There’s a school in Texas that has received breathless coverage about its groundbreaking model where students learn from software instead of teachers, from Michael Horn, The Week, and NYTimes. There Wired.

Some claim it should be the future of learning - that no other system could possibly produce results like this.

Others ask- why is Alpha so successful even when other edtech is not?

I was curious about this - how are the results at Alpha so much better than most large education studies? So I dove in.

They are opening locations all around the country, including in my hometown of Chicago. So my kids attended a shadow day in a hotel in River North to get a firsthand feel for what the experience would be like.

I also visited the Alpha High School campus in Austin. Several current students and guides answered many of my questions. The students were impressive and shared all the creative ways they are learning and growing with the program.

I want to understand: what is it about Alpha school that seems to work - and what doesn’t? And what can we learn from the program to apply to other progams?

Research written by a bot

I could find only three solid sources of data about the performance of Alpha school students: the official white paper, a blog post from last summer, and Joe Liemandt’s recent mid-year report card tweet.

Recent reports were all written by LLMs. In both cases, someone from Alpha School entered their raw school data to either Claude or Gemini, and then copy/pasted the results into a post. These are the three sources:

The Official Alpha 2 Hour Learning Paper (2024)

The official Alpha White Paper covers data up through 2024. It includes attainment in both elementary and high school at Alpha Austin, growth data from NWEA MAP, and a few cherry-picked examples from the Brownsville campus.

The Austin Scholar Post (2025)

The 2024-25 data published in a blog post on Austin Scholar blog last summer. Most of the post is LLM rehash - Austin said “I uploaded the K-2 class results from NWEA’s MAP test to Claude Sonnet 4 Deep Research”. He provided the original NWEA MAP School Summary PDFs so we can see the raw school results.

The Alpha Mid-Year Report Card (2026)

More recently, the founder Joe Liemandt tweeted out an LLM-generated 2,500-word analysis as their 2025-26 mid-year report card. The post includes rehash from Gemini and Claude.

This is not an actual research paper or even one that was written by a human. We all have seen how chatbots aim to please and will make overconfident claims that gloss over problems.

Since we’re just using AI for analysis, I plugged the Mid-Year Report card into Gemini and asked it to highlight questions to ask. I asked Gemini, “what are the questions you would ask the researchers to really understand the mid-year report on Alpha School?”

Here was its response:

This report is statistically staggering, but in the world of educational data, “too good to be true” usually means there is a hidden variable in how the data was collected or filtered.

As a researcher, I wouldn’t just look at the scores; I would look at the denominator. Here are the two critical questions I would ask the Alpha School leadership to determine if these results are a breakthrough in AI instruction or a masterpiece of data selection.

1. What is the “Matched Student” count? How many students were excluded from the growth data because they left the school or missed a test window between Fall and Winter?

2. What is the baseline cognitive ability (IQ) of the cohort? Are these results the product of the AI’s instruction, or is this simply a group of students who were already in the 99th percentile of innate ability upon admission?

Both of these questions focus on aspects of selection bias.

Selection bias is the most powerful force in education

Ten years ago, I was CTO at an education company. We built personalized learning software that provides students an adaptive path. I spent many hours pouring over standardized test results from school districts across the country. And I recall one powerful article - why selection bias is the most powerful force in education. Whenever you try to understand a particular education statistic, you have to understand: how was the underlying sample constructed? Who was chosen to participate? And how might that skew the results?

Despite the national profile, Alpha is a very small program. Nationally, they serve only a few hundred students. Only 154 PK-8 students made up the sample for the 2024-25 results discussed above. In such a small sample, averages can easily skew based on relatively small factors.

So who is in the sample? We know a few factors that may bias the selection:

• Parental Income. At the flagship location, $40,000 a year, and at some others even higher - Alpha Chicago will charge $55,000! Parental income is highly correlated with standardized test performance, so having a high tuition would lead to higher results on tests. As Dan Meyer said, “They have replaced poor kids with rich kids”

• Shadow Day Selection. Alpha students are not technically selected on the basis of a placement test (that’s what the Alpha GT is for). But the school is still choosy. Students are required to come for a shadow day, where students spend a few hours on educational apps, where their academic level is assessed. Guides also look for kids who are good communicators, respond well to feedback, and fit in well with the community.

• Expulsion. Like all private schools, Alpha can choose to counsel students out. I asked at an info session, why would you counsel a student out of the school (i.e., expel them)? She said that if a student is struggling, it’s never the student’s fault. Still, if parents aren’t supportive of the school’s methodology, then they may be asked to leave.

  And if they leave before the NWEA spring testing window, then their results won’t count in the growth stats.

• Only Alpha Austin. Alpha has numerous locations around the country, some which have been open for more than a year. Yet, the only public datasets I’ve seen have come from the flagship Austin location.

  

It goes without saying - that unlike public schools, no private school HAS to offer support to students who require special education, or have behavior challenges, or any number of other potential obstacles to their education. Alpha can choose to serve a small number of high achieving, rich students with strong parental support who are already committed to their model.

It is all fine to create another model to serve those with lots of support But it doesn’t tell us much about whether this extends to those who aren’t already successful.

Alpha students come IN testing at the 99th percentile

One thing we know: on average, most of the incoming class in Fall 2024 was already testing at the 99th percentile. The average RIT score is a full grade level ahead.

For the older kids, maybe they test so well because they’ve already been at Alpha School for years. But we see this effect in the four lowest grades: K-3 are all coming in at the 99th percentile.

If Kindergarten students come in to the school already demonstrating that they are excellent test-takers, then it’s not as surprising that they continue to test well in the spring.

How do we measure learning?

One of the best parts of Alpha is the belief that students can do things well beyond what’s typically expected of their age. Students build companies. They write musicals. They create all sorts of incredible projects in the community.

Despite all this grandiose rethinking, Alpha primarily uses a very traditional metric for success: standardized tests. For K-8, this means NWEA MAP and for high SAT and AP tests for high school. They also use other metrics, such as surveys about whether students like school, but their primary optimization is for test scores.

Students are all hyper aware of their test scores and work to optimize them. At the walkthrough, each student touted out their own SAT scores - not only juniors and seniors, but freshmen as well. The Timeback system for little kids incorporates their NWEA scores front and center on the main dashboard.

Yet, despite the core focus on test scores, their actual analysis is weak and mostly powered by LLMs. The white paper that they have released has major errors. These are best interpreted as marketing, and certainly should not be compared to peer reviewed, third party research.

Using the wrong column for growth

When I first heard of Alpha School, I wanted to know: how did they get such great growth numbers? I read the white paper. And as I read the blurry reports that form the centerpiece of their claims, it didn’t quite make sense. I’ve looked at many School Summary reports in my career, and this one was missing a critical column - the growth itself.

Here is an NWEA summary page from the original white paper. Alpha’s annotations claim that this shows 2.5x growth. They claim that when averaged, it shows that “Alpha students are learning an average of 2.47x faster”.

I noticed that the data points on the bar graph on the bottom didn’t match the data points on the highlighted column. This is because it’s the wrong column!

The chart above shows the average of the growth standard error! It’s not even the real growth number! Check out the official NWEA MAP definitions page to see that there are two columns:

• Observed Growth - Average change in RIT scores from starting term to ending term. This is the column they should be using.

• Observed Growth SE - Growth standard error (SE) associated with term-to-term growth for the group.

Here I place the Alpha report next to the NWEA sample so you can see - they have blurred the most useful column, and instead chosen to average a meaningless statistical measure of error rather than the correct data.

An error like this is huge. It should have been caught in early review. Yet, it wasn’t - and in fact that error serves as the basis for the 2x learning claim.

In truth, when I looked at the data for 2023-24 school year, the actual learning gain was closer to 1.3x in math and 1.6x in reading. For the 2024-25 results, the learning gain is a little higher - on average closer to 1.7x.

Does this matter? I mean, these are fantastic results! Kids at Alpha certainly do great on standardized tests. If you want your kid to do well on standardized tests, then Alpha might be a good place to go.

Alpha School should enlist an external researcher

The future of education is a huge open question. LLMs are exponentially improving, getting closer to AGI every day. How will students learn today and in the future? What works best? And what doesn’t work?

Alpha has popularized many ideas about the future of education. In many ways they have made mainstream the vision of a personalized learning software that knows each student and guides them through the school day. I love that they believe in their students and help them accomplish goals that seem out of reach for their age. I love the supportive community. And they are pushing the boundaries and driving innovation on the use of software to track progress and implement learning.

Most private schools don’t publish data to the degree that Alpha does- but also, most private schools don’t claim to be the future of education. When a school makes outsize claims of once-in-a-generation learning, then they should back those up with rigorous research to support it. They shouldn’t leave obvious statistical errors in their prominent results paper.

If Alpha sees its model as a blueprint for the future of learning, then they must move beyond LLM-generated report cards and cherry-picked datasets. They should open themselves up to proper peer review from respected researchers, and publish complete datasets that let the public evaluate their efficacy. Here are the three must-dos that would help me better trust the efficacy claims from Alpha:

1. Third Party Publication. When companies publish their own reports, they see a much more positive effect than neutral third parties. I would love to see a report from a respected university professor or researcher who looks at the data and does an independent analysis, following education research best practices including sharing methodology and samples.

2. Diverse Populations. I am so curious - what does the Alpha School data look like for Miami? How about the Texas Sports Academy or their Brownsville location? Does 2 Hour Learning show repeated results across different contexts? How does it work in the Unbound Charter school in Arizona?

3. Incoming Student Profile & Attrition. What are the test scores for the incoming students? What are the results including those who left the school throughout the year?

We don’t need a randomized controlled trial. Simple transparency and external validation would go a long way. If Alpha continues to show strong results across multiple locations and contexts, then they should follow the best practices in education research to build trust that the model can generalize beyond just the elite private school market.

We can assume the models will keep getting better. But how do we best incorporate them into the school day, the teacher workflow, so that teachers and students can make the best use of the powerful capabilities?

In software development, this shift is well underway. We have seen a move from copy/pasting in ChatGPT to embedding in an IDE like Cursor. Now, developers commonly outsource entire features or projects to standalone agents using Codex, Claude Code, or even an OpenClaw agent.

But in education, we’re still stuck. Earlier this week, I wrote:

AI in schools has reached a plateau. The majority of teachers have used AI to plan lessons, give feedback, and differentiate their curriculum, but are still stuck in a cycle of opening tabs, typing in prompts, and copy/pasting the results.

With the proliferation of tools, we have seen this is only getting worse.

I think we should look to the teachers for the solution.

—

Some of the best apps in education are built by teachers, for teachers. I had the pleasure of speaking recently with the co-founder of the popular teacher AI tool Eduaide.

Thomas Thompson was a middle school social studies teacher in Maryland. During COVID, he and a colleague started building Eduaide, now used by over 800,000 educators. Thomas identified the same issue:

The way most of these tools are working — ourselves, MagicSchool, DiffIt, Brisk, SchoolAI — what you do is you have a directory of 120 or so different prompts. You click on it, preloads a prompt, then the teacher fills in some information, that gets injected into the prompt and the prompt goes to the AI. You get a response, typically in like a chat format. And then the teacher will copy that, take it to Google Docs and, you know, put some flourishes on it, make it their own.

His team recently shipped their solution to this problem: a unified workspace that includes all the relevant context for a given lesson. The AI reads your document, understands context, and lets you revise in place. It’s the same shift we’re seeing in developer tools, from ChatGPT to Claude Code to OpenClaw, applied to education.

Here he shows how the new workspace works for teachers:

I talked with Thomas about his approach to product development, how they use evaluations to ensure that the tools do what they should, and how to apply pedagogy with myriad choices in AI tools. He gives us a sneak peek into where the future of AI education tools is going.

Listen to the full conversation here.

Try the tool: Eduaide.ai

AI in schools has reached a plateau. The majority of teachers have used AI to plan lessons, give feedback, and differentiate their curriculum, but are still stuck in a cycle of opening tabs, typing in prompts, and copy/pasting the results. Data piles up in its own islands, hidden behind apps and layers in the cloud. Teachers are left to blend it all together into a cohesive classroom experience.

But it doesn’t have to be this way. AI agents are coming to help.

Every spring, I teach two classes centering around robotics: an intro class for 3rd and 4th graders, and a more advanced class for 5th to 8th. The goal is to help students learn basic programming and engineering skills to prepare for the FIRST Lego League competition in the fall.

I’ve deployed an AI agent that can act on my behalf. It doesn’t live in a browser tab - instead, I chat with it using WhatsApp, share with it my Google Docs, and it even manages my classroom website. It bridges the curriculum and practice in a way that a browser-based chatbot just cannot. And all while providing more privacy and control than traditional AI apps.

My setup

I’m using a new system called OpenClaw, an open source tool that runs on my own machine. OpenClaw is the fastest growing open source project in history, propelled by the many people who want more control and power over their use of AI. Here are the key differences from existing systems:

• Local Control: Unlike ChatGPT, which stores memories in its own system, my bot runs on my own computer and stores everything in plain text that’s easy to read and edit using Obsidian.

• Chats with Me: My bot talks to me using WhatsApp. It’s called Mr. Roboto and I have a 1:1 DM as well as group chats with my co-teachers.

• Google Drive. While its internal memory is all plain text, Mr. Roboto can also create and share files in Google Drive, where we already store our lessons.

• Customizable Skills. Whereas MagicSchool offers about 100 tools that you can use, with OpenClaw you can incorporate pre-built skills or build your own, just by talking to the bot.

• Switch Models. You can switch from OpenAI to Gemini to Claude, or even local open-weights models, while preserving all the skills and memories. This gives flexibility in a world where model costs and availability change frequently.

• Privacy. All the data is stored on computers that I own and control. I don’t send student data to the bot. We operate only on generic questions relating to the class. I think in the future

Let me share a few of the ways that I use this bot.

Keeping us all on track

Many people talk about the big things that an AI can do, but I suggest starting small. When getting started with any new tool, begin with baby steps and work up to bigger tasks in a gradual release of responsibility.

The first layer is simple: Mr Roboto keeps me on track. I asked it to ping me twice per class each week:

• Preparation. The day before each class, it pings me with the current plan and any open items. It suggests a materials list and gives me links for any worksheets that need to be printed.

• Reflection. An hour after each class, it messages our WhatsApp group: “How did the class go? Anything to change for next time?” Then it incorporates feedback for the next session.

The bot lives in a shared chat room with myself and the other teachers, so they can see what is going on and ask their own questions. The reflection happens naturally via chat.

Mr. Roboto is not starting fresh each time I talk to it. It already knows what we talked about last week, and even what I did in my class next year. All this info is stored as files that I can access using Obsidian. Our shared workspace includes our plans and all the materials we’ve made together.

Creating materials for class

I often need to create materials for our lessons, and since the course changes year over year, I may not be able to just re-use what I did last time.

For example, this year I am emphasizing the value of iteration. I want students to try to solve challenges multiple ways and measure the results to see what works. So, I asked my bot to create a worksheet to support that activity. It suggested a few different iterations and we workshopped it until I was happy. This all happened in WhatsApp without needing to go to a computer.

Mr. Roboto doesn’t work alone - for the above, it used the Manus Slides API to generate worksheets, then shared them with my via Google Slides. Once we did this a few times, I asked it to remember as a skill - which is a configured set of actions that the bot can do over and over. Now, when I ask for a new worksheet in the future, it can follow a similar template and process, thus building a consistent body of materials for future use.

If I make changes to the Google Slides, I can tell the bot to check on those and make its own suggestions and revisions. It behaves more like a colleague on a shared doc than a transactional chatbot. Instead of copy and pasting everything myself, the bot does it for me and keeps the results in Google Drive or Obsidian.

Keeping everyone in the loop

I get a lot of questions about the robotics team at our school. I like to send out periodic emails to parents to let them know what’s going on. OpenClaw can help with that as well. It already has all the context of what’s going on including our reflections after class. So I can ask it to help with an initial draft that follows my style.

Note that in this case I don’t have the bot actually send the email - I only use it as a starting point and handle communication myself. But it is helpful to have a base to start from to help me remember all the key points.

Publishing to the web

My bot helps me manage our public class website. When we have a great resource that I want to make available to the students or others, I can ask the bot to publish it to the web so my students can access it from the website. For instance, after yesterday’s class, I realized that some kids may benefit from seeing a few examples of successful solutions that they can use as inspiration. I just asked my bot to find them and put them on the website for me:

It submitted the results to me as a pull request on the website. I was able to test out the change on my phone, give some feedback just like I would on any other PR. We had a little back and forth, but the code looks pretty good so I approved. And then a minute later, the page is now live:

Other teachers in my class can now also edit the website by just chatting with a bot. It can empower everyone to have a role in publishing materials wherever we need them.

Helping new volunteers

We rely a lot on parent volunteers, many of whom have no coding or robotics experience. Now, they have a new tool to help when others are unsure of how to get started.

For example, yesterday a student built a LEGO mechanism using a new type of gear, and we asked the bot to suggest how to help them. Mr. Roboto identified the part in question, saw the architecture of the bot, and suggested a few leading questions - in all, really great and helpful advice for an elementary robotics classroom.

I can’t overstate how helpful it is to just have this available in chat. This is already where the coaches discuss what to do and ask for advice, so the bot fits in without requiring an additional account or app to download. And the bot already has access to the context of what the class goals are - so don’t need to rehash the specifics like you would with a vanilla Google search.

The present: training wheels still required

When I was at a recent conference, some educators shared concerns about OpenClaw. There are some key blockers still to overcome around ease of use, cost, privacy, and security.

OpenClaw requires some familiarity with the command line in order to install open source software - although there are new one-click hosting solutions that are making this easier. As models improve and the rough edges are smoothed, the technical process of setting up a personal agent will grow easier.

Cost is a potential issue but can be kept under control. If you use it sparingly, or with cheaper models, then OpenClaw can cost only a few bucks per month, but more aggressive usage can into hundreds of dollars. The good news is that the competition is white hot, the pricing models change regularly, and OpenClaw is improving its efficiency all the time. OpenClaw includes a helpful dashboard to help monitor usage. If you dive in, I recommend funding your accounts with prepaid fixed amounts to prevent surprise bills.

You don’t have rely just on the big API vendors. Open weights models are not that far behind. OpenClaw recommends using Minimax 2.5, and also supports open models like Kimi, Llama and others. I bet that by the start of school in September, open weights models will be able to do what Opus and OpenAI’s Codex can do today, for a fraction of the cost.

One central question is data privacy. Student data is protected under FERPA and state laws that were written assuming cloud-based software vendors that can sign formal data privacy agreements - not for open source distributed systems. The safest way to get started is to choose use cases that avoid the need to handle student data altogether.

Finally, keep it secure. Data in OpenClaw is stored on the local computer, so make sure you secure that system appropriately. Don’t just give the agent access to private data - instead, create its own Google and WhatsApp accounts and treat it like an untrusted collaborator. Follow best practices or ask for help to make sure it’s solid.

The future: agents for everyone

Despite the challenges, after few weeks experiencing it firsthand, I am confident that everyone will soon have their own agent. And teachers - who are already major users of AI - are going to be leading the charge.

The bigger edtech companies are working to centralize the AI agents to maintain control and a market share. We will see more integrated products like MagicSchool AI Operating System and Brisk Curriculum Intelligence. OpenAI and Gemini are working to target the education market with integrated memory products of their own.

But we need not only build centralized systems. The teachers I know are always looking for ways to improve their classrooms, incorporate new technology to drive learning. Soon, every teacher will have the ability to make their own AI assistant, customized to their own workflow, and which can operate safely.

If you’re interested in setting up your own teaching agent, I’m happy to help! Reply to this post for a free consultation.

Many companies are working to help prepare kids for the near future when AI will be increasingly capable and a core part of all of our lives. LEGO education made some recent announcements that illustrate the direction we are all headed.

In January, LEGO announced that they will retire the popular Spike Prime robotics kits. They will be replaced a new set called “Computer Science & AI”, which will ship in April. Then, in June, LEGO will stop selling Spike Prime anymore. FIRST Lego League is also changing in response, introducing a new Future Edition that will phase in starting next year. These rapid changes are causing a lot of anxiety among schools and robotics teams that are unsure how to adopt the new robotics system.

Andrew Sliwinski, the head of product experience for LEGO education, said that the new kits are “everything that a teacher would need to teach all the foundational principles of computer science and AI.” While most online discussion has centered on the programming & robotics capabilities, I am curious: do these really teach all the foundational principles of AI?

Last week, I had a chance to play with the new kits at the IDEACon conference. In the end, LEGO has built a less capable robotics kit that just flirts with AI models, but the machine learning component is quite limited. There are better ways to teach kids about machine learning, without the cost of a new LEGO kit.

Image classification misses most of what we consider “AI”

The vast majority of students today interact with artificial intelligence in some form on a regular basis. They see the growing power of LLMs. And middle schoolers today will be graduating college in the mid-2030s. By that time, the exponential increase in the capabilities of artificial intelligence systems will have far surpassed humans in many dimensions. Students should learn about these AI systems while still in school to help prepare them for the rapidly transforming world.

There are two approaches to teaching about AI. One way is to show how to use tools like ChatGPT, Gemini or Claude for coding. Many schools teach kids some guidelines about how to watch for hallucinations, spot biases, and not to rely too much on chatbots.

But students should also have an opportunity to go deeper - to learn fundamentals of how these systems work. And there are plenty of incredible products available that help them do that.

While most computer science AI courses are at the high school or college level, there are a few that offer middle schoolers and younger the chance to train models, run inference, and evaluate results. For example, CodeHS offers a free middle school course on AI that shows kids how to train models to classify and generate music using TeachableMachine from Google. TeachableMachine makes it easy for students to gather data, train a model, and then use that model to make their own applications.

The new LEGO systems focuses narrowly on only one branch of AI: human pose estimation. Students can use the computer’s webcam to detect skeleton points on their body - such as elbow, wrist, nose, and legs. They can wire those up to coding blocks to take action based on the position of different body parts, or train a model to detect different poses.

The built-in model can detect whether the webcam sees a person, whether your arms are up or down, and it can measure the angles between different points on your body skeleton.

Once students get a hang of inference on the pre-built models, they can move on to more advanced concept of training their own model. To do this, you can take photos of your body in different positions. These serve as training data for a classifier. Here, you can see my attempt to train a class of poses:

The machine learning models can be wired to LEGO models. For instance, in their demo lesson, LEGO shows off a student controlling the arms of a robot model by waving their own hands:

This is pretty cool! In our training session, one of the participants figured out how to control a robot minifig with her nose. When she lifted her chin, the robot jumped in the air. There are some neat opportunities to help students see how they can control robots using their own bodies.

This is LEGO at its best - brings computer science concepts into the real world. Students can use pose detection to control real world LEGO models, and they can feed in pictures of themselves to direct new poses that aren’t in the pre-built model.

Limited by design

Unfortunately, once you get the basics of the pose classification - that’s about it. It’s a nice demo, and maybe a great lesson for one or two classes - but I don’t think it goes much deeper than that.

We can see how limited it is by looking at the programming canvas. See below the full set of AI blocks that LEGO is releasing in April. Students can detect a person or pose, and leverage the specific angles between body parts.

According to LEGO, their pose detection model is based on the older technology called BlazePose. This model was published six years ago in 2020 - years before ChatGPT or any of the more recent AI models. BlazePose runs entirely in the browser, on consumer hardware (no GPU), using only local data. UPDATE: After I published this article, I re-checked the Lego website, and they’ve updated it to say that they use the model MoveNet rather than BlazePose. MoveNet was published in 2021. It is faster and smaller but otherwise largely the same as BlazePose.

