machine learning software relatable, we need to use design thinking to intentionally build in mechanisms for users to form their own mental models of how the machine learning software works. Failing to include theses helps cultivate the common sense that machine learning is a black box for users.

I gave three different versions of this talk at Quant UX Con on June 8th, 2022, the Royal Institute of Anthropology’s annual conference on June 10th, 2022, and Google’s AI + Design Tooling Research Symposium on August 5th, 2022.

I hope you find it interesting and feel free to share any thoughts you might have.

Thank you for the conference and talk organizers for making this happen, and I appreciate all the insightful conversations I had about the role of design thinking in building relatable machine learning.

Data Science and Game Design: Conversation with Clayton Sisson (Part 3 of 3)

During the final part of our conversation, Clayton discusses his journey from game design to data science, including what inspired them to study data science and what it has been like learning and working in this new field. Clayton Sisson is a game designer and aspiring data scientist, passionate about how data science can shed light on human behavior.

This is the next installment in my Interview Series. During Over the course of the three parts of our conversation, we discuss how game design thinking can help develop usable and useful machine learning products within data science.

Here is Part 1 and Part 2 of our interview.

Resources:

Data Science and Game Design: Conversation with Clayton Sisson (Part 2 of 3)

In Part 2, we discuss how to apply the design concept shikake to machine learning systems. Clayton Sisson is a game designer and aspiring data scientist, passionate about how data science can shed light on human behavior.

This is the next installment in my Interview Series. During Over the course of the three parts of our conversation, we discuss how game design thinking can help develop usable and useful machine learning products within data science.

Here is Part 1 and Part 3 of our interview.

Resources:

Data Science and Game Design: Conversation with Clayton Sisson (Part 1 of 3)

Clayton Sisson is a game designer and aspiring data scientist, passionate about how data science can shed light on human behavior. For the next installment of my Interview Series, we discuss ways to use game design and UX design to develop usable and useful machine learning products and their experiences transitioning from design into data science. In this first part, we discuss the connections between data science and game design.

Here is Part 2 and Part 3 of our interview.

Resources: