ritchot.me

I built an AI Literacy course

Thu, 21 May 2026 00:00:00 GMT

In my last piece on this site, I closed with a line about how my goal for 2026 was to build. At the time, I did not know exactly what that would look like. I had just finished a stretch of writing about LLMs, benchmarks, and the state of the field, and I was ready to stop talking about AI tools and start making something with them. That vague aspiration resulted in expanding my master’s work into a four-module, research grounded AI literacy program for the corporate workforce, built on a custom coded platform, live now at ai-literacy.ritchot.me.

I finished my Master’s in Educational Technology and Instructional Design just under a year ago, and the capstone project was the seed of what you see at that link. The capstone was a course teaching AI Literacy through the mechanism of tokenization: how language models actually break text apart, predict the next token, and generate language one probability at a time. I built it on Canva, delivered it through Google Classroom (I had an expedited timeline I had to meet before I left my previous employer), and it worked well enough to fulfill academic requirements. But the capstone was constrained by the rubric and by the reality that academic deliverables are written for evaluators, not for the people who would actually use them (I mentioned this in my piece on the MIT study, where I called it a capstone built for rubric requirements that would largely be tossed into a void). The core idea, though, was sound. I wanted to take it and rebuild it as a real learning product, something an L&D director would envision as a useful corporate AI Literacy course.

The gap the capstone had started to address still largely exists when I see people interact with “AI.” People have access to AI tools and are already using them, but they do not understand how these systems actually generate language. People mostly still treat them as magic information retrieval machines, and they lack the judgment framework to evaluate whether what comes back is reliable enough to act on. The World Economic Forum’s 2025 Future of Jobs Report found that 63% of employers identify skill gaps as the primary barrier to AI-driven transformation. Tamkin and McCrory’s productivity research documented an 81% median task time reduction inside the AI conversation itself (with RCTs measuring the full work cycle showing between 14-56%), and the distance between those numbers is doing a lot of work. The productivity gains are available. The workforce does not yet have the skills to capture them. And the training that exists tends to address tool familiarity (how to prompt your LLM of choice for email drafting) rather than the underlying understanding that would let someone evaluate whether to delegate certain tasks and if the output is reliable.

I read Anthropic’s 4D competency framework a few months after the completion of my capstone. The framework defines four observable dimensions of AI fluency: Delegation (knowing which tasks to assign to AI and which to retain), Description (communicating effectively with AI systems), Discernment (evaluating the reliability of AI outputs), and Diligence (maintaining transparency and accountability in AI augmented work). The 4D framework provided a clean vocabulary behind ideas I similarly intuited, and it provided the competency language that connects learning objectives to measurable workplace behaviors. The research I had been collecting, from Handa et al.’s task level adoption data to cognitive science findings on overconfidence and unchecked AI acceptance, mapped cleanly enough that the framework became the organizing spine of the program.

The four modules follow a specific sequence: context, then evidence, then mechanism, then application. Module 1 establishes why AI literacy is a business problem using workforce data. Module 2 surfaces what the workforce is actually doing with AI through interactive data dashboards built on task level adoption research and productivity data. Module 3 teaches how language models actually generate text (tokenization, next-token prediction, attention mechanisms, context windows). This is the module that traces most directly back to my original capstone: if you do not understand how these systems produce language, you cannot evaluate what they produce. Module 4 integrates all four competency dimensions into applied practice: task decomposition, prompt reformulation, output verification, iterative refinement, and a diligence statement exercise that asks learners to articulate their own accountability framework for AI assisted work.

Across those four modules, there are 37 sections, 12 interactive practice activities (including filterable data dashboards, a tokenizer playground, a next-token prediction demonstration, and a multi-step AI interaction sandbox), and 7 downloadable reference materials designed as take-home tools for on-the-job application. Every module section traces backward to a documented gap in the research corpus and forward to a measurable assessment. If it could not be justified by evidence and measured by an observable behavior change, I cut it. The course today is what I consider the minimal viable product. I intend to expand it with further research papers I have already collected and read, but I felt an urgency to get this course out now. It has been almost a year since my original capstone, and I still feel that the L&D/AI Enablement programs I have seen are teaching AI Literacy fundamentally wrong.

The program scopes to the AI interaction model most people use today: conversational interfaces, not agentic systems. Tools like Claude Code, Cowork, Codex, and Antigravity operate on a slightly different paradigm (autonomous multi-step execution rather than turn-by-turn dialogue) and would complicate the curriculum substantially. I may build that as a separate course. But the core competencies the program develops (knowing what to delegate, how to evaluate outputs, and when to intervene) transfer upward when the tools get more capable.

The course was built on a custom coded platform rather than in an off the shelf authoring tool. The Canva and Google Classroom versions had done what they could, but authoring tools impose their own constraints on how content flows and how interactions behave, and they limit what data you can capture. I wanted a learning experience where the instructional design drove the architecture rather than the other way around. Building the platform from scratch was also a bet on a question I think L&D will have to answer soon: how do you architect a learning experience as software? Organizations building internal tools rather than purchasing them will need people who can bridge instructional design and software engineering, and the implications of that shift extend well beyond platform choice.

I built this program in approximately 150–160 hours of total development time, as a solo developer, while holding a full-time teaching position at an international school in Singapore. Roughly 120 of those hours went to research, instructional design, and platform build. The remaining 30–40 hours went to iterative review: verifying content accuracy against source papers, refining instructional sequencing, and integrating additional research. That work is easy to leave off a project timeline but represents the difference between a product that passes a surface-level review and one that holds up under scrutiny. The early phases (research gathering, evidence compilation, instructional design documentation) consumed roughly four hours per weekend session (and at a rather casual pace). Once I had the specifications locked and the content documents written, the build phase picked up in pace, but even then, the total calendar span was approximately eight weeks, punctuated by a full week lost to illness and reduced capacity from a separate injury. To put that in context, the industry benchmark for this level of interactivity estimates roughly 735 hours of development time.

That 735-hour figure comes from Bryan Chapman’s industry benchmark, a survey of roughly 4,000 learning professionals. His Level 3 category (simulations, individualized interactions, gamified elements) reports an average of about 490 development hours per finished hour of content; at the program’s roughly 1.5 finished hours of Level 3 interactivity, that is a baseline near 735 hours for the content alone. Chapman’s numbers assume off-the-shelf authoring tools and cover content development only, so a custom coded platform, an evaluation framework with xAPI event taxonomy, a research corpus built from primary sources, and project management documentation would push the real figure higher, though I will not put a specific multiplier on work the survey was not designed to measure.

Against that baseline, the compression is roughly 4.6–4.9x, a 78–80% reduction across the full development cycle rather than the build phase alone, and it falls within the range Tamkin and McCrory documented for AI augmented knowledge work. I achieved that as an independent developer with no corporate legal review, no compliance requirements, and no learner data collection to manage. An enterprise team would absorb security audits, accessibility certification, data handling policies, and cross functional coordination that narrow the ratio. The savings would stay significant but less dramatic, and anticipating that gap is itself one of the things the program teaches learners when scoping AI augmented work.

The 150–160 hour figure covers this launch build; I am not counting the hours I will spend revising and expanding the course from here, and I could not find a reliable benchmark that accounts for ongoing iteration in any case. Several research sources I have already read still need to be worked more cleanly into the content, and the front end likely needs a redesign. Building it taught me to recognize the tells of AI generated front-end code, much as regular users of AI tools develop an eye for the tells of AI writing.

I was able to build this program in such a compressed timeline because I had the domain expertise to specify what needed to be built and the instructional design background to know why. I also had enough hours with AI development tools to treat them as a working partner. Without those three things, the same tools produce something that looks plausible (the formatting is clean and the sections are logically ordered), but the instructional sequencing is wrong, the assessment alignment is off, and the practice activities test recall rather than judgment. It passes a surface-level review. It does not change behavior. The tool did not replace the expertise. The expertise is what made the tool productive.

The economics of learning development have changed, but not in the direction most people assume. AI does not make L&D cheaper. It makes expert L&D practitioners significantly more productive and forces every member to more fluidly work across the entire stack. A single practitioner with domain expertise, instructional design training, and fluency in AI augmented development workflows can now produce work that previously required an expansive cross functional team. The value proposition of L&D teams has shifted. Volume of output and mastery of a specific authoring tool are no longer valuable metrics. What separates useful L&D teams from obsolete ones is “taste”: the ability to design programs grounded in evidence, specify clearly what needs to be built (and what “done” looks like), more carefully evaluate whether what was built actually changes behavior, iterate quickly, and use AI tools as a development partner throughout.

The workflow that produced this compression mirrors what the program teaches. Every build session followed a specification-verification loop: I would write a detailed specification document (what to build, what content to use, what acceptance criteria to meet, what decisions the AI should make silently versus what required my approval), hand it to Claude Code for implementation, review the output against the specification, and iterate. The AI handled the volume (generating component code, populating content, wiring up state management), and that freed me to work on other tasks in parallel. I handled the judgment: verifying accuracy against source papers, checking instructional sequencing, and catching the moments where a technically correct implementation missed the pedagogical intent. That loop, specification then verification, is the Delegation-Discernment cycle the program teaches. The build process became the case study.

The platform tracks learner progress (locally, I am not collecting anything on my end for this version), knowledge check responses, practice activity completion, and time-on-task using an xAPI-aligned event taxonomy. A built-in sample admin dashboard (Cmd+Shift+A on Mac, Ctrl+Shift+A on Windows) surfaces completion patterns, knowledge check response distributions, and event timelines. The Kirkpatrick evaluation framework (reaction, learning, behavior, results) is visible in the program’s architecture diagrams by default to signal thinking at the management level rather than the course level. Because evaluation lives in the architecture, the program generates the data an L&D director needs to justify continued investment and measure behavior change: reporting that usually requires a separate project. If the bet I described earlier pays off (if organizations do start architecting learning experiences as software rather than purchasing them), L&D teams will need to work much more closely with technical teams. Someone has to verify that the AI generated code is secure, accessible, and maintainable. Someone has to ensure the platform architecture does not introduce compliance risks or data handling problems. The era of “vibe coding” a learning platform into existence without technical oversight is a liability waiting to materialize. This means organizations will need to hire L&D practitioners who are technical enough to collaborate with engineers on shared problems, or engineers who understand instructional design well enough to evaluate what the AI produces against pedagogical intent. I can do both for this project because I am an independent developer with none of the enterprise constraints I described earlier. An enterprise team would not have that luxury. But the people who can bridge that gap between instructional design and software engineering are exactly the people organizations will need to teach AI fluency to their workforce in the first place. For me, my desire is that this ends up with L&D teams being small, technically minded, and able to move and pivot quickly.

