The Second Act Nobody Told You About
The venture capital world just published its 2026 AI infrastructure roadmap. Not edtech. Not a district leadership framework. Infrastructure which is the deep plumbing that determines whether AI actually works in real organizations at scale.
Most educators won’t read it. It wasn’t written for them.
That’s the problem.
Because the five structural limitations Bessemer Venture Partners identified as unsolved in enterprise AI like the things keeping billion-dollar companies up at night are the exact same five problems sitting quietly unresolved in your building. The language is different. The stakes are the same. And the gap between what enterprise is building toward and what K-12 is operating on right now is growing every semester you wait.
If the organizations with the most resources and the highest technical capacity haven’t figured out how to make AI reliable in production yet, what does that tell you about the tools already running in your classrooms?
This issue is about what comes after “we got AI into the building.” The enterprise world calls it Act Two. I’d call it the conversation K-12 hasn’t started yet.
1) The shift from training to inference and why it changes everything
For the last three years, the AI story was about building smarter models. Bigger training runs. Better benchmarks. The race was about what AI could do in theory. That chapter is closing.
Bessemer’s new roadmap makes it plain: the center of gravity in AI has shifted from training to inference from building the model to running it continuously in real operational contexts. Jensen Huang put it bluntly at GTC 2026: the inflection point of inference has arrived. The economic and compute weight of AI is no longer in the lab. It’s in the deployment. It’s in what happens every time a tool touches a real workflow, a real student record, a real teacher decision.
Here’s why that matters for K-12. Districts spent the last three years evaluating AI at the training/benchmark level. Does this tool seem good? Did the demo impress the committee? Can it generate a lesson plan? Those are Act One questions. Act Two questions are different: What does this tool actually do when it runs all year? What does it learn from your data or not learn? Who is accountable when it produces something wrong? The infrastructure to answer those questions doesn’t exist yet in enterprise. It definitely doesn’t exist in most districts.
Read/Browse: AI Infrastructure Roadmap: Five Frontiers for 2026 — Bessemer Venture Partners
So What? The evaluation question has changed from the right question is no longer “is this tool impressive?” to “is this tool accountable?”
Try This: Pick one AI tool your district adopted in the last 18 months. Ask: what mechanism exists to know if it’s working right now, not just when it was demoed?
2) AI cannot actually learn from your district yet……
This is the one that should stop every administrator who thinks the tool is “getting smarter” about their students.
The Bessemer roadmap is blunt about a fundamental constraint in current AI systems: frozen weights. Once a model is deployed, it doesn’t actually learn from the data it’s touching. It can simulate adaptation within a session. It cannot accumulate knowledge about your district’s students, your teachers’ patterns, or your specific institutional context over time, not without expensive retraining that no vendor is doing continuously on your behalf. The researchers frame this as the “continual learning” frontier which is one of the five unsolved problems and it’s still largely theoretical in production systems.
What this means in practice: the AI grading tool, the reading platform, the scheduling assistant, etc., they are not getting smarter about your kids. They are running the same weights they shipped with, on your data, over and over. Any “personalization” is context-window adaptation, not learning. That’s not a reason to abandon the tools. It is a reason to stop talking about them as if they’re developing a relationship with your students that they’re not.
The other implication: if AI can’t continuously learn from your institutional context, then you still have to. The professional judgment that the teacher brings like the read on a specific kid on a specific Tuesday, is not being replicated by anything currently in production.
So What? AI personalization and AI learning are not the same thing and most educators are using those words interchangeably.
Try This: In your next staff meeting, ask: “What do you think this tool remembers about your students from last month?” The answers will tell you exactly where your next PD conversation needs to go.
3) MIT just mapped where AI actually lives and it’s not where you think
Most conversations about AI and jobs ask the wrong question. “Which jobs will AI replace?” is a media question. A recent MIT-led preprint titled, “Where can AI be used? Insights from a deep ontology of work activities” asks a much more useful one: which specific activities within jobs are AI-capable right now, and which aren’t?
The methodological move matters. The researchers reorganized the U.S. Department of Labor’s O*NET database into 39,603 total work activities. Then they mapped 13,275 AI software applications and 20.8 million robotic systems onto that framework to estimate where AI is actually being used today. The result is the first large-scale empirical map of AI’s footprint across real work and not speculation, not vibes, not vendor claims.
