”Learning is not a one-size-fits-all process; it should be tailored to each individual's needs” - Sal Khan
Throughout 400+ interviews with AI founders we’ve conducted at Edtech Insiders, one theme comes up more than any other: in the future, learning will be personalized. We in Edtech have a long history of dreaming about how tech can enable personalized learning, and many related ideas that follow: individualized learning, adaptive learning, customized learning pathways, and good old classroom differentiation to name a few. Much of this history has been disappointing, often due to our inability to customize materials and learning experiences at scale. This is why AI feels like such an exciting innovation for the industry.
However, personalization comes in multiple flavors. As AI models progress at incredible speeds, they are ever more capable of acting like an intelligent, informed educatorthat can respond to any query or prompt with nuance and acumen. Because these queries or prompts come directly from users, they contain tacit or explicit information about the users’ desires, preferences, interests, and personality. In other words, every prompt is personal, which nudges an LLM to craft a personalized answer.
Here’s how a LLM responds to a question that begins to call out the user’s identity and affect:
Q: “I’m a high school junior in America and I’m learning about the unification of Italy. Why should I care?”
A: “The unification of Italy isn’t just about dusty old wars; it’s about how people with different cultures, languages, and beliefs figured out how to become one nation—which is exactly what’s happening now online, in your world, as cultures and ideas clash and combine on TikTok, in group chats, or even when different friend groups mix. Learning how power, identity, and connection worked in the past helps you recognize those same patterns in your life today—so you can understand them, question them, and shape them.”
That’s… fine. Probably as good an answer as many educators could give if they were challenged by a question like this in a live classroom. But surely this answer would be a lot better if the AI knew (or remembered) more about the user?
And of course, that’s the bigger part of what we mean by personalization. If a LLM is provided with a set of characteristics, interests, and motivations (in this case, it was relatively generic information about modern 16 year olds), the answer looks more like this.
A: The unification of Italy might seem random, but it’s actually a story about people with totally different styles, goals, and histories figuring out how to work together without losing who they are—like if Olivia Rodrigo, Sabrina Carpenter, and Tyler, The Creator had to form one band.
It’s full of drama, loyalty shifts, and power struggles, just like the worlds you move through every day: balancing different friend groups, posting your story vs. keeping things private, deciding who you want to be and what matters to you.
One of the leaders, Garibaldi, gave up control of the land he conquered—even though he could’ve kept the spotlight—because he believed a united Italy was more important than his personal fame.
History isn’t about memorizing dates—it’s about understanding how people make hard choices in messy situations, and learning something real about yourself in the process.
Definitely better, if still a bit cringey and overly sincere. And long. Ask the LLM to be a little snarkier and you get this:
A: Congrats, you’ve unlocked Unification of Italy, aka the historical equivalent of a group project where no one texts back but everyone still wants credit. This was a whole bunch of messy men with main-character syndrome (Garibaldi was basically giving chaotic good energy) trying to squish a bunch of regions into one country. Backstabbing, weird alliances, major drama. Honestly? It’s giving friend group fallout, content creator beef, and "who are you even trying to be right now?" vibes. If this happened today, Emma would’ve podcasted it and Sabrina would've dropped a song. So yeah—this isn’t boring history. It’s just vintage drama.
It would be very informative to put these different answers in front of actual 16-year olds and see what resonates. However, the point is that LLMs are highly capable of changing their tune and style in response to what they know about the asker. So what can we do with that?
Early Personalization in LLMs
One of the deep ironies of LLMs is that they know almost everything, but remember almost nothing. You can hold an hour long conversation with Claude in one chat, and in the next chat, it’s like you were perfect strangers.
In the 2000 film Memento, the hero, whose short-term memory is fully erased, must tattoo important information on his body to pass it to his future self, often in cryptic formats (e.g., “Never Answer the Phone”, “Don’t Trust Your Weakness”). This is a fun (?) metaphor for how LLMs work: they must try to encapsulate important information from one chat and pass it to future chats in a condensed format that can be used in conjunction with the LLM’s extensive existing training.
You may or may not have noticed that a personalization menu has quietly appeared inside ChatGPT over the last few months, which allows you to tailor your ChatGPT experience across sessions by:
Giving ChatGPT a nickname to refer to you by
Tell it a bit about yourself (using the classic cocktail party question “What do you do?”)
Assigning ChatGPT some personality traits (e.g., “Direct”, “Chill”, “Opinionated”, “Sarcastic”)
Giving it additional info of any kind (“Anything else ChatGPT should know about you?”)