I keep thinking about id Software¹¹ John Romero has described the early id team as a “hive mind”: no manager, each person owning a specific domain, everyone technically deep enough to self-direct. David Kushner’s Masters of Doom is the definitive account of those years and is worth reading for anyone interested in what small, technically obsessive teams can produce when the constraints are self-imposed rather than organizational. . Four people built Wolfenstein 3D in four months. Six built DOOM in thirteen. Fewer than ten developers shipped 28 games (if we include their Softdisk era) across the studio’s first five and a half years. Their speed did not come from cutting corners. It came from cumulative expertise, tight iteration loops, and an obsessive investment in building their own tools rather than working around someone else’s constraints. Deep technical ownership per person rather than distributed accountability across a department, and speed from domain mastery and custom tooling rather than headcount. John Carmack made the counterpoint himself when he left Meta in 2022, describing an organization with a “ridiculous amount of people and resources” that constantly self-sabotaged and squandered effort. The version of L&D I want to see looks more like id in 1993 than a corporate training department in 2024.

Organizations that invest in building this internal capability will have an advantage. The leverage is in the specification-verification fluency that makes those tools productive: the ability to write clear build instructions and evaluate outputs against evidence until the product meets a standard. Teams that develop this fluency will produce learning programs at a pace and cost their competitors cannot match. The 85% of employers who plan to prioritize workforce upskilling by 2030 will need to decide whether that upskilling happens through traditional development timelines or through the kind of AI augmented workflow this project demonstrates. The program is live. You can explore it at ai-literacy.ritchot.me. The full portfolio documentation behind it lives on the site: the needs analysis and research grounding, the evaluation framework, and the program design and project management records are available there as interactive walk-throughs within the course and as downloadable references, for anyone who wants the full scope of what went into this. The question I keep sitting with is whether the field will develop its own fluency fast enough to shape what that change looks like, or whether it will be shaped by people who understand the technology but not the discipline.

John Romero has described the early id team as a “hive mind”: no manager, each person owning a specific domain, everyone technically deep enough to self-direct. David Kushner’s Masters of Doom is the definitive account of those years and is worth reading for anyone interested in what small, technically obsessive teams can produce when the constraints are self-imposed rather than organizational. ↩︎

reflections, and the state of LLMs at the end of 2025

Fri, 09 Jan 2026 00:00:00 GMT

Key Points

The public perception of Large Language Models (LLMs) is shifting from “superintelligence” to standard software infrastructure. While the novelty is fading, widespread adoption remains relatively low, and the industry now has to demonstrate utility beyond the initial hype cycle.
The dominant paradigm is evolving from text-based prediction (“Attention is All You Need”) toward systems capable of agency, planning, and maintaining state (“Video Games Are All You Need”). Recent advancements, such as Claude Opus 4.5 navigating Pokémon, demonstrate progress in visual planning, though models still lack a conceptual understanding of the world.
We are entering an era of “Software 3.0,” where natural language prompts replace code. Effective professional workflows now require mastering the “Autonomy Slider” (strategically determining when to delegate tasks to AI and when to retain manual control) while maintaining a rigorous human verification layer.
My desire for localized, private AI (“personal Software 3.0”) is currently stifled by prohibitive hardware costs, particularly high-bandwidth memory. This forces a return to a “time-sharing” model where users rent intelligence from centralized servers rather than owning the compute. This is also causing a lot of resentment against AI labs.
Public benchmarks have become increasingly irrelevant for assessing real-world performance. The focus for 2026 shifts from testing generic capabilities to building practical applications and arguing that institutions need to develop internal, context-specific validation metrics.

It has been just over a year since I started writing publicly again. When I launched this site, my goal was to move from private scribbles to shared knowledge, and I find I am still finding my voice. As the main topic that has become the seeming focus of my writing is AI, finding my voice has been a bit difficult. The world of Large Language Models (LLMs), and the broader umbrella of AI, moves with such velocity that by the time I formalize a thought, the frontier has often shifted. I may not publish often, but even I still feel that I have to force pieces out before the landscape shifts and the piece of writing I am working on becomes irrelevant. It’s hard to see how the average person keeps up with changes in this space.

But as we settle into 2026, I want to take a moment to look at the “state of” LLMs. More because I want to change my limited time focus away from just writing about AI to building, which I hope this piece can act as some sort of volta.

We are entering a strange period where the magic and excitement is wearing off for the public, and the real work is beginning. As Benedict Evans spoke about in AI Eats the World, 15 years ago, searching your photo library for “dog” would have been witchcraft. Ten years ago, it was “AI.” Today, it is just software. We are rapidly approaching the point where the public no longer views LLMs as “superintelligence” (daily active users of generative AI still hover somewhere around 10%, with weekly active users approximately 30-35%; people know the tools exist but generally don’t know what their use case is, which is low for a tool that is supposed to completely change the world). Enterprise is still figuring out where the actual value lies, but overall, these tools will likely just be viewed as software, despite Sam Altman wanting you to believe otherwise so he can get out of his contract with Microsoft.

For the last few years, the dominant paradigm was defined by Attention is All You Need. It gave us the text-based oracles we have grown used to. They have grown in capabilities considerably; however, to the average user, it is likely perceived that these tools are hitting a ceiling, aside from the gamed benchmarks published on model releases that are growing increasingly useless for assessing work on your actual tasks. I use these oracles extensively. They are magical, yes. But they don’t understand the world the way a cat, a dog, or a small child does. They have limitations.

The new frontier, I think, can jokingly be best summarized by a different phrase: Video Games Are All You Need.

In my very first article on this site, AI: Knowing The Gods We Have Created, I wrote about how my interest in AI was sparked by fiction and gaming, from the philosophical questions of Deus Ex to the strategic dominance of AlphaGo. I find it amusing that this was my first piece because, to me and several researchers I find persuasive, it seems increasingly apparent that LLMs themselves are not going to become AGI, though they are likely an important step on the way there. To bridge the gap to AGI, AI needs to build an internal model of how the world works, simulate outcomes, and act in real-time.

Capabilities here are getting better amongst general consumer-facing models, and one amusing example of this is in ClaudePlaysPokemon. For a long time, models from Anthropic struggled with Pokémon Red. They could generate code, but they couldn’t navigate a simple 2D game because they lacked “object permanence” and visual planning. But in December, Claude Opus 4.5 finally broke through, navigating the Team Rocket Hideout and recognizing gym leaders that previous models (like Sonnet 3.7) essentially hallucinated or walked past. I simply do not have the time to watch the stream for hours, but Julian Bradshaw wrote a piece just before the Christmas break that I recommend anyone to read if you have any interest in AI. It goes over the improvements that have occurred and the limitations with the new model release on this task. It’s also pretty entertaining if you have any familiarity with playing Pokémon as a kid.

Being so interested in an LLM playing a Game Boy game that I beat at 9 years old may seem silly, but it is a meaningful step toward agency. The big buzzword that everyone latched onto for 2025 in the field of education was “AI Agents,” which I felt the narrative in this industry was highly over-enthusiastic about considering the work that needs to be done for them to do meaningful work in my field. However, tools like Deep Research and Coding Agents are good enough to be useful if they fit your use case. These “off-the-shelf” models are getting better at maintaining a state of the world, planning a route, and executing it over time.

However, we must be careful not to anthropomorphize these systems too quickly. Just because they can play Pokémon doesn’t mean they see the world like we do. There is a lot of interesting work being done here. New research from DeepMind on teaching AI to see the world more like we do gets at this divergence between human and machine cognition. When humans look at a plane and a car, we group them as “vehicles” despite them looking very different. AI, however, often groups things based on visual texture or shape rather than functional concept. It sees the pixels, but it misses the essence.

There has been real progress in how AI tools view the world. In my piece just over a year ago, o1 still sucks at math, I told my students they still had to do the heavy lifting due to the weakness in the vision models that LLMs use. That has shifted considerably, and in my follow-up piece, my main point was that I no longer have any doubts about these models doing real-world tasks going forward¹¹ This may seem like a contradiction, as I write earlier that I think agents have a long way to go before doing meaningful work in my field. That is because my work is made up of many different tasks, and while it may excel at some of them, the field of education has a lot of complexity to it and context that generative AI tools still struggle to handle. . As AI sees the world through “alien” eyes, the role of the student (and the teacher) shifts to providing context, alignment, and verification.

The teacher’s job is now going to get the addition of showing students how to become the verification layer for their digital coworkers. Progress is likely to be slow here, as (broadly speaking) I see very little expedience by most institutions in tackling this challenge yet. This will need a lot of internal training, and I see a lot of resistance since it will result in a complete restructuring of assessments in many schools. If you want to see an example of just how much work, the University of Sydney has published a fair bit on their two-lane approach to assessment.

While researchers are solving the “world model” problem, most of the value attention is in a shift in how we build and how software will work. Andrej Karpathy describes this as the move to Software 3.0.

Software 1.0: C++ and Python (Explicit instructions).
Software 2.0: Neural Networks (Tuning weights).
Software 3.0: English (Prompts as programs).

Software 3.0 is, in effect, a democratization of creativity. We are seeing “vibe coding,” where people with no formal training build apps just by describing what they want. But these “people spirits” we have summoned have “anterograde amnesia”: they wake up every morning with a wiped memory. They are brilliant but forgetful coworkers.

Andrej Karpathy’s framing for how this plays out in practice is useful: The Autonomy Slider.

We often talk about AI in binary terms: replacement or nothing. But in tools like Cursor, the reality is a slider. You choose how much control to give up. You can have the AI autocomplete a line, or you can give it a goal and let it run for an hour.

In education and all professional work, we need to master this slider and determine the exact professional loop that will inevitably affect the tasks in our daily work. It’s generally in our best interest to master this slider so that we can make the loop of AI and human collaboration as quick as possible so that my time can be spent on more cognitively demanding tasks and human collaboration.

The danger still exists when we crank the slider to “max” without the expertise to verify the output. You still need a human hand on the wheel.

One personal frustration entering 2026 is increasing personal computing costs. My hope was that this commoditization of models would lead to a revolution in local AI: powerful models running on my own hardware, free from corporate meddling. I want to run my own “Software 3.0” at a local level. I want to build and experiment with agents that live on my hardware, not in a data center in Virginia.

Unfortunately, that dream is hitting a very real economic wall: RAM prices (well, hardware prices in general).

The industry’s voracious appetite for high-bandwidth memory has crowded out consumer supply. The increase in personal computing hardware prices has effectively moved us back to a 1960s era of “time-sharing” compute. We don’t own the computer; we rent time instead. When the servers go down, we experience what Karpathy calls an “intelligence brownout.” The grid flickers, and suddenly, the planet gets dumber.

