Generative AI and a New Era for Education Innovation — Lumos Capital Group

When we started Lumos Capital Group a few years ago, we referred to artificial intelligence as the “apex technology of the information age,” poised to redefine every industry.

At the time, AI had already influenced our lives in innumerable ways through the consumer and business technologies we use daily. The future use cases for AI in education were exciting to imagine, but more narrow and discrete in real life. In many ways, we were more focused on the second order effects of AI over time — how it would impact jobs and the future of work, and require greater societal investments in reskilling the workforce.

However, we also knew the edge of innovation in AI was moving quickly. The ability for AI to efficiently assess exams, fix lazy grammar or recommend educational videos on YouTube was going to be just the very tip of the iceberg.

Three years later, the latest developments in AI have been breathtaking, and the impact on education profound. A few months ago we saw one application — ChatGPT by OpenAI — that has been adopted at unprecedented speed, drawing over one million users within 5 days and 100 million users within two months, faster than any other technology or platform in history. Since then, new breakthroughs, tools and use cases such as Auto-GPT have appeared with breathtaking speed. Bill Gates recently wrote, “The development of AI is as fundamental as the creation of the microprocessor, the personal computer, the Internet, and the mobile phone.”

If 2012 was “The Year of the MOOC” — as New York Times dubbed it then — 2022 was undoubtedly “The Year of Generative AI.” Like Netscape’s web browser ushering in the age of the internet almost 30 years ago, we believe generative artificial intelligence represents the vanguard for a set of technologies that will impact the world of education in ways that have never been seen before.

Generative AI — A New Epoch in Technology

Generative AI uses particular data sets and large language models to make predictions and generate novel content. The difference relative to analytical AI is the ability to create.

We have seen 1.0 versions of this technology in smart compose email and smart search. But the massive increases in model sizes, combined with the deep learning models (transformers, the T in ChatGPT) that made AI self-supervised learning possible, have dramatically improved previous limitations in areas like natural language processing and image processing.

These transformer neural networks address one of the biggest constraints in the development of AI — the availability and quality of large, structured datasets to train models. As a result, according to the CEO of NVIDIA Jensen Huang, “AI has jumped to warp speed.”

The creative abilities of generative AI is not limited to text, but also encompasses code, images, audio, video, and excel formulas. As such, everything from writing, coding, producing music — or even the province of investment bankers and private equity analysts, financial modeling — will be disrupted if not transformed in the years to come by generative AI.

In 2022, besides ChatGPT, applications such as DALL-E, MidJourney, Stable Diffusion and others utilizing generative AI exploded into the popular consciousness, in large part due to the remarkable quality and speed of what they were able to produce, relative to prior generations of AI.

The rate of improvement in generative AI over the last twelve months has become one of the most notable aspects of the story. OpenAI released ChatGPT-4 only months after ChatGPT-3.5, and the differences in what the application was able to accomplish were astounding. With 170 trillion parameters compared to the prior model’s 175 billion parameters, GPT-4 performed at the 90th percentile on a simulated bar exam, for example, compared to only the 10th percentile for GPT-3.5. While GPT-3.5 was generally poorer at STEM related prompts and prone to “hallucinations” (outputs that are deviant from normal or accurate) in certain contexts, GPT-4 was already up to 30% better at avoiding such errors.

*The magnitude of ChatGPT-4’s improvement of 3.5 in a few months has been astounding*

This acceleration in what generative AI can produce has alarmed even the AI experts. An open letter signed by over 1,000 AI experts and researchers in March of 2023 states, “Recent months have seen AI labs locked in an out-of-control race to develop and deploy ever more powerful digital minds that no one – not even their creators – can understand, predict, or reliably control.” The letter issued a call for an immediate moratorium on the creation of systems more powerful than GPT-4 for at least six months, so the capabilities and dangers of such systems can be properly studied and mitigated.

But with the incentives at play, there is no pausing or even slowing down. Recently, Google — which has been in crisis mode since the launch of ChatGPT and Microsoft’s adroit positioning around it — gave its engineers a code red, issuing a mandate for them to integrate generative AI technologies into all of its major existing products and services. As of this writing, Google’s rival chatbot, Bard, is still waitlist-only. Meta announced in late February that it now has a team dedicated to building tools powered by AI. Chinese tech giant Baidu launched its own generative AI chatbot, Ernie Bot, in March.

Meanwhile, OpenAI announced that it is now letting 3rd party developers integrate ChatGPT into their apps and services via an API, while also rolling out various plugins, extending its lead in the marketplace. Already, one-third of Y Combinator’s latest cohort of startups are planning to use this or competitive generative AI services in their applications.

With both the speed of adoption and the rate of technological improvement, there is no doubt that the impact of generative AI on the education sector will be vast. As Gates noted in his latest post in March of 2023, The Age of AI Has Begun: “It appears that AI will impact the future of learning and the future of work more quickly and more profoundly than many of us had been expecting.”