Access the “saved memories” it has about you (mobile only)
Tell ChatGPT whether you want it to reference your chat history (mobile only)
Erasing the saved memories (mobile only)
The ChatGPT Personalization feature on web (left) and mobile (right), along with some of the dozens of ‘saved memories’ it has about me
If you use the ChatGPT mobile app, I strongly recommend checking out the saved memories it has stored about you. Like the tattoos in Memento, some will be very clear and useful, others cryptic, and some just plain wrong: incorrect inferences from questions you’ve asked in the past. There is some logic going on in the background here to identify aspects of conversations that are worth remembering across sessions.
Google Gemini also offers levels of personalization: you can go to “Saved Info” in the Gemini app and tell it things about yourself, but with even less guidance than ChatGPT:
My product manager spidey sense tells me that these are incredibly powerful features that are probably being tragically underused… by design. They’re hidden in small menus that don’t draw much attention to them. There is not much to showcase the value of these tools, but more importantly, it puts the onus on the user to do much of the work of figuring out what information to tell the LLMs that would actually add value to your life.
It may go without saying that this is clearly not designed for edtech or education purposes: asking a high school student, or even most full-time college students, “What do you do?” is a sure way to get them to widen their eyes and stop talking (or in this case, close out of the menu).
Luckily, we’re not entirely reliant on general purpose LLMs to drive personalization. There are plenty of other startups that are also working to tackle the consistent memory problem. AI therapy apps, for example, wouldn’t be much good if they started every session with, “Nice to meet you. What brings you in today?”.
For edtech, the key is understanding what information we might want an LLM to know and remember about users as learners.Certainly, many AI tutors like Khanmigo and learning companions like Tolans must remember information across sessions and innovative companies, from incumbents like Powerschool and Instructure to research-backed smaller players like Mindprint Learning and Oluko Learning have been working on cohesive and comprehensive student profiles.
What does an AI have to know about a student to create meaningful and efficacious learning experiences? That’s a very deep and interesting question for the whole field. Here’s a humble suggestion about where we might start.
What’s Missing? The PEARL.
Imagine a Portable Learning Model… that is, a framework for how to understand a user as a learner that is contained outside of any single chat or learning tool. Let’s call this a learner’s PEARL: a Portable, Evidence-Based, Adaptive Representation of a Learner (credit to a long chat with Claude for this fun acronym).
A learner’s “PEARL” might include categories like their:
Cognitive Habits: Preferred feedback style, memory strengths and challenges, response to mistakes
Motivational Triggers: Need for autonomy, level of curiosity, need for relevance, relationship to competition
Learning History: What’s been mastered, what still needs work
Behavioral Signals: Engagement time, switching behavior
Goals and Values: Future goals, cultural values, and identity
Learning Blockers: Sources of academic anxiety, fixed mindset, misconceptions
Interaction Preferences: Content format preferences, preferred interaction mode, preferred time of day
Ideally, these categories would be rooted in learning science and real evidence (although, I imagine they would evolve quickly based on new findings). A learner’s PEARL would be dynamic and could be created and updated based on many different inputs: on-boarding questionnaires, chat histories, inputs from human educators, SIS data, and more.
It also could be tailored to different learning subjects and even learner moods: no consistent learning styles here! It could combine fixed categories with qualitative signifiers. Over time, you can imagine interesting relationships between a learner’s PEARL and their IEPs, but this is a legal and privacy quagmire over the horizon, at least for now.
And, lastly, but critically, just like those Memento tattoos, the pearl would be portable. They’d stick to you wherever you go. In an ideal world, you’d own your learner profile, have control over it and be able to port it not only between chat sessions, but between tools (LLMs, study tools, learning companions) and between educational institutions (school, tutoring, college).
Are you already making a version of this? Have ideas? Reach out to alex@edtechinsiders.org and let us know about it.
Thanks for reading Edtech Insiders! To receive new posts and support our work, consider becoming a free or paid subscriber.
Join Edtech Insiders+
If you love Edtech Insiders and want to continue to support our work, consider becoming a paid subscriber! By joining Edtech Insiders+ you receive:
Early access to all Edtech Insiders events. No more sprinting to sign up for our monthly happy hours, edtech summits, and online conferences.
Access to our exclusive WhatsApp channel to connect with the Edtech Insiders community and discuss the latest trends and news in our space.
Access to exclusive member content, including interviews, panels, videos and articles.
All proceeds from subscriptions will help us invest in our podcast, newsletter, events, and community. Our goal is to make Edtech Insiders an enduring and sustainable community offering that connects the edtech ecosystem!
We would be so excited if you decide to join us as a member of Edtech Insiders+.
This edition of the Edtech Insiders Newsletter is sponsored byTuck Advisors.
Since 2017, Tuck Advisors has been a trusted name in Education M&A. Run by serial entrepreneurs with over 25 years of experience in founding, investing in, and selling companies, we believe you deserve M&A advisors who work as hard as you do. Not all LOIs are the same. In fact they are often apples to oranges to cashews! Use our Deal Value Calculator to level set the real value of your LOI.