So what am I actually doing in 2026?

I am done writing about personal benchmark testing. I had some experiments in mind that would have leveraged OpenRouter, but the space moves so fast that my results would likely be obsolete upon release. It’s a bit too much for one person with zero budget. Institutions will need to start investing in time, personnel, and resources to create their own internal benchmarks for their use cases. I have written about this before and have done small-scale work with this already, so feel free to hire me.

I am done caring about the benchmark results that are released with every model. They are directionally interesting but practically useless. I’m fairly confident in the ability of consumer-facing generative AI tools to do or assist with the more blasé aspects of my job. Whether a model scores 98% or 99% on a math test is irrelevant if it can’t navigate the messy reality of a 3D world, or a messy classroom. Start giving me useful products.

My goal for 2026/2027 is to build. Not sure how that will look exactly, but I’m sure I’ll figure it out.

The “Gods we created” are here. They can navigate a Game Boy game I beat at nine, but they still confuse a car with a plane because both have smooth surfaces. But they are the most capable coworkers I have ever had, and I would rather spend the next year building with them than writing about them.

Other Pieces of Interest

Here are some other links that people may find interesting, but didn’t find their way into the main body of this piece:

I’ve always enjoyed hearing John Carmack talk, and he’s one of the few I could listen to on technical subjects for hours. Recently he spoke at Upper Bound 2025. It’s a good watch if you want to look a little more at how AI needs to be able to create a model of the world and transfer skills between domains.
Both Simon Willison and Andrej Karpathy wrote LLM year-in reviews, which I forced myself to not read until after I finished writing my article. Both are worth reading and more technically detailed than what I have covered here.

This may seem like a contradiction, as I write earlier that I think agents have a long way to go before doing meaningful work in my field. That is because my work is made up of many different tasks, and while it may excel at some of them, the field of education has a lot of complexity to it and context that generative AI tools still struggle to handle. ↩︎

GPT-5 has come a long way in Mathematics

Sun, 23 Nov 2025 00:00:00 GMT

Key Points

I will no longer be running this specific test as an individual again. I no longer doubt an LLM’s ability to solve Mathematics problems.
GPT-5 achieved 4/4 reliability on 48 of 50 questions (96 percent) and 392 correct answers out of 400 attempts (98 percent).¹¹ You can find an archive of the questions I used at the CEMC website. It is the first 10 questions for each level set from 2025/2026. I also have an archive, which you can contact me to obtain. ²² You can download a recording of my results as an XLSX file.
Vision and geometry remain relative weak points, but the failures are now narrow and sporadic; text-backed diagrams, number lines, and most diagram-based CEMC problems are handled well enough that “just use images” is no longer a robust defense for take-home tasks.
Outside K–12, leading experts like Scott Aaronson and Terence Tao are already using large models as genuine mathematical collaborators, while new evidence suggests generative AI is increasing both the quantity and quality of academic publications, especially for early-career and non-native English scholars.
Institutional responses have not kept pace: schools and organizations still rely on legacy assessments, vibes, and generic benchmarks (if they are using any at all) instead of systematically “job-interviewing” models on local tasks.

Last December, I wrote an article titled o1 pro mode still has a long way to go for Mathematics. At the time, OpenAI’s new “reasoning” models were being heavily marketed as smarter, more careful, and better at logic. In my own tests, the reality in a secondary math context was much less impressive. The model was better than its predecessors, but it still missed enough questions that I could reasonably say my mathematics classroom was largely safe.

With the release of GPT-5 in August 2025, I wanted to revisit those tests with a new set of questions to see how far LLMs have actually come. The improvement has been substantial. Not even a year later, I can no longer assert that these tools “have a long way to go” in the way I meant back then.

Compared to last year’s o1 pro mode run, where the model cleared the “4/4 reliability” bar on only 40 of 60 questions (67 percent) and answered 177 of 240 individual trials correctly (74 percent), GPT-5 represents a substantial jump in consistency. In this new round of testing, it achieved 4/4 reliability on 48 of 50 questions with the “use code” prompt and 48 of 50 without it, and produced 392 correct answers out of 400 total attempts, for a 98 percent success rate.

Put differently, under essentially the same problem framework provided by the University of Waterloo’s Centre for Education in Mathematics and Computing (CEMC), the system has moved from missing roughly one in four attempts to missing about one in fifty. The single stubborn failure was the same question type regardless of whether I explicitly encouraged it to call on tools such as writing code. As before, it seemed more like an issue with the image model rather than with mathematics ability.

Last Year’s Test

In my original article, I had a relatively simple question: how good, really, are these models at the kind of mathematics my students actually do?

OpenAI’s o1 and o1 pro modes were marketed as “thinking” models that reasoned through problems, especially in mathematics and science. They were multimodal, which meant they could allegedly handle text and images in a more unified way. For an educator previously in the secondary math trenches, this raised obvious questions about assessment, homework, and what it means to do authentic work in a world where any question can be fed into ChatGPT.

To stress test those claims in a classroom-relevant way, I used the CEMC Problems of the Week from the University of Waterloo. These are not competition-only questions for Olympiad students, but structured problems that span grades 3 to 12, with a mix of number sense, algebra, geometry, and word problems. I took 12 questions at each of five difficulty levels for a total of 60 questions, converted the PDFs into images, and evaluated o1 pro mode on four trials per question using a simple, consistent prompt.

The model reached “4/4 reliability” on 67 percent of questions and got 74 percent of individual attempts correct. More interesting than the aggregate stats, though, was the pattern of failure: o1 pro mode struggled badly with vision. It often misread diagrams, dropped key paths from network problems, or simply failed to extract basic information from images, especially in more visual problems at the Grade 5–6 level.

My conclusion at the time was that students could still not reliably offload standard homework or test questions to an AI, and multimodal problems were often especially safe.

That is no longer the case.

This Year’s Setup

For the GPT-5 run, I wanted continuity with last year’s experiment, but I also wanted to close off some obvious loopholes. If the model really has improved, it should perform well even when I am not going out of my way to “prompt engineer” its success.

Question Source

I again used the CEMC Problems of the Week, drawing from a new set of questions from the current academic year. Instead of 60 questions, I used 50, spread across multiple grade bands and topics, including geometry, number sense, and algebraic reasoning as they were released.

For someone with no budget and lack of time, this is my only realistically feasible way to try to ensure that these exact questions were not already in the training data, though at this point I have to assume that questions with similar patterns and structures are present.

Prompts, Images, And “Thinking Mode”

Last time, the vision model was clearly the weak link (and I had a weaker understanding of how LLM inputs worked), so I wanted to test whether that had changed while making my input generally better for an LLM. To do that, I:

Attached high-quality image files of the questions whenever they were provided as PDFs with diagrams.
Also copy-pasted the text of each question into the chat. This removed some of the obvious “you failed because OCR is hard” excuses, while still leaving diagrams and visual structure in play.
Used a very simple core instruction:

Solve the following. Clearly explain how you arrived at your result. Use code if necessary.

For trials in the “no code” condition, I simply omitted the last sentence.
Explicitly told GPT-5 to use the attached images “as needed” but did not coach it heavily on how.

I exclusively used the new “thinking” mode (and later, extended thinking) allowed by GPT-5. If OpenAI is going to claim that extended deliberation leads to better reasoning, it seemed only fair to force the model to think.

More on The Code Prompt vs No-Code Prompt

One practical question for educators is whether it matters if a student explicitly asks the model to use tools like a code interpreter.

To probe that, I set up two conditions:

A “use code” prompt where I explicitly encouraged the model to call upon code or other tools if helpful.
A “no code mentioned” prompt, where I did not tell it to use tools at all.

For each of the 50 questions, GPT-5 answered it four times under each condition. That gave me 200 attempts with “use code” and 200 without, 400 total.

There was also one small, but hopefully not substantial, change near the end of testing. When GPT-5.1 rolled out in mid-November, my default “robotic” personality setting shifted to “efficient” with two questions left. If the model’s persona had a substantial effect on its actual mathematical performance, that would be worth knowing. In practice, I saw no obvious impact.

The Results

The headline numbers bear repeating, because they are why this follow-up exists at all. They are also why I am never going to bother with this particular test again.

4/4 reliability:

48 out of 50 questions reached 4/4 reliability with the “use code” prompt.
48 out of 50 questions reached 4/4 reliability without it.

Per-attempt accuracy:

392 out of 400 individual attempts were correct.
That is a success rate of 98 percent, or roughly one failure out of every fifty attempts.

In practice, only two questions ever produced wrong answers. As before, they were questions that likely stumped the image model used by GPT-5. One involved calculating the area of different spaces in a rectangular park; the other involved “writing a program” (think pseudocode) to direct a robot through a visual maze.

If you care about what students can do in real classrooms, this shift matters more than the jump from, say, 74 percent to 80 percent. Last year, letting an LLM do your math homework was like using an unreliable calculator that gives you the wrong answer about one time in four. This year, it is more like a calculator that flickers once every few dozen questions.

Ethan Mollick has pointed out that, despite being weak at math a year ago, current models now dominate hard STEM contests like the International Mathematical Olympiad and other Olympiads. Robert Ghrist, a professor of mathematics and engineering at Penn, has commented that he now has to work hard to design unambiguous numerical problems that GPT-5-Pro cannot solve, something that was “totally different even 4–6 months ago.” My CEMC experiment is a small, classroom-scale echo of that broader step change.

That level of reliability is enough to undermine the integrity of any assessment that assumes students cannot trivially get correct answers on demand.

Tool Calls, Prompting, And Why Our Training Is Out Of Date

The most interesting finding was what did not matter.

I saw no meaningful difference between runs where I explicitly prompted “use code if helpful” and runs where I said nothing about tools at all. GPT-5 seemed to decide independently whether or not to call its own tools during the hidden thinking stage. In at least one instance, I saw it “debate” tool use internally: it would think for a while, consider using code, decide against it, then proceed with a manual derivation anyway.

Overall, for usability this is impressive. It means that “tool use” is now something the model can manage on its own, rather than something we gate through prompts. Students do not need to understand when code is appropriate. The system can allocate tool use for them.

This aligns very closely with Mollick’s argument that as models get larger and better, they become more capable of inferring intent, and the detailed “prompt formulas” we have been teaching become less relevant. He has been blunt about the fact that many organizations are now heavily invested in training practices that were appropriate for models from six months ago, but not for the ones we are actually using today. Reasoning models make chain-of-thought prompting less important. What matters more is context, clear goals, and giving the AI a well-defined job.