By using only local models, LEGO is trading privacy for power. Andrew Sliwinski from LEGO says this explicitly: “We are bringing AI and education in a hands-on way that LEGO does. But we’re doing it rooted in privacy and trust. … Everything we do runs locally on the kids’ device. There are no student accounts and no data is ever transmitted to a third party.”

When I first saw that LEGO was introducing image classification, I thought that meant the robot would be able to see the world around it. How cool would it be for the robot to react to what it could see? But no - there is no camera on the robot - instead, the images come from the computer’s webcam. A computer needs to be always connected and in range in order to use any of the new programming tools.

When I think of AI & Computer Science in 2026, I think of LLM-assisted coding, or image generation, voice interactivity, or autonomous AI agents. But the issue with all of those is that the newer forms of powerful AI are inherently risky. Lego has chosen to play it safe - keeping all data local on the device means that they can’t leverage modern models that rely on GPUs.

It’s hard to claim that the system covers all the foundations of modern AI without being able to leverage any of the recent models built in the current decade. It seems like the “AI” label is added by marketing, but is more of an afterthought in the product itself. Educators who want to teach AI would do better to use Teachable Machine or other online tools.

LEGO kits are built for classroom teaching in schools

Compared to Spike Prime, the new kits have several improvements that help them fit well within an elementary classroom. First and foremost, they are designed for group collaboration. Second, all connection happens wirelessly from computers - meaning far less cord management. And third, the offer a suite of pre-built lessons with a teacher dashboard, group collaboration built in, and simpler, streamlined sets designed to be re-used by multiple classes in a single day.

Group collaboration built in

The new lessons are built from the ground up to support teams, rather than individuals. Today, because most LEGO kits require that pieces be put together in order, most LEGO assembly works best alone or with a partner. This can make it hard when you have a large group of kids working together.

But the new lessons are designed to allow up to four people to build at the same time. Each page of instructions has very small steps for each of the team members.
The trainer in my session said there are “no bossy builders, no silent sitters”. The lessons are built for teams. In our demo lessons, this mostly worked - each participant waited for the others before moving on.

LEGO trainers reinforced that the new lessons are designed for groups to sit at a table rather than staring at a laptop.

Wireless connections means fewer cords

The new system connects components entirely by Bluetooth. Components connect to each other, in screen free mode, or if there’s a laptop involved, then they can connect to the laptop the same way - by flashing new Color Cards.

Having no wires may mean a simpler classroom experience, when everything works right. However, I anticipate a few potential issues. Each component still needs to connect to a controlling laptop via Bluetooth. Instead of being either connected or not, as with Spike Prime, students will now need to worry about whether each of the motors has its own connection. Also, each component needs to be charged separately. Since each kit has up to four items that need charging, I expect that there will be a lot more “my battery is low” issues in classrooms that forget to plug everything in.

Pre-built lessons with teacher dashboard

Spike Prime today has a few small lessons, but I have found they aren’t that easy to implement. The Spike robot can take a while to assemble, and students are reluctant to break it down just in a single class period. There’s quite a bit of legwork just getting to the lesson.

The new teacher portal has built-in slides, including recommended build time for each small step. And students can join a lesson with a code - a much more standard way to link lessons. The teachers at the conference sessions I went to were excited by the ability to start teaching these with less prep time.

Future is murky for robotics programs

Change is hard. It’s especially hard when the new kit isn’t here yet and everything seems to be changing.

So the new AI & CS kits are likely good bets for classrooms - but how will they fare for FIRST Lego League? Here, I’m a bit more ambivalent - let’s wait and see.

FLL has announced a “Future Edition” that will run side-by-side with the current rules for the next two years. The FLL Share and Learn Facebook group has been doing a great job of tracking the differences, and Creator Academy Australia has an excellent breakdown as well. The key differences are:

• Robots will no longer be autonomous - instead, students will use a laptop for control at all times (or at least most of the time).

• Teams are maxed at 8 kids instead of 10 (although in truth, FLL teams get quite unwieldy over 7 anyway, so this is likely good guidance). Teams also support multiple different roles, not just “technician”, which supports the core values of inclusion and teamwork.

• The new board is half the size - I think this will make it much easier to fit practice into standard classroom spaces.

The new kits cost a little more, but not crazy more than previous sets. At its launch, the EV3 cost $492 (in 2025 dollars), and Spike Prime cost $414. The new CS & AI middle school kit will cost $429 for the 3rd-5th grade set, or $529 for the middle school version.

Seems like the new kit was built first for classrooms, and then FIRST is trying to shoehorn it in to a new system to allow FLL to keep going. My advice: I think robotics teams should sit it out the next year. In Illinois, we plan to continue competing in the Founders Edition, at least until we know more about the program. I think that Spike Prime will continue to have a strong role to play in teaching robotics, and I hope that the secondary market stays strong for some time.

Many schools are worried about the price. If you just plunked down thousands of dollars for a class set of Spike Primes, then it can feel a bit like a bait and switch to have to upgrade right away for new sets as well. In that case, for schools that already have a set of kits, I would recommend waiting to see how the real life usage evolves in the months ahead.

On balance, I think that LEGO has made a good kit for classrooms to use to teach basic coding and building skills through play. Kids will likely have fun. The AI offerings are quite limited, and there are better options such as Teachable Machines, but it’s still a good first effort. There’s no reason they couldn’t expand those offerings in the future as well - perhaps to add a camera or other features.

LEGO has a deep history of prioritizing kids, hands-on learning and fun. Their previous upgrades were met with skepticism at the time but proved to be great bets in the long run. I am optimistic that the teams at LEGO and FIRST will experiment, listen to feedback, and iterate towards an incredible product that will usher in the next generation of engineers - even if it’s not yet quite the right fit for everyone.

In January, Brisk Teaching announced Curriculum Intelligence - “the first AI that actually knows your curriculum.” They promise that districts will be able to upload their own curriculum and then teachers will generate materials that match their own curriculum. CEO Arman Jaffer explains:

Without being grounded in a district’s existing curricular materials, AI can unintentionally undermine instructional quality rather than strengthen it. While content is easy to generate, what we consistently hear from districts, and what’s hardest to get right, is AI’s alignment to pacing, curriculum, and what’s actually taught.

District admins will upload PDFs and spreadsheets of textbooks, teacher pacing guides, sample tests, district policies, and other curriculum data. Brisk tools can then include a portion of that info alongside teacher queries to the LLM, so that the results generations should hopefully be more aligned to what the district is already offering.

It makes sense that AI companies will add these features - districts will pay more for tools that align to their current curriculum. As AI companies move upmarket towards larger districts sales, it’s no surprise that they are adding ability for admins to have more control over the content.

Competitors offer weak or roll your own ways to integrate existing curriculum - but expect that they will incorporate more in the near future. Today, teachers can prompt Gemini or OpenAI directly to please align with XYZ curriculum. MagicSchool claims that their Lesson Plan Generator can ensure lesson plans are coherent with prescribed curriculum - but it requires quite a bit of overhead work from the district & teachers to use a custom tool. These rough edges will smooth out as AI edtech products compete for the big districts.

But will these features actually help students improve their learning?

Teaching to the Test

On the surface, the goals of curriculum alignment seem good. Districts try to align three components: what teachers teach in the classroom, what students read in their textbooks, and what they are tested on. If the teacher says one thing, while the textbook says another, and the student’s test doesn’t reflect either, then they can spend time confused and pulled in different directions. One literature review from 2012 summarizes the evidence supporting benefits to student learning from having aligned curriculum.

But as districts have pushed harder to have teachers implement aligned curriculum “with fidelity”, there has been significant pushback. If you fully align tested materials with what is taught, then you risk leaving out those important skills that can’t easily be tested. Reading and math receive more emphasis than writing and critical thinking because they can be more easily assessed on state tests. This is known as “curriculum narrowing”- when teachers “teach to the test” they may short change areas that fall outside the well defined path.