We had an AMAZING time at our Co-Design Secrets webinar last week, thanks to all who attended for your engagement and amazing questions! And, thanks again to our incredible panelists:
If you couldn’t catch this one live, be sure to watch the replay. AND, keep your eyes peeled for more upcoming webinars and live events.
Cooley LLP is the go-to law firm for edtech innovators, from early childhood through workforce. Informed by decades of experience in the education vertical, Cooley created the first edtech practice to provide industry-informed, business-minded counsel to companies and organizations at all stages of the corporate lifecycle. Cooley provides a multidisciplinary approach to client needs, offering seamless collaboration across offices and practices.
To learn more about what Cooley can do for you, reach out to Naomi May.
1. Students are AI Optimists, but Women are at Risk of Being Left Behind
A new survey from WGU Labs finds that while the majority of students feel optimistic about AI’s role in education, women are significantly less confident using AI tools than men. This gender gap raises concerns about equitable access and inclusion as AI becomes more embedded in learning environments. The report calls for targeted efforts to improve AI literacy among women and to ensure that student voices help shape the development and deployment of educational AI tools.
2. Inside OpenAI’s Plan to Embed ChatGPT Into College Students’ Lives
A recent New York Times report reveals OpenAI's ambitious strategy to integrate ChatGPT into every facet of college life, providing students with personalized AI "accounts" from day one and enabling chatbots as tutors, study tools, and career coaches. Several large institutions, including Duke, the University of Maryland, and California State University, have already partnered with OpenAI, rolling out its premium "ChatGPT Edu" service with protections like enhanced privacy and faculty-designed bots. But critics warn of potential downsides, like hallucinations, weakening of critical thinking, and loss of human interaction.
In the wake of rising AI-generated plagiarism, some college professors are pushing back by reviving handwritten work, asking students to pen essays and responses in class instead of submitting digital assignments. These in-person handwriting exercises make it more difficult to rely on AI tools and encourage deeper engagement with the material.
Recent legislation under consideration would expand federal Pell Grant eligibility to short-term credential programs—lasting 8 to 15 weeks—even when offered by unaccredited or for-profit providers. While supporters argue this move could broaden access and support workforce training, critics warn that the proposal lacks sufficient oversight, relying on self-reported job-placement and graduation metrics and omitting accreditation requirements.
5. AI Is Still an Unknown Country — and Teens Are Its Pioneers
A new EdSurge report shows teens are exploring AI with curiosity and creativity but little guidance or structure. Most use it casually through tools like search engines, yet many struggle to define what counts as misuse or cheating. The article calls for clearer education policies to help young people navigate AI ethically and confidently.
6. Apple’s New Research Paper Says AI Reasoning Isn’t All It’s Cracked Up to Be
Apple's research team dropped a revealing paper called The Illusion of Thinking just before WWDC 2025, demonstrating that even advanced “reasoning” AI models—like OpenAI’s, Anthropic’s Claude, and DeepSeek—perform well on simple puzzles but falter when complexity increases.
7. US Court Approves Sale of Indian Edtech Startup Byju’s Assets
A U.S. court has approved the sale of Byju’s U.S.-based subsidiaries, Epic and Tynker, as part of the company’s bankruptcy proceedings. Epic was sold to China’s TAL Education Group for $95 million, while Tynker was acquired by CodeHS for $2.2 million. These sales follow Byju’s insolvency filing in Delaware and reflect steep losses from the company’s previous acquisitions.
8. New Course Offers a Guide to GenAI in Higher Ed
Coursera and Lund University launch Transforming Higher Education with GenAI, a free course for educators, technologists, and institutional leaders. The course steers clear of hype, offering practical tools and critical reflection for applying GenAI in real-world academic settings.
This edition of The Edtech Insiders Newsletter is sponsored by Starbridge.
Too often, outbound sales in EdTech feels like guesswork. Starbridge surfaces early buying signals - meeting minutes, strategic plans, competitor contract expiration dates, and RFPs - to help your team build an accurate, predictable pipeline. See why revenue teams at Collegevine, Amira learning, and Kiddom rely on Starbridge for predictable EdTech growth.
Interview: Juan Zavala, New Markets Venture Partners
We have had some amazing guests on The Edtech Insiders Podcast in the last few weeks. One of our stand-out interviews is with Juan Zavala, who joined New Markets Venture Partners in 2019 and is a Partner. He is responsible for sourcing, evaluating, and executing new investment opportunities as well as supporting existing portfolio companies and firm operations.
Here’s a deep dive on our interview with Juan, and we encourage you to give the full episode a listen for more!
Slow and Steady Venture Capital Strategy in EdTech
Juan describes a distinctive, long-term investment philosophy that contrasts with traditional fast-growth venture capital. He advocates for “slow and steady” progress toward sustainability and profitability, particularly because education is a complex, slow-to-change sector with limited market scalability. Rather than chasing unicorns, he suggests focusing on sustainable outcomes and realistic growth expectations.