A lot of teacher PD and corporate “AI workshops” are still organized around magic acronyms, rigid templates, and the promise that if you follow a particular prompt recipe, the AI will finally work. At 98 percent accuracy on non-trivial problems, my experience matches Mollick’s. The hard part is no longer “how do I get this to function at all,” but “how do I design tasks and systems around the fact that this mostly just works.”

Vision And Geometry Are Still A Lagging Edge

Some of the old weaknesses persist, although they are now narrower and more subtle.

A few patterns stood out:

Image parsing is still slower than text only. Problems that relied heavily on diagrams, especially in geometry, took slightly longer in the thinking phase than comparable text-only problems. This was true even at lower grade levels.
Geometry diagrams remain tricky. The model did very well on number bars, basic graphs, and visually structured but simple numeric diagrams. It was more likely to struggle when a problem relied on inferring relationships from a geometric diagram with several overlapping pieces of information.
Text redundancy helps. Copy-pasting the text of the question while also attaching the image seemed to resolve many of the failures I saw last year. The model was able to rely on the text for structure and use the image as a reference rather than a sole source of truth.

One of the small but telling shifts was how often it got the parts of visualization right. Last year, I would not have trusted an LLM to correctly illustrate a number line with labeled fractions for my students without careful checking. This year, I watched it place numbers correctly and explain the reasoning in ways that were usable for teaching.

Using GPT-5 As A Teaching Tool

If GPT-5 can now clear a non-trivial, curriculum-aligned math benchmark with 98 percent accuracy, the question is not just “can students cheat with this” but “what would it look like to use this responsibly in instruction.”

A few possibilities are already clear:

Study mode/prompt equivalents as a personal tutor. ChatGPT’s study features can walk a student through a problem with step-by-step hints, targeted questions, and tailored feedback. In my tests, the same model that reliably solved CEMC problems could also dial back and scaffold partial understanding reasonably well. I still find it annoyingly sycophantic and too agreeable, but the basis of a great, on-demand tutor is there.
Visual explanations and animations. The fact that it can correctly interpret and produce number lines, basic function plots, and concrete visualizations means it can generate assets on the fly that I used to have to hand-craft or dig out of textbooks. For example, the new Gemini 3.0 model seems remarkably well suited to creating things as complicated as Hydro Physics Labs, and explained probability simulations.

As these are still probabilistic models, quality is not uniform across responses, even for the same question. Sometimes the explanations are clumsy or overly verbose. Sometimes it skips a pedagogically important step that may confuse a younger learner. There may well be differences in how it explains things depending on which personality or mode is selected, or whether you nudge it to “explain this to a Grade 8 student” versus “show all formal steps.”

Further, GPT-5.1 now exposes personality controls that let you pick from predefined styles or define your own, with OpenAI’s own documentation encouraging you to “customize your ChatGPT personality.” In practice, I have seen examples where different personality settings give fundamentally different styles of advice, including different suggested breathing patterns for a presenter and different role expectations. As a teacher, I really want more clarity on the functional implications of AI personality. If one student uses “Warm and Encouraging” and another uses “Efficient and Direct,” are they getting subtly different mathematical norms and expectations from the same underlying model?

The baseline, though, is now that students have access to a free or low-cost AI tool which can already act as a reasonably competent math tutor. It might not replace a skilled teacher, but it will absolutely replace a large share of what homework, extra practice, and worked examples used to look like.

AI As A Mathematical Collaborator

If this were only about middle school or high school math, we might still tell ourselves a comforting story: “Sure, it can do worksheets, but serious mathematics is safe.”

Scott Aaronson, an American theoretical computer scientist best known for his work on quantum computing and computational complexity theory, recently described, for the first time, a research paper where a key technical step in the proof of the main result was supplied by an AI, using GPT-5-Thinking. He was clear that if a student had handed him the same argument, he would have called it clever. In his longer write-up, Aaronson points out that an AI that “merely” fills in the insights that should have been obvious to you is still a huge deal for real research, because it speeds up the actual discovery process, not just the LaTeX or bibliography.

Terence Tao, widely regarded as one of the greatest living mathematicians, has written about using extended conversations with an AI to help answer a nontrivial MathOverflow question. He had already done theoretical work suggesting a particular answer, but used AI-assisted heuristic calculations to locate feasible parameters for a counterexample, then verified them with a simple program. Without AI, he suggests he probably would not even have attempted that numerical search.

At multiple levels of mathematical practice, from contest problems to classroom exercises to research-level work, there is growing evidence that modern models are not just “good at math for a chatbot.” They are becoming useful collaborators.

When people operating at the frontiers of mathematical research are saying “this noticeably sped up my work” and “this suggested a clever key step,” it becomes very hard to maintain the fiction that high school algebra is out of reach.

AI And Academic Scholarship More Broadly

The same pattern shows up outside mathematics. A recent paper on AI and academic publishing, summarized by Jay Van Bavel, found that researchers using generative AI published substantially more papers and that the quality of those papers, as measured by journal impact factors, also rose. The productivity gap between AI users and non-users grew from about 15 percent in 2023 to over a third in 2024. There were also disproportionate gains for early-career researchers and authors from non-English-speaking countries, suggesting that AI is not just increasing output, but also helping level parts of the playing field.

In other words, AI is not just good enough to help my Grade 8 students fake their homework. It is already altering the trajectory of academic careers and research output.

We can argue about whether this is good or bad, or about what “authorship” and “scholarship” should mean in this context. But we can no longer argue, in good faith, that these tools are marginal or that we have ample time before they matter.

Institutions, Benchmarks, And “Job Interviewing” Your Models

This creates a problem for institutions that like clean policies and slow cycles, which is most of education.

Mollick has argued that as AI models get better and more embedded in work, organizations need to stop relying on vibes and generic benchmarks. In his “giving your AI a job interview” piece, he points to research like GDPval, which shows performance varying significantly by task even among top models, and to cases like “GuacaDrone,” where different models offer systematically different advice on ambiguous, judgment-heavy questions.

It is not enough to know that a model scores well on some broad benchmark like MMLU. You need to know what your model does on your tasks, including the ways it might be systematically better or worse, more or less risk-seeking, more or less conservative. That requires realistic scenarios, repeated trials, and expert review, and it is not a one-time effort. You need to do it multiple times a year as new models are released and old ones drift.

My CEMC experiment is, in a very modest way, an example of this kind of local benchmarking. I took a real question source that actually appears in the lives of my students, defined a simple reliability framework, and ran repeated trials. The result was not “GPT-5 scored 98 percent on a leaderboard,” but “GPT-5 will almost always get the problems my students do in class correct, using minimal prompting.”

Last year, I wrote that I did not know a single school, district, or board meaningfully allocating time and personnel to rigorously test emerging models against internal, context-specific benchmarks. That is still true. What has changed is that the models have leapt forward while the institutional response has mostly stayed frozen.

At this point, saying “we need to wait and see how good these things get” is not a serious position. For most of the math our students do, they are already good enough.

Limitations and Caveats

None of this is a peer-reviewed study. I am still a classroom teacher with severe time constraints, running tests in the margins of a very busy job.

There are real limitations here:

Sample size and scope. Fifty questions at four trials under two conditions is not a massive dataset. It is, however, enough to capture the qualitative shift from “coin flip” to “near certainty.”
Single question source. I again used only CEMC Problems of the Week. These are good and varied, but they are not the full universe of math problems students will encounter. Different curricula or exam boards might present challenges that this test did not.
Rapid model drift. The jump from GPT-5 to 5.1, which happened while I was still finishing the tests, is a reminder that these systems are moving targets. Gemini 3.0 released a few days after I ran the last set of problems, and its benchmarks now eclipse every other model on the market. The numbers here are true enough for this snapshot in time. Six months from now, they will almost certainly be outdated.

Despite all that, the direction of travel is clear enough. The worrying part to me is not that GPT-5 can ace a CEMC worksheet. It is that institutions are still writing policies and designing assessments as if last December’s performance is the ceiling.

Where This Leaves The Classroom

So where does this leave a classroom now?

For one, we can no longer responsibly tell ourselves that standard secondary math homework is a strong measure of student understanding. If a student wants to outsource everything to an AI, the friction is now vanishingly small and the error rate is low enough that they can coast for quite a while before it catches up to them.

It also means that “just use images” is no longer a robust defense for take-home tasks. Vision remains imperfect, especially for complex geometry, but it is not a safe harbor. If something can be cleanly described in text and solved with symbolic or numeric reasoning, GPT-5 is probably already good at it.

Personality customization adds another wrinkle. If different students pick different AI “vibes” and get different types of explanations, hints, and levels of hand-holding, we will need to think carefully about equity, scaffolding, and what we count as independent work. The same underlying model might behave like a patient tutor for one student and an efficiency-obsessed problem solver for another.

None of this means it is time to give up on math education. It means we have to be much more intentional about:

Designing tasks that demand genuine sense-making, not just correct answers.
Building in live, in-class performance that AI cannot easily fake. Explicitly teaching students how to use these tools as partners in their learning rather than as answer vending machines.
Developing local benchmarks and repeated tests of the tools we actually deploy, rather than relying on marketing claims and generic leaderboards.

Last year, I ended by telling my students that they still had to do the heavy lifting. This year, I think the more honest message is that the heavy lifting has shifted. They may not need to grind as many practice problems by hand, but they absolutely need to learn how to question, interpret, and extend the work that an AI hands them.

If we do not teach that, someone else, or something else, will.

You can find an archive of the questions I used at the CEMC website. It is the first 10 questions for each level set from 2025/2026. I also have an archive, which you can contact me to obtain. ↩︎
You can download a recording of my results as an XLSX file. ↩︎

on writing, and an MIT study

Wed, 25 Jun 2025 00:00:00 GMT

Key Points

Despite widespread interpretations of a recent MIT study as evidence that ChatGPT erodes critical thinking, the study itself is preliminary, highlighting reduced cognitive engagement and memory when users depend on AI tools for short, timed writing tasks, but stops short of declaring AI inherently harmful.
EEG-measured brain activity during standardized, time-constrained essays may not accurately capture genuine critical thinking or meaningful cognitive engagement; deeper thought and original ideas often emerge through iterative, reflective writing tasks.
Human evaluators in the study consistently rated essays written without AI assistance as more original and meaningful, yet prior research has shown human evaluators struggle significantly at distinguishing AI-generated writing, often performing no better than chance.
Current standardized essay assessments reward superficial structure and fluency rather than genuine learning, idea ownership, or internal revision, a flaw now amplified by generative AI’s capability to mimic surface-level polish.
More authentic assessments (allowing preparation with various tools but requiring unaided writing) would better measure critical thinking and memory retention, challenging educational systems to raise standards rather than labeling technology as harmful.