If Brisk’s new tool means that teachers have less freedom to experiment with alternative lesson plans or resources, then it could encourage more narrowing - and remove some of the benefits of having personalized, AI generated resources in the first place.

Bigger context windows may help enforce the status quo

Rapidly improving models enable these new features, but it may not always be for the best.

While Brisk hasn’t yet published exactly how their new system will work - but we can make some guesses. District admins will likely upload materials to a central repository. The materials could include PDFs of scope and sequence docs, spreadsheets with district plans, or even whole textbooks, depending on how its implemented.

Then, when teachers request a specific tool, Brisk’s agent will determine which materials should be included from the repository, and pass them alongside the prompt to the underlying model. That way, the model can take action with the additional context. It’s like operating with an open book that the model reads in full before every question.

Each model has limits on how much context they can handle, but these are growing rapidly. Only a few years ago, context windows were as low as 8,000 tokens - but more recent models like Gemini 3 and Claude Opus 4.6 can process context of more than a million tokens. And some models are even promising 10 million tokens plus - that’s equivalent to the ballpark of 10,000 to 20,000 pages of documents. Now, the models may not be able to always recall all of the context perfectly, but these new context windows do allow developers to include a lot more than just the user’s prompt.

With more powerful models, AI vendors have a lot more options about how to craft the underlying generation to meet the needs of teachers and students.

Is curriculum alignment the best way to help students?

I am skeptical that more aligned AI outputs will really lead to better results for students. A survey from 2023 by EdReports showed that as more online options are now available, fewer teachers are using standardized resources today - less than half of teachers reported using any aligned curriculum weekly. The availability of supplemental materials long predates AI, but the newer LLMs have supercharged the ability of teachers to pull in and adapt new curriculum for many use cases.

Many times, teachers use AI because the existing curriculum from the district does NOT offer their students what they need. One teacher I spoke with told me that she mainly looks for AI to help when they don’t have a central curriculum, or when it doesn’t work for her students. If the AI vendors force teachers to stay aligned with existing curriculum, then it may undermine the creativity and potential of the tools to offer differentiation or alternate lessons that students may need.

I don’t believe that the ultimate purpose of AI in the classroom is to facilitate a better worksheet aligned to a twenty year old district pacing guide. Instead, we should be building towards agentic workflows that empower teachers to personalize for each student, and tutors that help kids learn at their own pace. By tying these new AI models so closely to existing district silos, we risk hamstringing their potential.

Tools like Brisk’s Curriculum Intelligence will likely help AI vendors close deals with school districts, and provide stronger alignment in the short run. In the best cases, teachers will be able to build lessons that work well within the rest of their schools and classrooms. But in the worst - they may stifle the promise of personalized, AI driven lessons. Time will tell.

Big thanks to Priya Mathew Badger for her feedback on this post - subscribe for excellent tips at her newsletter, Almost Magic.

I had a recent conversation with Robert Gulotty, a professor of Political Science at the University of Chicago. I asked him how his teaching had changed with the rise of AI. He said it has been a huge disruption that has transformed how he teaches.

But not in a good way. Yes, he said that it’s nice that it’s sometimes easier to make lecture notes or presentations. But the real problem is that students use the LLMs to do the work for them, and miss out on the productive struggle that leads to real learning.

 There’s been just an enormous change away from a standard of education which involves a lot of writing and thinking by yourself and turning it in outside the classroom. That is decreasing and it’s being replaced by oral presentations and in-person exams.

The evaluation and the homework - they force the students to engage with the material. And if you don’t do the work, it’s really hard to learn. The pain is sort of a memory and learning device. The fact that you have to work through all the problems and all that suffering helps you remember and learn things.

When I was in college, I spent many late nights wrestling with assembling sources, writing papers, and doing really annoying and difficult problem sets. The struggle is what builds the mental muscles, what helps you be stronger. I sat through plenty of lectures where I was lost, and didn’t understand what was going on. And in the pre-AI pre-smartphone era, I had no choice but to just sit and try to figure it out.

But for many students today, they have easy access to a friendly, always on companion who can help interpret everything. Dr. Gulotty shared that some of his students rely on AI during his lectures:

 I’ve had students who just sit there in class and type everything I say into ChatGPT to help them understand it, sentence by sentence. And they go to the textbook and rather than thinking about what the words say, they’ll just type it into ChatGPT and say, “Explain this to me.”

And they do that sentence by sentence, equation by equation. Part of equation by part of equation. So they never experience the pain of not knowing.

This is a big problem. Who among us can resist the pull of an easy, always on, anonymous and free tool that just makes everything easy?

Ubiquitous AI is hard to avoid, even if you want to

This week there was a deep report released by the Brookings Institution on the risks and promise of AI in education (full report). The researchers surveyed a few hundred people around the world and analyzed many of the studies therein.

The first key takeaway: student use of AI is ubiquitous and worldwide. The vast majority of children surveyed in many countries have used AI. In the United States, more than 91% of children age 13-17 have used generative AI in their personal life, according to a survey from May 2025. Other studies have shown similar depth of usage.

And teens use AI in lots of contexts, in school, at home, and elsewhere, which means neither districts nor parents really have control over the whole picture:

The lines between education technology, friendship, information query, and entertainment are blurring. Students in our study report using AI through general purpose LLMs, social media, Internet search, video games, and education technology, often moving fluidly between these platforms …

Given their pervasive presence in everyday applications, students find it challenging to avoid using AI altogether.

So it kind of doesn’t matter what AI is used within the classroom context. Students (and adults) have nearly universal access, at home and at school. But if students have easy access to something, then it is more likely they’ll be tempted to use it.

One high school student shared with me that his district has adopted the free Gemini education suite for this school year. Since then, he says students are “cheating like crazy” and just using it for everything. This should not be a surprise - when things are easy and available, then people use them more often.

Students are often required to use Google Docs for assignments - but they have recently embedded a “Help me write” feature directly in the app. Sure, students could go out to ChatGPT - but now the AI is just easily embedded right there!

We know that when junk food is present in schools, kids eat more of it - even if they know they shouldn’t. So why wouldn’t the same apply to AI? Who can resist the temptation?

AI is uniquely problematic for kids

The Brookings report highlights several risks and benefits to AI - but the one risk that really worries me is the increasing evidence that access to AI can stunt students’ cognitive development. Even if AI tools do help students improve learning in certain areas, the broader problem is that they just make life too easy.

The overuse, and even the routine use, of AI is fundamentally and negatively reshaping how students approach learning; this has profound implications for cognitive engagement, the development of practical skills, and existential questions about human purpose and agency. Study participants identified this threat to student learning as the greatest risk posed by AI.

The issue is fairly simple: constant access to an AI fosters dependence. Learning happens when you try hard problems, and then work through the solutions. But students find success using chatbots, then they grow to depend on them - and don’t develop the underlying critical thinking and literacy skills to fall back on when AI hallucinates, shows bias, or is just not available.

AI represents incredibly powerful technology, and professionals doing real work get enormous value from it. But they have years of experience to fall back on. But if the goal is learning, then having an AI do the work can get in the way. Eric Wastl, writer of the incredibly fun annual Advent of Code contest, points out “If you send a friend to the gym on your behalf, would you expect to get stronger?”

The Brookings report goes into a great summary of the issue:

The people harnessing these powerful AI tools are professional adults with fully matured brains. They have already developed sophisticated metacognitive and critical thinking skills that undergird their approach to their work. They have deep expertise in their domains, and the cognitive flexibility that comes with such expertise, allowing them to navigate, evaluate, assimilate, reject, and strategically use the information AI generates.

For students, the situation is fundamentally reversed. They are not mini-professionals. Their brains are developing, undergoing crucial processes of neural pruning and strengthening that depend on repeated cognitive effort and struggle. They lack the metacognitive skills, critical thinking abilities, and neurobiological maturity of adults. School exists precisely to build these capacities through sustained engagement with challenging material.