“Don't look at raising a lot of money at high valuations as the metric of success. The metric should be to raise the right amount of money from the right source and get to a cashflow positive business… You don't really have a business unless you're generating cashflow, or else it's just you spending money to make less money.” - Juan Zavala
The Double Bottom Line — Financial and Social Impact
Juan highlights how New Markets Venture Partners integrates social impact with financial returns. He emphasizes selecting companies that serve underserved populations and whose solutions are backed by third-party research. Education companies must deliver not only revenue but also measurable efficacy.
“We're trying to find a way to make this accessible to the other folks… We're very careful to invest in things where there is efficacy and that it's not going to experiment on students. We have an impact framework that we use to monitor what's the reach, what's the depth of the impact, who are the stakeholders actually being impacted?” - Juan Zavala
Guidance for Entrepreneurs on Capital and Valuation
Juan gives candid advice to entrepreneurs: be wary of raising too much capital at inflated valuations. He explains how misaligned incentives and unrealistic growth expectations can hurt founders, employees, and even customers. Instead, he recommends targeting the “right capital” for your stage and vision.
“Venture capital isn't necessarily the right capital for every company, especially in education… The metric of success should be to get to a real business and then sell — and have everyone on your cap table and everyone who ever supported you and yourself make a return… not just chase media headlines about raising a hundred million at a $2 billion valuation.” - Juan Zavala
Company Spotlights – Real-World Applications of Strategy
Juan discusses a range of portfolio companies including NEXford University, BetterLesson, Brains in Motion, and CreatorUp. He explains how each balances financial sustainability with impact — from enabling affordable, skills-based higher education to providing data-driven professional development and safe, structured after-school programs.
“With Brains in Motion, for example, the thesis was: kids are only in school until about two or three, but there’s a lot of opportunity for learning from three to six. One principal told me, ‘My kids are on the streets after school — if I can keep them from three to six in these formative years, I can prevent a lot of that.’ That’s impact.” - Juan Zavala
Navigating the AI Wave in Education
Juan shares cautious optimism about AI's role in edtech. He emphasizes the importance of integrating with existing systems instead of trying to replace them and sees long-term value in AI's potential to structure and interpret data rather than simply automate instruction.
“Partner with the existing infrastructure and institutions… try not to rip it out. In K–12 especially, no one's really incentivized to take risks. AI may be great for data structuring and inference — like what we're seeing with Sensia and Workday — but if we experiment too much, we risk creating negative experiences for students.” - Juan Zavala
Curious to Learn More?
You can listen to our full interview with Juan Zavala, as well as interviews with many other edtech founders, investors, and thought leaders at The Edtech Insiders Podcast! Check it out, and as always, we’d love to hear what you think!
I sense many of us build similar profiles right now, stored in Notion (or equiv), then pasted into [tool.ai]. I think the question becomes, in what Use Case / context is such profile being ported?
E.g., by the learner for the learner in tools subscribed to by the learner?
OR, by an institution for the learner across variety of enterprise-provided tools (i.e., at scale)?
OR, ....similar thinking.
And then layering on regulatory compliance layer on such portability.
Each such vector has its own +/-. E.g., while "by the learner for the learner" breathes fresh of empowerment, it requires users not to be lazy FLOBT about maintenance. When blockchain-rooted IDs first came out, startups realized that users still trusted big-tech to do the hard work of preserving their credentials (and do the trust-signaling).
On the other hand, speaking of blockchain, if a By/For Learner profile could exist, it could be authenticated in a privacy-preserving manner by *any tool being used by *whatever institution such learner eventually attended. That might solve for two use cases in one: personal and at scale portability.
Brilliant, Alex. A great analysis and clarification on what many of us are working toward.
Love it. A few thoughts:
I sense many of us build similar profiles right now, stored in Notion (or equiv), then pasted into [tool.ai]. I think the question becomes, in what Use Case / context is such profile being ported?
E.g., by the learner for the learner in tools subscribed to by the learner?
OR, by an institution for the learner across variety of enterprise-provided tools (i.e., at scale)?
OR, ....similar thinking.
And then layering on regulatory compliance layer on such portability.
Each such vector has its own +/-. E.g., while "by the learner for the learner" breathes fresh of empowerment, it requires users not to be lazy FLOBT about maintenance. When blockchain-rooted IDs first came out, startups realized that users still trusted big-tech to do the hard work of preserving their credentials (and do the trust-signaling).
On the other hand, speaking of blockchain, if a By/For Learner profile could exist, it could be authenticated in a privacy-preserving manner by *any tool being used by *whatever institution such learner eventually attended. That might solve for two use cases in one: personal and at scale portability.
Fascinating topic, Alex.