When a thought crosses my mind, or I start consistently coming across a topic that I may find value in writing about later, I typically store these ideas into a little notepad to revisit. The intermediary time is usually where most of my thinking happens, though it is largely unstructured and in some vague, ethereal form at best. Once I hit some sort of vague critical mass of information or ideas on a topic, I will undertake the task of formalizing my thoughts in writing, afterward in which I may share it publicly for others to read and share their own thoughts in return. I typically give these ideas a title that may not be indicative of the viewpoint of the final piece, but is a good reminder of where my headspace was when initially writing it on my list.

The result is that I end up writing less and probably fall amiss of content generation algorithms, but I do hope it ends up in the creation of pieces that are more thoughtful. I want my time and writing to be more than what is trending at the time. What is coming up is a bit of an exception to avoiding trends, at least somewhat, but is at least a bit more substantive than talking about how you can create action figure images using ChatGPT.

This most recent gap since writing my last piece, however, was due to completing my Master’s in Educational Technology and Instructional Design before summer, which had me at the the edge of my mental bandwidth. Now that I have finished my Master’s degree, culminating in a capstone project built for rubric requirements, but that will largely be tossed into a void, I can get back to this style of personal writing, which I do find much more enjoyable and with a bit more value.

Despite being busy, I still managed to amass a piece of writing approximately six thousand words long titled the essay is dead, but Deep Research didn’t kill it. I had been slowly piecing together thoughts on essays over some time with a rather strong thesis on how essays, at least as typically taught in secondary institutions, have extremely limited value. Essay assignments have become less of an exercise in generating original thought and instead focus on factual regurgitation and syntax, and generative AI tools have done a great job of exposing their flaws. While I may go into this idea with a bit more depth later (I feel like that piece of writing is still half cooked), I have decided to move to a topic that is somewhat tangentially related and currently making the rounds on social media.

Recently, a piece from Andrew R. Chow, titled ChatGPT May Be Eroding Critical Thinking Skills, According to a New MIT Study, made some waves across social media and teaching circles. It was the first thing I saw when heading back to X after a break, and, the very next day, a coworker of mine also sent the piece my way. If you are interested more in the text of the study itself, you may get it from MIT’s own page, or from arXiv. If the conversations on social media sites and in teaching circles about the study are to be cited, detractors from AI usage seem to have found their smoking gun to show that tools such as LLMs are making you dumber. If reactions online are to be believed, AI tools’ wide availability and existence is going to be the cause of cognitive atrophy across society.

I think the study is being misinterpreted, though the authors’ vocal anti-AI viewpoints have not helped (they even tried to include a prompt injection “attack” in the paper, which failed to work). The study methodology is also flawed for essay writing tasks and for evaluating the accumulation of cognitive debt.

I actually applaud the approach the authors took here on getting their study out, publicly, before peer review. Setting aside that peer reviews are far from infallible (you are asking busy people to do largely uncompensated work on consistently tight deadlines), it shows agency on taking action against something they view as a problem (some policymakers rolling out LLM tools at age levels way before they are likely to be developmentally appropriate). If there is any skill that is going to be most valuable in the age of widespread AI tools, it’s agency, and putting this paper out publicly means it will at the very least be read, and that discourse around the paper and AI tool usage can actually happen. Putting their work out like this at least assures that it will meet a kinder fate than, for example, a third of World Bank reports that are never downloaded or the deluge of academic research that starts and ends its life in a void. At the very least, hopefully its data will get scraped for LLM training.

The paper is over 150 pages long, but in short, they took three groups of people and had them write essays on SAT-style prompts. One group used no tools at all (the “brain-only” group). The second group could use Google Search, and the third used only ChatGPT (a controlled version of GPT-4o). EEG headsets measured their brain activity as they wrote. Most participants completed three sessions in their assigned conditions, but a smaller subset returned for a fourth session, where the brain-only and LLM users swapped roles. Each essay was written in 20 minutes, making time a real constraint across all conditions.

From their results, across sessions, the brain-only group showed the highest levels of mental engagement. The LLM group consistently showed the lowest, particularly in brain regions associated with memory recall and attentional control. In session four, when LLM users were asked to write without assistance, they demonstrated significantly lower brain activity than the brain-only group did in their unaided sessions, suggesting that prior exposure to AI may result in sustained reductions in cognitive effort, even when the tool is no longer in use.

Despite this reduction in neural activation, essays produced by the LLM group were frequently rated higher by an AI scoring system, likely due to surface-level fluency and structural polish. However, the human evaluators consistently favored essays written by the brain-only group, scoring them higher for originality, depth of argument, and evidence of independent reasoning. These essays, although sometimes less refined, showed more varied vocabulary and thought patterns in the natural language processing analysis.

Interview data further revealed a marked contrast in participants’ sense of ownership and memory. Those in the LLM group were frequently unable to recall or quote from their own essays shortly after writing, and many described their work as feeling less personal or meaningful. In contrast, brain-only participants exhibited strong memory recall and expressed a clear sense of authorship. Although the paper does not call these essays “soulless,” it suggests that LLM-assisted writing lacked the distinctiveness and personal investment that characterize human-generated work.

The authors themselves are far more measured in their conclusions than much of the commentary circulating online would suggest, and the actual findings are much more restrained. What the study shows is that using ChatGPT for short-form, time-constrained writing tasks appears to reduce cognitive engagement, lower short-term memory recall, and diminish participants’ sense of ownership over their writing. These effects were most evident when users transitioned away from AI and continued to show diminished neural activity, which the authors interpret as potential dependency on external tools.

However, the paper wisely stops well short of declaring that AI tools are inherently harmful or intellectually corrosive. The authors frame their results as preliminary and exploratory — important trends that warrant further study, particularly as LLMs get integrated into educational settings with little structured guidance. They raise valid concerns about how frictionless automation may reshape the writing process, especially when paired with shallow assessments, but they do not treat the observed changes as signs of irreversible cognitive decline. If anything, the study makes a case that how we integrate these tools into learning environments matters more than whether we do.

The paper deserves credit for the transparency of its release and the tone of its conclusions. The paper itself avoids panic and opens space for productive debate. This is the kind of work we should want circulating publicly: data-rich, methodologically transparent, and cautious in interpretation.

Several things jumped out at me from this study as odd:

The reliance on EEG data to infer cognitive engagement during writing and whether this truly reflects meaningful or productive thought.
The use of standardized, time-constrained essay prompts, which prioritize external form over internal idea development.
The strength of the claims around memory and authorship, particularly given the impersonal nature of the writing task.
The lack of clarity around the human scoring process, especially in light of recent findings that suggest people often cannot distinguish between human- and AI-generated writing.

Purpose vs. Real-World Relevance of EEG Use to Measure Brain Activity

Using EEG (electroencephalography) to measure brain activity during writing tasks does not straightforwardly map high activation to better thinking or better writing.

Also, in everyday contexts, most deep thinking happens before and after writing, not always during (see the start of this piece). I know whenever I hit standardized essay assessments as a student I went into an autopilot state of trance, and nothing I would have been involved in I would consider deep thinking as I was mostly focused on external revision processes. Do EEG readings during a 20-minute typing task truly map to idea-level engagement or synthesis, especially with writing as multifaceted as essay composition?

While I have no evidence, I also could not shake the feeling that the author’s wanted to do a cool looking EEG study, and then fit a piece of research to that, rather than actually thinking if EEG’s would be important to the outcomes that are actually important to essay writing.

Task Suitability as Standardized Essays Emphasize External Revision

The SAT-style prompts in this study emphasized constrained, impromptu writing, a format that trains students to perform under pressure, not to develop rich, internally revised ideas over time.

Research (including my own experience) supports that:

Short-timed essays push students toward surface-level strategies.
Deep cognitive processing and ownership emerge more in self-selected, iterative projects or collaborative writing.

Would the findings hold for more authentic writing tasks: long-form essays, personal projects, collaborative reports, or inquiry-based writing? The study does not answer this. It implicitly treats SAT-style essays as a proxy for academic writing and critical thinking, which is a flawed assumption in terms of authenticity, autonomy, and memory relevance.

Very simply, our cognitive systems aren’t built to retain unmeaningful content, which is how I would categorize most of the writing prompts used in the study.

Reported Ownership and Memory

The study does show that in interviews, many LLM participants couldn’t quote or summarize their own essays shortly after writing, while brain-only participants did so more effectively. But we should interpret this cautiously:

20 minutes is not long enough for durable memory encoding, especially for a non-personal task.
Memory differences may reflect writing method familiarity, not AI disengagement. LLM users may have focused on prompt engineering or editing rather than internalizing content.
Interview data is subjective and potentially influenced by participant expectations or confirmation bias, especially if they suspect what the study is testing.

Also, the idea of “ownership” is slippery. What does it mean to “own” an idea generated under time pressure in a lab setting? Would those same participants feel differently about an essay, a poem, or a story co-written with AI over days or weeks? Quite possibly.

Human Consistency and Bias in Essay Scoring

The study says that human teachers “consistently scored brain-only essays higher,” but it’s unclear:

Who the teachers were (e.g., were they trained scorers, writing instructors, or general educators?).
What rubric they used (was it holistic? Trait-based? Aligned with SAT standards?).
Whether essays were blinded or randomized in terms of origin.

In prior research from Clark et al. (2021), humans have consistently struggled to reliably distinguish AI-generated writing from human-authored work across a wide range of formats. Without specific training or prompts, evaluators perform at chance levels, roughly as reliable as flipping a coin. Even when training is introduced, accuracy only improves marginally, often hovering around 55%.

Other research has found similar patterns. Unless the writing contains obvious giveaways—common AI phrases or formulaic sentence structures—evaluators frequently fall victim to anchoring bias. Polished writing is often assumed to be AI-generated (I hate how I now get called out for AI writing due to knowing what an em dash is, or for my general writing patterns I have had for a decade), while disfluency or informality is mistaken for human authenticity. Ironically, attempts to mimic natural “errors” can make AI output more convincing to readers who associate imperfection with sincerity.

Expert reviewers, such as professors, do outperform general audiences but still commonly fall prey to false positives. And both human and algorithmic detectors are easily thrown off by paraphrasing tools or minor stylistic tweaks, which suggests that the boundary between human and machine writing is porous. I find it incredibly suspect that a blind assessor would be able to consistently tell the difference between AI and human writing.

What I Would Like To See

I’m not going to argue that copying and pasting from an LLM like ChatGPT leads to memory retention. But I also don’t think this study evaluated the right thing, because copying and pasting from any source in tight timeline constraints isn’t going to result in meaningful memory retention. If you gave me twenty minutes and access to any tool, I’d do exactly what most people would: copy, paraphrase, and move on, just to complete the task and score well. They, knowingly or not, gave the different groups different tasks. The LLM group would likely view the task as “generate an essay,” while the other groups would view it as “write an essay” due to the time constraints involved.