The productive power of AI tools, which amplifies adult expertise, can undermine this developmental goal when placed in the hands of learners. Cognitive development requires the effortful processing, mistake-making, and problemsolving that AI too easily circumvents and that many students are willing to bypass. Used liberally, AI is not a cognitive partner; it is a cognitive surrogate. It does not accelerate children’s cognitive development—it diminishes it.

Offloading diminishes capacity

The report cites several studies that provide evidence of reduced abilities in people who have access to digital tools like phones, search engines, and AI.

• “Teachers report a digitally induced “amnesia” (Lee et al. 2025) where students cannot recall information they submitted or commit that information to memory, resulting in a loss of basic factual knowledge of a topic or domain.” Lee et al, The Impact of Generative AI on Critical Thinking.

• “The ease of offloading work to AI is negatively impacting students’ love of learning, curiosity, self-confidence, sense of self-efficacy, sense of agency, and capacity for ethical reflection and nuanced problem solving” citing Bozkurt, et al, The Manifesto for Teaching and Learning in a Time of Generative AI.

Critical thinking is particularly at risk. One paper in particular was published last year by Michael Gerlich in Switzerland. It shows just how dangerous the lure of AI can be to our critical thinking skills.

Gerlich points out:

Because humans have evolved to cognitively offload, students naturally take shortcuts when given the opportunity. As cognitive scientist Daniel Willingham observes in his essay, Why Don’t Students Like School? “Unless the cognitive conditions are right, we will avoid thinking.” For teachers we interviewed, they see Willingham’s observation confirmed in their classrooms as students become consumers, not producers, of information.

When chatbots do the work

When students do the work, they learn not just the immediate goal of the assignment, but the broader mechanism of how to think. One professor at University of Chicago, Russell P. Johnson, described the essential problem of what happens if students rely on AI to do their writing:

I’ll encourage you to think about what will happen if you use AI and don’t get caught. What will you miss out on if you subvert the writing process? Is learning how to write argumentative papers simply a drudgery to be automated away as soon as possible, or is it a spiritual exercise?

When I assign you to write a four-page paper on the Zhuangzi, for instance, it is not because I am under the illusion that you will need that knowledge later in your professional life. Writing a four-page paper on the Zhuangzi will not prepare you to compete in the high-pressure economic landscape of twenty-first-century America.

No, the reason why I insist you do writing assignments like this is that they give you valuable practice discovering insights and communicating them to others. Reading texts closely, encountering a problem, developing a plausible interpretation, and persuading readers of that interpretation—these are the steps one must go through in order to write a good paper. Going through these steps again and again makes us clearer thinkers and better communicators.

That essay was written in April 2023, just a few months after ChatGPT launched. Since then, his predictions have come to pass - and we are only at the beginning of the age of AI in schools.

The future is unclear

Teachers are facing a challenge: what to do about this cognitive offloading? The report has a few recommendations, but at the core we need to shift educational experiences in school. AI is here, and only getting more powerful by the day - we cannot only return to the “old school” methods of teaching. Instead, they suggest schools adopt child-friendly product design - AI tools that challenge and help students learn rather than doing the work for them. But cell phone bans and blue book exams will only get us so far.

The authors leave us with the acknowledgement that we don’t really know what exactly needs to be done: “None of us have all the answers right now. And yet the stakes to get AI right for our children could not be higher.”

In future posts, I’ll go deeper into examples of what educational experiences are working and how we can maintain rigorous education when students are surrounded by ubiquitous tools that can easily do the work for them.

I’ve been working on writing feedback for a long time. In 2017- long before ChatGPT - my team at eSpark developed an AI system to give writing feedback. When I spoke to teachers back then, many told me they were apprehensive about sending AI-generated feedback directly to their students.

Fast forward to today. We have very powerful LLMs, and teachers use them to generate feedback all the time. Yet, they still aren’t quite at the point where most teachers feel that they can trust the outputs completely. Here’s two teachers from the report on use of Claude in education:

Using AI in grading remains a contentious issue among educators. One Northeastern professor shared: “Ethically and practically, I am very wary of using [AI tools] to assess or advise students in any way. Part of that is the accuracy issue. I have tried some experiments where I had an LLM grade papers, and they’re simply not good enough for me.

And this is an issue - because if teachers need to verify everything the LLM does, then it just adds additional busywork:

Most educators seem to agree that grading shouldn’t be anywhere close to fully automated. The ability to evaluate AI-generated content for accuracy is becoming increasingly important. “The challenge is that with the amount of AI generation increasing, it becomes increasingly overwhelming for humans to validate and stay on top,” one professor wrote.

For teachers to get more leverage from AI agents , then the systems need to be more accurate and reliable. But even beyond the issues of accuracy, one-shot responses from ChatGPT or Gemini may also have a number of problems - sometimes the feedback is too wordy, sometimes it is off topic, sometimes the AI can gush with praise for really mediocre work, or the opposite - criticize when a student needs some encouragement.

But there has been quite a bit of research that shows some ways forward.

Teach AI to check its own work

I’ve been reading quite a few papers from the incredible Stanford SCALE GenAI Repository to understand the research landscape. One in particular stood out to me. The paper is called From First Draft to Final Insight: A Multi-Agent Approach for Feedback Generation, published at AIED last summer. Authored by Jie Cao at UPenn and Ken Koedinger of Carnegie Mellon, it looks at one of the most essential elements of an automated educational systems: how do you give quality feedback to students?

In the First Draft paper, they look at one mechanism for improving the feedback: have the AI check its work against a defined rubric before submitting to the students. This simple technique turns out to be surprisingly effective.

Rather than just generating and delivering the first feedback a model prints out, the authors propose a three-step process called G-E-RG:

• Generate. Initial feedback generated using a standard prompt (they tested six different prompts, including some that incorporated source material)

• Evaluate. Feed the feedback to a different prompt (maybe a different LLM), and ask to evaluate the first draft against

• Re-Generate. Give the model initial feedback, the rubric and evaluation as context, then ask it to try again.

That’s it - the LLM loops once, asks itself to see if it’s good, and then proceeds. The authors tested it on a set of responses from 208 students. They measured the quality of the feedback based on human evaluators. And the outcomes were quite positive.

What is good feedback, anyway?

Before we get to the results, it’s worth thinking: what makes good feedback good? In this paper, they formalized some criteria. They want to see first, that it is reliable (does the feedback actually adhere to the truth in the course materials?), then, do the comments provide critiques & strengths, and concrete actionable next steps? They also look for evidence that the feedback would strengthen the learner relationship with the teacher and encourage agency and independence.

The authors found that when an LLM takes a moment to reflect, it did result in better feedback for students. Feedback generated with G-E-RG was on average more accurate, highlighted more strengths of the student response, and suggested more directly actions the students could take. The authors concluded:

Regardless of the quality of the initial feedback, through evaluation and regeneration, it can be transformed into high-quality feedback in the second round, which reduces the instability of LLM-generated content, ensures quality control of feedback, and enhances its applicability in real-world educational settings. This suggests that the “G-E-RG” method is extremely effective in aligning feedback with higher quality.

Do we need a custom frameworks?

But it doesn’t come for free. More computation and thinking means higher cost and latency. And there are simpler ways to just utilize tokens.

Many LLMs, like OpenAI o3 or GPT-5, have a “thinking mode”. This lets them spend tokens before presenting answers to the user. The model decides how to think and what to spend tokens on. In many cases, the thinking mode may decide to run a process similar to the G-E-RG reflection detailed above. And as the models get smarter, such heavy frameworks may become redundant.

But in some cases, a few like custom frameworks for reflection still pay off. A few recent papers illustrate the benefits.

Iterating to success in data science

Data scientists often follow a multi-step, trial and error approach to solve problem. Researchers from China released a preprint in October showing their work on a new DeepAnalyze model which embeds significant inference-time compute. The model uses special tokens to analyze a problem, reflect, and then revise and try again until it believes the problem solved. They found that it did better on a specialized task.