Panic in the Ivory Tower

Understandably, many educators have reacted to ChatGPT and its peers with pessimism and fear. The Atlantic declared, “Nobody is prepared for how AI will transform academia.” Noting that the essay has been the center of humanistic pedagogy for generations — how we teach children to research, think and write — the writer concluded that the “entire tradition is about to be disrupted from the ground up.” In a recent Inside Higher Ed piece, Professor Jeremy Weissman called ChatGPT a “plague upon education,” analogizing it to COVID, and saying that this plague “threatens our minds more than our bodies.” Alex Lawrence, professor at Weber State University, described the technology as “the greatest cheating tool ever invented,” as reported by the Wall Street Journal.

This reaction was not unreasonable. According to survey data from Study.com of 1000 students aged 18 and up in January 2023, only 2 months after the release of ChatGPT to the public, more than 89% of students who responded said they had used ChatGPT to help with a homework assignment, 48% admitted they had used it for an at-home test, and 53% had already used it to write an essay.

Professors and teachers also wonder how this technology will affect their jobs. An analysis of professions “most exposed” to the latest advances in large language models like ChatGPT found that eight of the top 10 are teaching positions.

One immediate reaction was to attempt to block or ban the technology. In December 2022, the Los Angeles Unified School District “preemptively” blocked access to ChatGPT (on their own WiFi) while a “risk/benefit assessment is conducted,” a district spokesperson told the Washington Post. And in January 2023, New York City Public Schools banned access to ChatGPT from devices and networks that the school owns. A spokesperson for the NYC Department of Education said that the decision was made “due to concerns about negative impacts on student learning and concerns regarding the safety and accuracy of content.”

Already, the race is on to create technologies to detect AI-generated content. The leader in plagiarism detection, Turnitin, announced that their new AI writing detection tool would be released in April 2023. Turnitin has claimed that they are able to detect AI-generated writing because it is “extremely average.” A Princeton computer science student created GPTZero in January 2023 to detect AI text, and already claims over a million users. Yet many technologists have already predicted that as quickly as Turnitin and others have released AI to detect AI, there will be yet another wave of AI that will specialize in avoiding AI-based detection, ad infinitum.

Implications for Teaching and Learning

Beyond the practical impracticalities (and frankly impossibility) of restricting students’ access to technologies such as ChatGPT, some educators have observed that the ability to now automate many aspects of writing actually presents an opportunity to rethink how we teach and learn. Pointing to the calculator in a prior era, many believe that educators will eventually have to incorporate the reality of students’ access to generative AI into their approach to the classrooms, which will benefit the education process in the long run.

Indeed, because of the outsized influence of high-stakes, standardized writing assessments like the ACT/SAT and AP exams, many teachers today have aligned their writing instruction to simply mimic the model output of the 40-minute writing exam, reducing the art of writing to generic, five-paragraph essays — and effectively teaching students to write using an algorithm.

Granted, there are differences between a tool such as a calculator in math and a tool such as generative AI chatbots like ChatGPT. For calculators, the mechanical operations employed mimic the labor of the student, more accurately and infinitely faster; this may in fact free up students to practice the mathematical thinking that solving a complex or multi-dimensional math problem requires. By contrast, the labor of writing is the process of thinking. Frequently we only know what we think — or discover how constipated our thinking has been — after we go through the difficult process of putting pen to paper.

But if computers can now make the standard 5-paragraph essay obsolete, what kind of writing — and what kind of teaching and learning about writing — should teachers and students engage in? Is it more personal, creative writing, helping students develop their unique voice? Is it more deeply researched and planned writing, writing that encourages students to take more nuanced positions than what a generic web-scraping technology can produce (and which merely conforms to standard conventions while avoiding all controversial positions)?

Beyond writing, the world of generative AI holds enormous promise to advance teaching and learning. For example, AI technology can help teachers create personalized learning experiences for students with much greater ease. With more advanced AI tools than we have today, teachers can analyze student data and tailor their teaching to the specific needs of each student with content that caters to students’ learning styles.

Imagine tools that can help a teacher create 2-minute explainer videos in a matter of seconds on the concepts of speed, distance and time — but with 25 individual versions that are personalized to each student. One video might use a basketball example while another presents outer space travel or a ballet dance floor as the context; one video might be narrated by a famous historical figure, while another utilizes anime, with each designed to maximize the individual student’s engagement and therefore learning.

Even before the release of ChatGPT, relative to 2019, education institutions were already reporting 15-25% greater usage of AI across a range of core academic functions. The below data from HolonIQ summarizes institutions’ responses as of September 2022:

Routine tasks in the classroom, from generating quizzes and learning materials, to grading assignments and providing feedback on student work, can be elevated and automated with generative AI. AI tools can also analyze student performance data and provide teachers with insights into how students are learning. Many of these functions today are already enhanced in some ways with software and digital content that incorporate basic analytical AI, but the possibilities for dramatic improvement relative to today’s technologies are enormous.