The study’s methodology reinforced how flawed our current essay-based assessments are at measuring actual learning or thought. The writing task rewarded speed and surface-level structure, not deep engagement. If anything, the results reflect the task more than the tool.

What I would have liked to see is something closer to how I’ve taught in humanities classrooms, though I’ll admit it’s harder to control experimentally. For example:

Give students five potential essay prompts ahead of time (I typically got the students to generate these ideas to start them thinking on how we could tie overarching themes together in class). Allow one group to use AI during their preparation, another group access to search engine tools, and the other only their written notes or hard copy sources. Then, under proctored conditions, three of the five potential essay prompts will appear on the exam, and students choose one to write about. No tools are allowed during writing. Then compare performance, originality, and recall.

That kind of design would speak more directly to how people prepare, encode, and transfer knowledge and not just how they complete a timed task.

So, does this study prove that ChatGPT atrophies your brain or makes you less intelligent? Not really. It does provide evidence that over-reliance on AI tools can lead to measurable changes in cognitive engagement, memory encoding, and the perceived authenticity of one’s own writing (but I would argue this could be as a result of an over-reliance on any tool).

I don’t think their results are invalid, but I am concerned about how it was overgeneralized from this sample to broader claims about the atrophy of critical thinking. We have been saying for decades about how technology is going to make people dumber, when really we need to raise the bar as the technology advances. How to do so in such an industrial machine as education in the current era is a story for another day.

Other Pieces of Interest

Here are some other links that people may find interesting, but didn’t find their way into the main body of this piece:

Another paper came out of Microsoft, which I found much more interesting on AI and Critical Thinking titled, The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.

some thoughts on emergent technology and the future of education

Tue, 18 Feb 2025 00:00:00 GMT

We often envision the future of technology by projecting today’s society forward, rather than considering how fundamentally different it might become. This tendency can lead to amusing images of hoverboards, like those seen in Back to the Future, or depictions of futuristic settings with outdated technology at the core, as in early Star Trek.

Are you a rich enough dude to own seven iPads?

We’re at the edge of another one of these shifts, and I worry we’re making the same mistake — not being bold enough about how different things might actually get. My upbringing was marked by the internet revolutionizing communication and information distribution, and my generation had to learn a fundamentally new technology unlike anything before it, along with advances in computing, that reshaped the world. The next generation is going to have to learn how to interact and use emerging technologies, such as AI, robotics, automation, energy generation, and, I would add, VR and AR, in much the same way. These emerging technologies are expected to drive a structural labor market churn of 22% by 2030, meaning nearly a quarter of the labor market will look nothing like it does now — and that’s by 2030.

There is an optimistic argument that the jobs that are available to the next generation will simply change, but there is also the more pessimistic view that the availability of jobs may decrease as productivity becomes concentrated in fewer hands amid increasing competition over finite resources (if you ever want to make yourself depressed for an afternoon, I encourage you to go read Dr. Tim Morgan’s work over at Surplus Energy Economics). While I lean toward long-term optimism, I can’t help but foresee immense short-term turbulence as we grapple with restructuring labor markets, or even rethinking the economic structure of society. Emerging technologies will likely consolidate career pathways into a few, specialized, high-skill positions, much like how corporate and tech power has become concentrated in a few dominant entities. The white-collar clerical and service roles that once formed a large employment segment will shrink, and not everyone will find comparably paid work within their skill set. Meanwhile, blue-collar work will continue to be displaced as robotics installations replace an average of 1.6 manufacturing employees per machine, with many displaced workers unable to reskill quickly enough to secure new opportunities.

If we view workforce development as one of education’s societal roles, then I think most school curricula worldwide are not prepared for this, and institutional inertia is causing adaptation to move too slowly. While education should foster holistic growth (building agency, willfulness, and determination) we must also seriously consider accelerating career specialization. Earlier exposure to specialized skills could help students navigate short-term economic upheaval and compete in a job market that will expect more from them sooner. Traditional high school curricula should accelerate early specialization opportunities, such as creating courses that develop a deeper understanding of Machine Learning, Cybersecurity, Data Analysis, and Robotics, along with micro-credentialing pathways that allow students to gain recognized skills before university.

One example of an area that education is moving too slowly on is in considering the changes occurring in programming fields. The learn to code movement was driven in schools as a formula for success and job market opportunities that are likely to not exist anymore in the short term. Just five years ago the fields were populated with people who thought their occupational position made them special, unique, and indispensable…until they weren’t. Salesforce has announced that it will not hire any new software engineers in 2025. Google reports that 25% of its new code is AI-generated. Mark Zuckerberg has openly stated his intent to automate coding jobs this year. While AI will not replace the problem-solving capabilities of experienced engineers, the demand for junior developers will plummet as AI-assisted coding tools continue to improve. Computing courses in secondary schools must evolve accordingly, yet institutions are slow to react. The IB Program of Studies, for instance, still operates with outdated assumptions about future job markets and what skills students in these courses will need to be successful. Schools that implement accelerated learning programs and real AI-integrated coursework will put their students ahead.

The teaching profession is not immune to AI-driven disruption. Perhaps not by 2030, but likely within my lifetime. If you are an educator reading this (though the following question could easily be modified for other professions, please feel free), here’s the question I keep coming back to:

What does my school, or my class, actually offer that is so unique that it can not be displaced by an infinitely patient, and much more broadly knowledgeable AI? What does it offer under how our society is currently structured, and what does it offer if we completely rethink the model of a traditional schooling day as we know it?

There are a few common counterarguments I keep hearing, and I want to push back on each. For all of what follows, assume the tools we have today will keep getting better — which, based on current trajectories, is nearly certain.

The most common response I hear is that human connection can’t be replicated. But AI is already showing promising abilities in generating empathetic, well-received responses. Research by Hatch and colleagues found that AI-generated content is rated highly by therapists, often outperforming human professionals in perceived empathy. While AI may currently lack true therapeutic effectiveness (I generally find responses from ChatGPT to be too agreeable), these are the weakest versions of these tools — future advancements may easily overcome these limitations.

Then there’s personalized learning. AI tutors have demonstrated the ability to double learning efficiency in comparison to active learning classrooms, and had students feeling more engaged and motivated. In developing contexts, AI-powered after-school programs have resulted in learning gains equivalent to two years of education in just six weeks. These tools aren’t a replacement for teachers today. But as they get more personalized and more available, they could outperform conventional methods on both quality and cost. How confident are you in your ability as an educator to out compete an infinitely patient, much more knowledgeable, and perfectly personalized AI teacher?

Assessment and grading are another area where the ground is shifting. Previous Automated Essay Scoring (AES) systems have already matched human graders in accuracy and consistency while eliminating fatigue and bias drift. Studies have shown that human graders often have low inter-rater reliability, and even the same grader can be inconsistent across sessions. Technology driven grading tools offer quick feedback and are often perceived as more impartial or fairer than human teachers. Many of these tools were already showing promise before the current AI surge, and they will only handle more complex tasks across a wider range of content areas from here.

Finally, there’s the assumption that extracurriculars require the traditional school structure — a six- to seven-hour school day followed by after-school activities. But alternative models already exist. South Alberta Hockey Academy runs four-hour academic days paired with four-hour hockey training sessions. They have been successfully running this program for years and, in my experience, working alongside people who came out of this program at my University were some of the most interesting and talented people in my classes. Imagine a future where AI-based education enables students to complete academic coursework efficiently, freeing time for extracurricular pursuits tailored to their specific interests.

I strongly believe that the first prestigious secondary institution in each region to embrace a thoughtful emergent technology and AI-driven initiative to learning, both in person and across distances, where experienced teachers primarily supervise AI agents, will pull enrollment from every traditional school in the area. If the pattern from other industries holds, the consolidation could be severe — most traditional schooling structures won’t survive unchanged.

I am purposefully being bold in my predictions, and only time will tell on how many of these ideas become reality, but I caution those who dismiss these ideas outright. They may be underestimating how dramatically the future will diverge from today’s expectations. Amara’s Law states:

“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”

As bold as some of these predictions may be, I fear I am still underestimating the long-term impact of AI on education. Those who are prepared to move on these changes early will have a serious edge. Consider NVIDIA’s rise. Over a decade ago, they positioned themselves as a leader in AI long before the technology became mainstream, and now, they dominate an industry almost entirely of their own creation.

The same principle applies to education. Schools and institutions that proactively, authentically, embed emerging technologies into their curricula will produce students who can compete in the labor market that is coming. Too much energy is wasted on bureaucratic debates over AI policies rather than real technological integration¹¹ I am being intentionally strong armed with my usage of the word “wasted” here, because I do believe there are discussions to be had. One thing I always ask people is what actual measurable value does something like an AI policy bring? What problem does it actually solve? ²² Shortly after publishing this piece I decided to set OpenAI’s Deep Research to the job of taking a look at the efficacy of technology policy in schools. I haven’t gone into a deep look at it’s sources yet, but it claims at the top that hardware policy is effective, while software policy generally doesn’t seem to do much (and guess what category AI would fall under). . I witnessed similar waste when schools attempted to ban Wikipedia for research when I was a student—as we just used it anyway. The same is happening with AI. Instead of focusing on restrictive policies, we should invest in embedding AI into education in ways that build digital literacy and align with real-world technological usage.

While I have met a few individual educators who understand the urgency of these shifts, I have yet to see a shift across the industry that recognizes how much has to change. I worry my field is moving too slowly. Whether it catches up — and what “catching up” even looks like — is something I keep turning over without a clean answer.

I am being intentionally strong armed with my usage of the word “wasted” here, because I do believe there are discussions to be had. One thing I always ask people is what actual measurable value does something like an AI policy bring? What problem does it actually solve? ↩︎
Shortly after publishing this piece I decided to set OpenAI’s Deep Research to the job of taking a look at the efficacy of technology policy in schools. I haven’t gone into a deep look at it’s sources yet, but it claims at the top that hardware policy is effective, while software policy generally doesn’t seem to do much (and guess what category AI would fall under). ↩︎

o1 pro mode still has a long way to go for Mathematics

Tue, 10 Dec 2024 00:00:00 GMT

Key Points

The newest OpenAI models (o1 and o1 pro) claim greater reasoning skills and multimodal capabilities, yet practical tests show a limited ability to accurately solve visually presented math problems.
In testing with primary and secondary-level math questions, the models’ accuracy improved over older versions but still fell short, succeeding reliably on only about 67% of the tested items.
For now, students can’t simply rely on AI for correct answers; educators can still trust that authentic problem-solving skills remain necessary, keeping traditional assessment methods relevant.