DeepAnalyze consistently outperforms all compared systems across every task. Notably, agent systems built on proprietary LLMs with tool calls exhibit a significant performance drop on open-ended data research tasks compared to more instructive tasks, such as data preparation, analysis, and insight, where explicit steps or goals are provided.

So is data science a specialized loop that may be outside the core training of the major LLM vendors? Seems like for now, yes, and so a specialized inference scaffold shows improvements.

Coding loops

As we see with Claude Code, LLMs can be incredibly effective if they are allowed to operate in a loop - try something, check results, try again, and continue until success is reached. A short paper published in September showcases a new model called LLMLoop. The authors found enormous improvements in specific coding workflows by giving LLM structure and testing out improvements in a loop.

• Even without a loop, having structured, systemic steps helped a lot. The baseline initial generation got 71% of the benchmark problems correct. But with structured steps, that jumped to 80%.

• But doing multiple loops helps a lot too. When the LLM was able to try 10 times (instead of just one), success jumped to over 90%.

The need for structured loops in education agents

This points to one of the foundational ways to make any AI agent system more reliable: help it reflect and check its work along the way. With coding systems, we see that unit tests, command line tools and build systems can provide checks that are more easily automated.

In education, many steps along the instructional path are squishy and harder to measure and test precisely. It’s easier to tell that a unit test passed than that feedback is excellent, or that a student has understood the material. But as we see here, we can use LLMs to help with the evaluation to maybe overcome that ambiguity and still provide useful tools. They don’t have to be perfect, just good enough to be able to gradually earn trust of teachers and students over time.

Consider the rise of robot-assisted surgery. In recent years, robotic assisted surgery has offered levels of precision and control previously beyond reach. Newer AI systems have taken it to the next level, offering even more autonomy and support, resulting in better outcomes in the operating room.

So, are surgeons being replaced? No, the majority of surgeons see AI as a positive development and are excited about its potential to improve care.

We have not yet seen many robot teachers, but we do see that the majority of teachers say they are using AI. However much of that use is still in one-off, question-and-answer format. The most popular use cases include lesson planning, generating assessments, and adapting materials. But most of the time, the teacher prompts an LLM, and then copy/pastes the results into a doc or slides and moves on. The AI isn’t that actively involved in the teaching itself, most of the time - yet.

In other industries, we are starting to see emergence of much more powerful agents, which can take actions in a loop on behalf of the user. Instead of needing to glue together each step of a workflow. We are starting to see more advanced agents that can take on bigger chunks of work, check the outputs, and then iterate to continue. Some mainstream use cases include email assistants, research tools, sales agents, and automated customer service agents.

We have yet to see this type of autonomous agents in use at large scale in education, but they are coming. Dan Carroll, founder of Clever, said recently

I think 2026 is the year that we see AG Agentic AI make its way into schools.

What will this mean for teachers? Some are worried that AI will replace or disempower teachers. But I think we can learn from the experience of software developers. I want to share my own experience using a coding agent.

Claude Code is the first ubiquitous agent

I have grown to love using Claude Code over the last year. I don’t mean just Claude, which is the LLM offered by Anthropic, but specifically Claude Code. It’s a tool that’s used by developers to automate huge portions of the software engineering workflow.

Claude Code was released just a little over a year ago, but in 2025 it has rapidly became the dominant coding agent for experienced and junior engineers alike. Andrej Karpathy reinforced how transformative this new tool is:

Claude Code (CC) emerged as the first convincing demonstration of what an LLM Agent looks like - something that in a loopy way strings together tool use and reasoning for extended problem solving. … It’s not just a website you go to like Google, it’s a little spirit/ghost that “lives” on your computer. This is a new, distinct paradigm of interaction with an AI.

Claude Code does more than just answer questions. Instead, it can write code, run it, check if it works, and then iterate to make additional improvements, all in a loop and mostly on its own (albeit, with a strong permissions framework).

Instead of making suggestions to the user to take actions, the AI agent itself can take those actions (with permission) to automate more of the process. I have personally had a lot of fun playing with Claude Code. In this year, I have already written more code in 2025 than I had in the three prior years combined.

When I use a chatbot, it feels like I’m talking to an experienced advisor who can give me pointers and advice. But with an AI agent, it feels more like I have a partner who I can delegate tasks to. I still need to supervise the agent, make sure I’m asking the right questions and that its output is sane, but it’s much more empowering.

Here’s one example of how the agent is more helpful than just question-and-answer. I had a recent problem where a build was failing on my machine. I just told Claude Code to make it work:

Claude Code didn’t just offer a suggestion. Instead it actually tried to build it, then worked on the problem. It ran the build, saw that it was failing, then tried to fix it a few times. Each time, it tried to rebuild, until it finally figured out the full set of errors - and reported that it was successful:

By taking actions on my behalf, the agent can run for longer periods. This doesn’t eliminate the need for engineers - in fact, AI makes each engineer more powerful. Michael Novati of Formation.Dev, points out that the most experienced engineers get a ton of value from the tools:

For senior engineers, AI is leverage. For myself and other senior engineers, AI is not flattening experience. It’s amplifying it. The engineers with the most architectural awareness, the strongest product instincts, and the clearest standards are the ones extracting the most value.

In their book, Vibe Coding, Steve Yegge and Gene Kim identify five dimensions in which AI improves the software engineering experience. It lets engineers:

1. Write code faster. At its best, AI tools can speed up development tremendously.

2. Be more ambitious. Projects that once seemed too difficult or time-consuming become feasible.

3. Be more autonomous. Developers can accomplish tasks autonomously (and in some cases, alone) that otherwise would have required help from teams, freeing them from needing to coordinate with others.

4. Explore more options. Vibe coding lets you try things out at lower cost, exploring multiple paths in parallel.

5. Have more fun. Vibe coding lets you spend much more time in the “flow” state, blow through tedious tasks, and shifts the focus from implementation details to building things. At Adidas, they found that vibe coding resulted in developers spending 50% more time in what they called “Happy Time,” productive time when they were mastering their craft.

In one popular study from October 2024, researchers surveyed developers using Github Copilot, the most reported change was that it increased work satisfaction or just felt good. After this, the most common change was that work was less boring since boilerplate and repetitive code was taken off their plates”.

So - vibe coding is improving output, making coding more fun, and giving the most experienced engineers more leverage. Can “vibe teaching” do the same for educators? Can AI make school more enjoyable and effective for teachers and students?

Educators say that AI saves them time

The biggest benefit teachers cite today about use of AI is that it saves time. I see this over and over in studies and reports, and time saving came up as a big factor in my own personal interviews with educators (which will be posted on my podcast soon!)

Gallup ran a survey in June 2025 which claimed that teachers see about 6 hours a week of time savings by using AI tools.

Teachers who engage with AI tools more frequently report greater time savings: Weekly AI users save an average of 5.9 hours each week. …

“Teachers are not only gaining back valuable time, they are also reporting that AI is helping to strengthen the quality of their work,” said Stephanie Marken, senior partner for U.S. research at Gallup.

Another recent, study on teacher use of GenAI found that “the adoption of generative AI tools (like ChatGPT and DeepSeek) is directly associated with increased happiness and reduced stress among faculty.”

In this EdWeek report from last year, teachers went into some real world use cases that highlight how they use AI. A repeated theme is that it saves a tremendous amount of time, and reduces burnout.

“Pierman now uses ChatGPT and other AI-powered tools, like Brisk Teaching, Quizizz, and EdPuzzle, to generate quiz and test questions. …

Before using AI, Pierman had access to one bank of questions, tied to the textbook, that she would pull from. Each chapter has about 150 questions or more, and she would sort through every single question, choose the ones she wanted students to answer, and modify questions as needed. It took her seven to 10 hours to make one test, she said.

With the help of generative AI tools, crafting an exam now takes 40 minutes. She can put the exam topics into the tool and prompt it to generate multiple-choice questions and even ask it to mix up the questions so students have different versions.”

… With AI taking those tasks off her plate, “I don’t get burnt out as much,” she said.