In higher education and workforce development, we believe that generative AI will push colleges and universities to enhance their curriculums, and make their educational experiences more interactive, higher quality and perhaps even more affordable. And because generative AI is likely to impact all knowledge fields and knowledge workers, it will drive a whole new wave of reskilling imperatives, both job training programs funded by government as well as reskilling initiatives supported by businesses. For the individual, and especially the individual in fields where their core skill sets are being automated away, AI will push every person to keep learning in order to stay relevant.

Implications for Investors and Companies in the Sector

From a market perspective, we believe that generative AI will unleash a new wave of innovation in the sector, and create enormous value for the companies that are able to utilize these technologies effectively to better serve students, parents, teachers and schools. At the same time, many existing companies will have their business models threatened.

Already, many prominent edtech companies and education nonprofits have begun to incorporate integrations with ChatGPT. Quizlet announced Q-Chat, which they dubbed “the world’s first AI tutor built with OpenAI’s ChatGPT.” Duolingo, the language learning app, introduced a new subscriber tier, Duolingo Max, with GPT-4 integrations to power features such as “Explain My Answer” and “Roleplay.” Even Khan Academy has begun to use GPT-4 in an AI-powered tutoring platform called Khanmigo. Numerous startups in edtech are also scrambling to catch the latest generative AI wave.

Nonetheless, any company that depends on a Google search to reach its audience and drive its top of funnel – for example, direct-to-student companies that are trying to use the homework help pain point to get a learner to its landing pages and subscribe to its services – will be hurt in the short term by the rapid adoption of technologies such as ChatGPT. For any college student today it is simpler, cheaper and more convenient now to just ask a question to a smart chatbot.

We are entering a world where machines are increasingly able to produce quality content at scale — videos, articles, art. Lowering the barriers to entry for content creation by decreasing costs and overall friction will greatly increase the velocity at which content will be created. This will have a dramatic impact on both creators and platforms, such as publishers, course authors and global content marketplaces. Historically, education publishing was one of the segments of the global education market where a lot of market capitalization was concentrated, in part because the expense of creating content such as educational textbooks required scale. As a result, these segments of the market in both K-12 and higher education generally produced oligopolistic market structures (for example, for decades just three publishers controlled over three quarters of the K-12 publishing market).

Until generative AI is able to produce content that passes rigorous academic standards on a consistent basis, there will still be a role for trusted institutions; indeed this role may be amplified in a world of content abundance with dubious quality. Indeed, the well-known limitations of AI with issues of bias and unknowable data inputs and outputs (the “black box problem”) has generated a deep skepticism by many members of the education community.

Yet over time, especially with the personalization capabilities of AI, the market is evolving towards the democratization of content creation. This will be a trend that Lumos will be following closely. Similarly, because of the increasing capabilities of AI to interact intelligently with students, tutoring companies are racing to see who can first solve Bloom’s famous “2 sigma problem,” and design an AI tutor that can truly deliver personalized tutoring at scale.

It remains to be seen whether OpenAI, Google or another company will create the underlying platform or large language models used by companies. The challenge many edtech AI companies face is how to build differentiation and a lasting competitive moat if they do not own the underlying model. Most likely, much of the value for this generation of companies will come from owning and deploying unique, proprietary data sets that can produce ever more “fine-tuned” recommendations. On that front, some of the most scaled edtech companies today that already collect billions of learner data points will have a significant head start.

Regardless, in the near-to-medium term, having AI capabilities built into an educational product or service will become table stakes for competitiveness. Value will migrate once again from one part of the market to another, away from incumbents who stand still and toward the innovators who can deliver more accessible, higher quality and more engaging educational experiences.

The Most Valuable Company in the World?

The world of generative AI presents as many challenges as opportunities. There is much we do not understand about the likely path of future technology development. There are even calls today from credible sources for regulation that would ban AI completely; indeed, as of this writing Italy has completely banned ChatGPT, citing privacy concerns. Nonetheless, we believe the most likely outcome over time is that the market utilizes these new advances in AI technology to reshape the future of learning and work in exciting ways. Companies that utilize this technology well will gain competitive advantage; entire market segments may undergo creative destruction over time. As an investor in this sector, it is an extraordinary time.

In 2016, futurist Thomas Frey predicted that by 2030, the largest internet company in the world will not be Google or Facebook, but an education company that no one has heard of.

At the time, this was certainly a contrarian claim. Up to that point the world had seen education companies from a prior generation peak at $20 billion in market capitalization, and even the most successful edtech software and consumer subscription companies reach only high single digit billions in value. The notion that an online school could be the largest and most important company in the world seemed far-fetched.

To do so, such a company would not only have to deliver the dramatic transformative impact that we know the best educational experiences and most wonderful teachers can have — but be able to scale in a way that no education company has been able to scale.

With the new developments in technology and artificial intelligence in 2022, this prediction appears much less implausible today than ever before.