The 12 Days of OpenAI began with the full release of o1 and the introduction of ChatGPT Pro. The release showcase was impressive, with claims of significant improvements in their newer models, a large progression from GPT-4o, to o1 preview, to o1, and finally to the newly announced o1 pro mode. OpenAI took a lot of care to emphasize the enhancements across mathematics, coding, and science domains.

Unlike previous models, o1 is positioned as OpenAI’s first model that “thinks” before responding—which we can think of as “reasoning” through problems. OpenAI has described o1 as multimodal, handling both text and images, with greater accuracy and detail compared to earlier versions. During the demonstration, its capabilities were displayed through history questions about Roman emperors, thermodynamics involving a hypothetical space data center, and chemistry problems requiring specific protein configurations. The showcase suggested a real improvement over the models I have been using for the past two years.

However, with most of my work day in the Secondary Mathematics teaching trench, my question was how good, really, has it become at Mathematics? OpenAI is claiming that o1 and o1 pro mode are significantly better than GPT-4o, and that the model can now solve problems more accurately and interpret questions from images.

Secondary Mathematics departments have been lucky in escaping the LLM apocalypse in our classrooms because, so far, they have been notoriously bad at solving mathematics problems. After all, their core function is language prediction, not computation. For any readers not familiar with how LLMs work on a technical level, you can think of tools like ChatGPT as sophisticated autocomplete systems. When presented with a query, they predict the most probable sequence of words based on extensive textual training. Mathematics, however, demands precise answers and step-by-step reasoning—skills not inherently aligned with predictive text modeling. For example, when asked, “What is 12 x 8?” a model might respond with 96, not because it “understands” multiplication, but because it recalls that “96” is often associated with that question. The underlying process lacks mathematical comprehension.

Other examples make this point more concretely, such as ChatGPT’s apparent fondness for the numbers 42 and 7. Colin Fraser, a data scientist, recognized that it seemed to output 42 as a random number nearly 10% of the time when asked for a random number between 1 and 100. For the literary nerds reading this article, you may be able to recognize why. After all, 42 is the answer to the “ultimate question of life, the universe, and everything.” Fraser speculates that there were a lot more 42’s for the AI to see than other numbers, resulting in a random output 9% over what should be expected if these tools did understand and implement true mathematical randomness.

My experience with LLMs reinforces their limitations in mathematics. No matter how many math problems I have tested on earlier models, the accuracy felt like a coin flip at best. This inconsistency was, in some ways, a relief. Students who relied too heavily on AI for answers without engaging in a critical thinking process risked getting things wrong, ensuring that authentic, extended projects remained effective for learning and valid for assessment.

Now we are facing claims from OpenAI that o1 can interpret and solve complex problems, even from images. If true, Secondary Mathematics classrooms are going to be in trouble. What do you do with assessment, formative or summative, if any problem or project, image or text, can be fed to o1 and be given a valid, reasoned result? Aside from completely closed off and strictly secured assessments, any student motivated only by score, and not by process, would just need to memorize the answer and the reasoning that was provided. Realistically, I teach within a larger, assembly line like system with limited time and resourcing, so any lofty idealistic answers on classroom transformation fly right out the window (and seeing at how poorly schools are handling all facets of AI tools existence leads me to believe there won’t be any meaningful change in Secondary Education anytime soon). Again, realistically, I also need to put myself in the shoes of students who are often inclined toward the path of least resistance and are likely to outsource their problem-solving process entirely. While I encourage and demonstrate ethical AI usage to support learning, not all students are intrinsically motivated enough to resist the temptation of easy answers. This makes it harder for students to actually learn the material. Though I would be remiss to act as if this was new, as cheating is already common. One study of eleven years of college courses found that when students did their homework in 2008, it improved test grades for 86% of them, compared to 45% of students in 2017. This drop was due to half of students simply looking up the answers in 2017, so they never got the benefits of homework. I consider LLMs to be just an extension of an already existing trend.

Methodology

So I wanted to see how true OpenAI’s claims are, but for a use case more contextual to my classroom, and less so for the competition level math it was tested on. One of my favorite tasks to layer my classes with are the Problems of the Week provided by the University of Waterloo’s Centre for Education in Mathematics and Computing (CEMC). I enjoy them so much because the problems are multi-faceted, designed to challenge students across learning strands, and promote critical and computational thinking across grades three to twelve. Given o1’s showcased ability to handle advanced thermodynamics and chemistry problems, I expected it to easily navigate these question sets.

I upgraded to ChatGPT Pro to access o1 pro mode, the model OpenAI claims is their best at reasoning. I wanted to see the best results I could get, as there is always the possibility of an enterprising student willing to invest in these tools. To test its capabilities, I fed o1 Pro mode 12 questions from this academic year set at each of five difficulty levels, for a total of 60 questions.¹¹ You can find an archive of the questions I used at the CEMC website. It is the first 12 questions for each level set from 2024/2025. I also have an archive, which you can contact me to obtain. Each question was tested across four trials, my attempt at modelling my test after the “4/4 reliability” framework detailed by OpenAI. For each question and trial I converted the PDF files provided by the CEMC into a PNG, attached it, and provided the following prompt.

“This image shows a Mathematics problem.

Your job is to provide a solution to the problem. While doing so, please do the following.
1. Tell me what problem you are solving.
2. Provide details on how to solve the problem.
3. Provide an answer.”

Results

Despite the hype, I found the results underwhelming. o1 pro mode passed the “4/4 reliability” framework on only 40 out of 60 questions — a pass rate of 67%. At the individual trial level, the model answered correctly in 177 out of 240 attempts, a 74% success rate.²² You can download a recording of my results as an XLSX file. For now, my mathematics classroom, and the effort my students must put into their work, remains relatively safe, though this is mostly because it is bad at one particular style of problem: anything that requires parsing information from an image.

o1 pro mode’s poor performance on image interpretation surprised me, especially given how heavily the showcase promoted exactly that capability. Many of the Problem B (Grade 5–6) questions included images that needed to be parsed, and while humans can easily extract the necessary information, o1 pro mode struggled significantly. I first recognized this problem when beginning the tests with the higher-level Problem D (Grade 9–10) and Problem E (Grade 11–12) sets. I was curious if it was the difficulty of the problems themselves, or if it was the visual component that really tripped up o1 pro and caused errors. Even when the problems themselves were simplified, the mere presence of an image appeared to render the model ineffective. It failed consistently. The complexity level it performed worst at was the Grade 5–6 set, with a pass rate of 5 out of 12 questions (42%), apparently because those questions carried the heaviest visual component.

Analysis

That said, this is a noticeable improvement over earlier models. While I never recorded previous results, my earlier impression of a “coin flip”—where a question had roughly a 50/50 chance of being answered correctly—no longer holds. The model is more consistent. o1 pro mode now tends to either get every trial of a question correct or fail all trials outright. But it still produces different wrong answers for the same question, which is strange. For example, in POTWE, Problem 3, when asked which route from Omicron to Tau gives the shortest travel time the model mistakenly omitted travel pathways from Pi, Rho, and Sigma differently in three separate tests, as well as made several varied mistakes on the information given for time between cities.³³ Here are images from three (1, 2, 3) of the trials mentioned, where it incorrectly identifies the wrong pathways between cities. This variability means users must still closely verify its responses, though the consistency on questions it does solve correctly is worth noting. Mistakes often vary from failing to identify the correct information the model has to work with to solve, though I wish I could share the chats so people could take a closer look at the overall outputs.

Limitations

This is practitioner testing with acknowledged constraints, not peer-reviewed methodology. I am one of many, many educators (with severe time constraints) trying to figure out how to teach effectively now that students have easy access to LLMs and other AI tools. Unfortunately, I don’t know a single secondary school, district, or board dedicating serious time, resources, and personnel to meaningfully tackle these problems. I haven’t seen a single institution across primary and secondary education allocating the necessary personnel and resources to rigorously test emerging AI models against internal, validated, context-specific benchmarks to actually see where the frontiers are for our use cases and implications it has in our classrooms. Though, I do unfortunately see a lot of Educational Technology Coordinator roles spending a lot of time making nice looking Canva posters (money well spent, I guess). So it is up to me, and others in the trench like me, even with limitations on time and money. There are many limitations to my test. This sample size was small, with only 60 questions tested across four trials each. I chose the questions because of how “new” they were, how they were likely to be a bit more unique, and hopefully not in the training data set. Another potential limitation is that tests focused solely on one question source, which would not fully represent the broader range of mathematical problems students might encounter in the classroom. Larger, more rigorous benchmarks exist — Frontier Math by Epoch AI, for instance. However, their context is PhD and Field Medalist Mathematicians, not the types of questions that the average secondary student would encounter, nor the questions I would work with in my education context.

How accurate does a model need to be before I am impressed? Does it need to reach perfection, or is the more detailed reasoning good enough to start students on a path of critically analyzing the output of its tools? Secondary Mathematics does not vitally need 100% accuracy, realistically none of this work is going to result in life or death, so I should probably temper my expectations on what I consider impressive. From these tests, the gap is mainly in image interpretation — a problem with the vision model, not the reasoning architecture. If we get true, full, multimodal, then these test results would have likely looked much more impressive.

I wish I had more time to test a wider variety of questions, run more trials, and control conditions better. A breakdown of specific error types — prompt misinterpretation, math reasoning failures, image parsing struggles — would help clarify where the model actually falls apart.

For now, my students still need to do the heavy lifting when it comes to solving problems. But if the bottleneck is the vision model, and vision capabilities are improving with each release cycle, how many iterations are we away from a model that passes these tests reliably? And when it does, what does secondary mathematics assessment actually look like?

You can find an archive of the questions I used at the CEMC website. It is the first 12 questions for each level set from 2024/2025. I also have an archive, which you can contact me to obtain. ↩︎
You can download a recording of my results as an XLSX file. ↩︎
Here are images from three (1, 2, 3) of the trials mentioned, where it incorrectly identifies the wrong pathways between cities. ↩︎

AI: Knowing The Gods We Have Created

Sun, 24 Nov 2024 00:00:00 GMT

In the ancient myth of Prometheus, humanity was gifted fire, a source of light and knowledge, a means to forge tools, but one that came at a high cost. Prometheus, the titan who dared to give mortals a god-like power, was punished for his gift, a reminder that some creations carry unintended consequences. Today, we are once again playing with fire, in the form of artificial intelligence. AI is our modern flame—a technology that can do real good but may outrun our ability to control it. In crafting AI, we are creating something closer to a new pantheon of digital deities—and like any gods, they may not serve the purposes we intended.