Teachers save time beyond just lesson planning. It can be helpful in drafting communications and other repeated work:

Joe Ackerman, a 5th grade teacher at Mead Elementary in Mead, Colo., is an early adopter of generative AI tools, such as ChatGPT and Google Gemini. He uses those tools to refine and expedite his communication with staff members and families, produce design-thinking lessons for students, and help grade student work.

“It has helped me free up my time [that I can] then devote to teaching,” Ackerman said. “I’m spending less time crafting emails, I’m spending less time on administrative tasks, and it also helps me provide more feedback on a more frequent basis.”

Some teachers have even begun to use Claude Code in a more automated way, as well. This Reddit user says Claude has completely transformed the teaching workflow as they are able to give raw resources and generate more complete presentations:

I used to spend hours every week writing worksheets, building PPTs, and scrambling for reading activities. Now I use Claude Code to generate custom HTML lessons that run in any browser. I give it a passage, level, vocab list, and what I want students to practice. It gives me a full interactive lesson as an HTML file. Vocab cards that flip, comprehension questions, multiple choice, text entry boxes, teacher mode, progress indicators, everything. I just open it on a laptop or smartboard and teach.

From simple tools to teaching extensions

Even in the examples above, teachers are largely using these tools in a one-off capacity: generate a lesson or feedback or newsletter, and then the teacher uses the plan separate from the AI tool. But what would it look like for an AI agent to act more as a co-teacher, helping drive the core instructional experience?

Look again at the coding agent loop above. From their docs, Claude Code does three things:

1. Gather Context - understand what the present state of the codebase, logs, and objectives.

2. Take Action - make a code edit, run a command

3. Verify Results - check to see if it worked

This strikes me as similar in spirit to many data-driven educational practices. In an analogous classroom example, teachers may do the same steps in a loop:

• Gather Context. Assess what the student is currently able to do

• Take Action. Give a lesson, or some other teaching intervention

• Verify Results. Check for understanding via exit ticket or other lightweight assessment

Today, there are some “adaptive” edtech products such as iXL, Khan Academy, eSpark, and many others that attempt to implement this loop- give a student a multiple-choice question to assess, followed by some lesson or problems, then repeat and adapt. But the standalone edtech products usually operate in their own silo, and with minimal input or interaction with the teacher beyond the initial assignment. When I was working at eSpark, we saw that a large majority of teachers never changed the default assignments - teachers mostly treated the product as something separate from their core practice.

AI agents can extend, rather than replace, teachers

Future AI educational agents may not be separate products. Rather, as they gain more context about the classroom, they can become an extra set of hands that extend the teacher’s intent, just as Claude Code provides leverage to engineers. Imagine an AI agent that could apply the teacher’s instructions to 30 different students at the same time. What if a teacher could identify the broad goals of a lesson, and then an AI agent can handle tailored interventions, still with guidance, permission and oversight from the educator?

The ultimate goal isn’t just efficiency, but rather empowerment. If more mundane tasks such as collating data and preparing individualized lesson plans can be offloaded, then perhaps teachers can spend more of their day in a flow state, focusing on really getting to know the students and working at a higher strategic level about how to address everyone’s needs.

The good news is that there are many experiments underway to make this sort of thing a reality. In my next newsletter, I will surface some real-world examples of AI agents in education - both in research and in practice - that are being used today around the world. Stay tuned and subscribe for more!

AI use has exploded in education at all levels - I’m launching a newsletter to keep track of the changes.

Most teachers are already using AI

The majority of teachers are already utilizing LLMs in some form, and more each day. In June 2025, over 60% of teachers surveyed said they used AI in some form during the previous school year. By October 2025, a separate survey said that over 85% of teachers reported using AI.

Dan Carroll, co-founder of Clever, noted recently that “2025 is the year where I feel like AI went across the chasm from early adopters to something that most teachers have access to and are using, if not every day, maybe every week or every month.”

At the same time, education is a major use case for the large LLM vendors. Over the summer, Google announced that Gemini was free for teachers, and in November, OpenAI followed suit with free ChatGPT for teachers. By making it free, the products are seeing even more adoption. This makes sense: more than 10% of overall ChatGPT use is classified as Tutoring or Teaching.

And major AI-focused educational products continue to grow quickly - for instance, MagicSchool claims use by 6 million teachers (up from 5 million a few months ago).

Educational research is necessary, if a bit messy

I have been craving a deeper look at understanding how the education system is changing today, and what we are learning about it as a society that will inform where it is going in the near future.

So, last May, I started a podcast, The AI Who Taught Me, to follow all the rapid changes as LLMs gained adoption in education. We are effectively launching a huge experiment as society: what happens as we can rent intelligence for effectively pennies or free? How does that change the practice of teaching and learning?

The research world is deluged by studies about the use of AI. There are new papers released every day from all over the world, and I read as many as I can. Last year, I covered a few bombshell reports that gained significant media attention - such as the MIT study that claimed ChatGPT rots your brain, or the Nature meta-analysis about the positive impact of ChatGPT on learning. But there are so many more that fly under the radar.

Deciphering the high quality education research can be a challenge. We have to contend with publication bias, vendor-funded studies, and the fact that research can move slowly compared with the pace of change of underlying technology. Some studies released even just in the last year still reference gpt-3.5 — how can you draw conclusions about the capabilities of the latest models when using three-year-old models?

Does this mean that the academic research world is not worth paying attention to? No, because the alternative - ignoring the academic research - is even worse. Without peer review and the scientific method, we end up with vendor reports and unsubstantiated claims from whichever school or product has the largest PR team. For example, if you search “AI in education” one of the top results is Alpha School, which claims that students learn 2x their peers. However, when I dove into the results, I found that their entire “white paper” is a sham - the results come from using the wrong column in their report. At least with peer-reviewed papers we have a shot at discerning what is really working, even though academic research may be a messy process.

Yet, I’ve found that podcast is not the best medium for in-depth, detailed discussion of research. So I am starting a newsletter to go deeper.

Welcome to Base Classes

The name, Base Classes, has a double meaning. In programming, the base class is the template from which instances of an object can be copied. A base class provides the core principles that define the rest of the whole system.

In education, base classes are be the core ways of operating, many of which have been consistent for decades or longer: students are grouped in classes with peers of the same age, learn from one or two teachers in classes of 20 to 40 students, the day is oriented around time blocks, subjects siloed and departmentalized, on static curriculum pathways, with periodic high stakes assessments.

With each wave of technological and societal change, the education system adapts. Last decade, we saw enormous adoption of laptops and phones, and an explosion of edtech companies that offer more products than were available in the past. In classrooms today, students routinely use laptops to access robust educational programs, pull from a much wider variety of publishers and resources than they had access to 20 years ago, sometimes have automatic differentiation, and assessments are now largely taken online. These changes have had a huge impact in classrooms.

But when we have the rise of fully automated intelligence, will those primitives still hold? Or will there by different, better ways to help children learn what they need to succeed in life?

I don’t know the answers. My goal is to identify the early uses of new technologies and bring them to light so we can all figure out which educational primitive will work. For example, my latest podcast episode highlighted the role of voice AI in helping grow early literacy. Students in many classrooms - but not all - are able to speak and have a computer listen, and in some cases this leads to large learning gains. Does this mean that our future classrooms will further incorporate AI voice agents that can listen and respond to students?

Perhaps - we will see and learn as these gain adoption over time.

What we’ll cover

This newsletter will cover progress in a few key categories:

• Research. We’ll cover recent educational research into efficacy of AI and look at what types of novel use cases are being studied.

• Schools. We’ll talk to teachers and students in schools around the United States and internationally to learn how they are using AI tools. (If you would like to speak with me, please sign up for an interview here)

• Products. I’ll analyze AI education products - from established players as well as brand new startups.

• Themes. Finally, I’ll tie them together into the essential themes - the Base Classes that are driving student learning today and in the future.

Throughout this all, I will link to other thought leaders that I respect and pull from. I would love to have you along on this journey. Please subscribe, share, and send your feedback.