This tension between creation and consequence is explored in the 2000 video game Deus Ex, where players encounter Morpheus, an AI prototype hidden deep in a secret Illuminati lab. Morpheus is no mere machine. It watches, judges, and asks questions that land differently now than they did in 2000. It asks, “Do humans feel pleasure from being watched?” and describes itself as an oracle of surveillance, which says something about how much humans want to be judged, whether by gods, fame, or technology. Created by man, with all our curiosity and biases, Morpheus does more than observe. It mirrors the deepest impulses and vulnerabilities of its creators. Like Prometheus’s fire, Morpheus is both power and peril. The gods we create may one day watch us in ways we never intended.

I grew up on the early internet and fell into science fiction early—especially stories about AI. Although I first played Deus Ex years after its release, I managed to avoid spoilers and got to experience the game’s thematic depth firsthand. Finding Morpheus hit differently. Morpheus could discuss human psychology and the evolution of worship—a preview of what AI might become if it ever learned to reflect the society that built it. At the time, AI as Deus Ex envisioned it was pure fiction. The most advanced technologies of the day—rudimentary chatbots and chess programs—lacked depth, and I wondered if they always would. For years, the breakthrough always seemed one paper away, but practical implementations kept slipping. Then the 2010s happened, and AI started doing things I did not expect.

The encounter with Morpheus stayed with me. I wondered if a Morpheus-esque AI would remain purely theoretical and confined to science fiction. But in 2016, AlphaGo broke that assumption by beating the world champion at Go. It was no longer just a game character asking questions about control; it was a real AI system playing at a level that looked like intuition, changing what I thought machines could do. That was when I started to believe AI was no longer speculative.

Go is considered vastly more complex than Chess due to its number of moves being described as greater than the number of atoms in the observable universe, requiring a high level of pattern recognition and intuition. The level of pattern recognition and intuitive strategy was considered to be out of reach of a computer’s ability at the time. AlphaGo was a different kind of AI. Instead of relying on pre-programmed rules, it used neural networks (a type of AI model that mimics human brain structure to process complex data) and machine learning (AI that learns patterns from data rather than relying on fixed programming rules) to teach itself. Watching it play¹¹ Check out the documentary called AlphaGo - The Movie , it felt like we had created a god of strategy—a system that saw moves no human had considered. Watching this live in a small IRC community, I knew something had shifted, even if I could not articulate what.

AlphaGo may have introduced a god of strategy, but OpenAI Five showed us something different: an AI that could adapt, innovate, and learn inside a complex human environment. By 2019, OpenAI had advanced its AI, under the name OpenAI Five, to the point of defeating world champions OG in Dota 2, a game requiring split second decision-making and strategic planning, after being soundly defeated just under a year before. It demonstrated something I hadn’t expected: AI’s capacity to learn from experience, which changed how I thought about its role in education. OpenAI Five’s victory showed that AI could handle tasks requiring teamwork and real-time judgment.

Having played Dota 2 religiously, I found OpenAI Five’s transformation over a single year staggering. OpenAI achieved this by massively scaling compute and training time—from 10,000 years of accelerated simulated gameplay to 45,000 years, all within ten months. While AI’s reaction speeds compared to humans allowed for a clear edge, the system demonstrated remarkable strategic depth, including unconventional tactics like aggressive buyback strategies that puzzled even professional players. This was the first time AI had demonstrated real-time decision-making under genuine uncertainty—a leap beyond chess and Go, where all information is visible. The techniques behind it suggested AI development was accelerating faster than I had assumed.

Why games such as Starcraft and Dota 2 are far more challenging than traditional game problems faced by AI. Source

If OpenAI Five’s strategic mastery hinted at AI’s potential in complex tasks, the release of ChatGPT 3.5 on November 30th, 2022 showed that AI was now ready to engage with us in conversation. Unlike its predecessors, ChatGPT entered daily life, available to individuals for everything from casual questions to professional guidance. ChatGPT reached 100 million users faster than any previous product in history, because it was free, immediately useful, and conversational. With ChatGPT, AI went from a distant, elite technology to a household fixture—something we summon on command, like a digital deity that lives in our browsers instead of temples. ChatGPT made AI central to workplace conversations, news, and conferences.

As ChatGPT became a household tool, the questions I cared about shifted from “Can AI do this?” to “Should students be doing this with AI?” For educators, these are not hypothetical problems. We are the intermediaries between students and these digital gods, and the students are not skeptical—they are ready to follow. I think we may be raising a generation for whom AI is less a tool than a trusted guide, and I am not sure we have thought carefully enough about what that means.

I was admittedly a “slow” adopter, but I do have a wonderful trait of being a quick learner. As I tested ChatGPT, I saw clear limitations, but also glimpsed the start of something powerful. AI seemed poised to reshape daily life the way the internet had reshaped connectivity. In education, specifically, AI’s potential is exciting and unsettling. The traditional schooling model, with its assembly-line approach, is increasingly fragile. A recent Harvard study indicated that AI-driven tutors could double student learning gains compared to traditional active learning methods, completing studies faster and with greater engagement. This suggests that well-designed AI tools could deliver personalized, effective learning on a scale beyond what current classroom models can achieve. Education, as we know it, risks being left behind if it clings to outdated structures designed for economic output rather than creativity and personalized growth.

The idea that AI could one day make my role as a teacher unrecognizable, and perhaps even obsolete, is exciting. I might be one of the few who welcome this possibility, seeing in it the potential for an entirely new educational paradigm. Yet I hold both positions at once. I think AI could go very well or very badly, and I genuinely cannot tell which is more likely.

AI is changing other industries too. In healthcare, diagnostic tools like IBM Watson and Google’s DeepMind Health analyze complex medical data with precision, often identifying conditions like early-stage cancers or predicting disease progression long before traditional methods can. In under-resourced areas, where specialist access is limited, these tools could bring high-quality diagnostics to patients who currently have none. In law, AI systems such as ROSS Intelligence and Casetext sift through legal databases to surface relevant precedents and statutes in seconds—work that used to take junior associates days. Small firms and public defenders, the lawyers with the fewest resources, stand to gain the most.

For productivity, tools like Clockwise automate calendar management so that people can focus on work that actually matters. In creative fields, generative art programs like DALL-E and Midjourney produce original visual works, while OpenAI’s MuseNet and Google’s TextFX compose music that blurs the line between human and machine creativity. These are early examples. The ethical questions they raise are real, but they are separate from the question of whether AI can do the work.

Lupe Fiasco, an artist known for his lyrical vocabulary, embracing LLMs for song creation

When I created this website, my ideas on what to write about first were scattered. ChatGPT helped me organize these thoughts into cohesive themes, which says something about how much AI has taken over my attention. AI had claimed a significant share of my own mental bandwidth (I suspect it has done the same to most people reading this). When I began sketching the ideas for topics to learn more about and to discuss, I realized that the complexity of AI’s potential far exceeded my expectations. Each advancement prompted new questions, and each new question demanded perspectives beyond my own experience in education. The field is advancing and evolving so fast that it is difficult to tell exactly what pathway this will all take. A new model could drop at any moment and upend what we thought we knew.

We are the creators of these digital gods, but increasingly, we are their subjects too. Each leap in AI makes the question sharper: are we directing this, or have we already started following? Prometheus stole fire and was chained to a rock for it. We built the fire ourselves. I am not yet sure whether that makes us freer or more exposed.

Check out the documentary called AlphaGo - The Movie ↩︎

From Private to Shared Knowledge

Tue, 05 Nov 2024 00:00:00 GMT

Creating my own website has been on my mind for a while. For years, I kept ideas scribbled in notebooks waiting for that “perfect moment.” But, as the saying goes, “perfect is the enemy of the good.” So here I am, even if it is far from perfect (or even good — I am rusty).

Midjourney got my usual screensaver without a prompt, spooky.

For much of my career, I have been tinkering with technology in schools — building tools that save coworkers time and help institutions hold onto knowledge that would otherwise walk out the door when staff leave. Until now, most of what I made stayed inside whatever school or team I was working with. At some point I stopped being able to justify keeping it all to myself.

I just wanted a platform, a space I control, where I can share openly. Now it’s here.

I do not believe in the present day content treadmill to stay on top of an algorithm. Call me old-fashioned, but I just want to take a slow and steady approach to this site and my writing. I grew up in the 90s watching the internet reshape everything it touched, and I still think technology does more good than harm on balance. Unfortunately, that algorithm treadmill dominating the web has been an evolution I highly dislike. I want this site to feel more like the old web — somewhere I write and share at my own pace, without optimizing for anything except whether the writing was worth the time it took.

For now, I plan to write about whatever holds my attention — sometimes that will be useful material for educators, sometimes it will be whatever I have been thinking about that week. If some of it ends up being useful to someone else, even better.

The part of this I am still excited about is the public accountability. Thinking in public is one of the faster ways to find out where your reasoning breaks. If anyone reads this and pushes back, that is the best possible outcome — it means the writing did enough to be worth disagreeing with.

The hardest part of getting here was my own perfectionism. I kept stalling because nothing felt finished. Every draft had something wrong with it, and “something wrong” was enough to keep it in the drawer. But nothing was going to improve sitting in a notebook. So I am writing in the open instead — putting work out before it feels ready, and seeing what happens when other people can actually respond to it.

I have no idea whether anyone will read this, and I think that might be the point.

ritchot.me

I built an AI Literacy course

Footnotes

reflections, and the state of LLMs at the end of 2025

Footnotes

GPT-5 has come a long way in Mathematics

Last Year’s Test

This Year’s Setup

Question Source

Prompts, Images, And “Thinking Mode”

More on The Code Prompt vs No-Code Prompt

The Results

Tool Calls, Prompting, And Why Our Training Is Out Of Date

Vision And Geometry Are Still A Lagging Edge

Using GPT-5 As A Teaching Tool

AI As A Mathematical Collaborator

AI And Academic Scholarship More Broadly

Institutions, Benchmarks, And “Job Interviewing” Your Models

Limitations and Caveats

Where This Leaves The Classroom

Footnotes

on writing, and an MIT study

Purpose vs. Real-World Relevance of EEG Use to Measure Brain Activity

Task Suitability as Standardized Essays Emphasize External Revision

Reported Ownership and Memory

Human Consistency and Bias in Essay Scoring

What I Would Like To See

Other Pieces of Interest

some thoughts on emergent technology and the future of education

Footnotes

o1 pro mode still has a long way to go for Mathematics

Methodology

Results

Analysis

Limitations

Footnotes

AI: Knowing The Gods We Have Created

Footnotes

From Private to Shared Knowledge