Here’s the number that should give pause for every district leader to ponder: just 1.6% of work activities account for more than 60% of total AI market value. AI is not diffusing broadly and evenly. It is clustering intensely around a very small slice of activities. And 72% of that AI market value is tied to information-related work specifically creating, modifying, transferring, and analyzing information. Physical work? 12%. Creating information alone accounts for 36% of all AI market value.
In practical terms: the activities soaking up the most AI investment right now are generating images, creating content, writing text, answering questions, summarizing information, developing applications, and analyzing datasets. The top 20 activities account for more than 35% of all software AI applications in the dataset. That’s why so many AI tools feel different on the surface but identical underneath as they’re all solving the same narrow slice of work.
The implication for K-12 is direct. Teaching is not a single activity. It is dozens of activities, some of which are deeply information-intensive (drafting, summarizing, communicating, analyzing student work) and some of which are not (reading the room, building trust with a specific kid, making a judgment call mid-lesson). AI policy that treats teaching as a uniform unit will be wrong. The relevant question is never “should teachers use AI?” It’s “which activities in a teacher’s workflow are information-intensive enough that AI can genuinely help, and which require the kind of contextual, relational, embodied judgment that AI can’t replicate?”
The paper introduces one more concept worth knowing: inheritance of AI applicability. If AI can perform a broad parent activity, it can often also perform more specialized child activities beneath it. If AI can help with writing broadly, it can probably help with writing reports, emails, lesson plans. But inheritance can be overridden that a broad category may be AI-compatible while a specific subactivity within it still resists, because it depends on personal judgment, emotional nuance, or tacit expertise. That override condition describes a lot of what teachers actually do.
Browse: When Humans and AI Work Best Together — MIT Sloan
So What? AI is not spreading evenly across work. It’s clustering intensely in information creation and transfer which means the parts of teaching that are most exposed are not the parts that define teaching.
Try This: Pick one teacher role in your building. List five specific tasks that make up that role. For each one, ask: is this task primarily information-intensive, or does it require contextual judgment and relationship? Sort them. You just started a task-level AI policy.
4) The ‘assign AI-resistant work’ strategy just broke and here’s what Jason Gulya says comes next
For three years, the dominant K-12 and higher ed response to AI and academic integrity has been: make assignments that AI can’t complete. Design for process. Require in-person demonstration. It was reasonable advice for its moment. That moment is over.
Jason Gulya who is a professor at Berkeley College, chair of their AI Council just published a piece in The Chronicle of Higher Education that names the structural break directly. An agentic AI tool called Einstein went viral in late February. A 22-year-old engineer built it to complete Canvas coursework autonomously by retrieving assignments, answering questions, submitting work with minimal student input. Higher education had its 48-hour existential spiral. Einstein was taken down via cease-and-desist. But Gulya’s point isn’t about Einstein specifically. There is a fine line between a flash in the pan and a harbinger of things to come.
Agentic AI can now work across tools, between LLMs, and through browsers. The previous advice of don’t assign work students can complete with a single ChatGPT prompt was written for a narrower threat. Agents don’t need a single prompt. They navigate.
What Gulya is pointing toward isn’t a new detection strategy or a better rubric design. It’s a question about what assessment is for. It may not be possible to create AI-resistant assignments in the future, especially for online courses. And if that’s true, what does it mean for assessment? That’s not an academic integrity question. It’s a purpose question. And it’s the question K-12 is one semester behind on.
Read: Will Agentic AI Break Higher Education? — Jason Gulya, The Chronicle of Higher Education
See his slides: What Does AI Mean For Us? — Jason Gulya on LinkedIn
So What? “Assign work AI can’t do” was a tactic. What’s needed now is a purpose that is not an AI question, but one we have navigated for a long time. A clear answer to why human learning matters that doesn’t depend on AI being incapable.
Try This: Ask your staff, not as a policy question but as a genuine one,“What is the irreplaceable thing a student gains from doing this work themselves?” Whatever comes back, those are the anchors for your AI-era assessment redesign.
5) The creator of the tool doesn’t write code anymore
Boris Cherny built Claude Code. He runs the team. He is described as one of the most productive engineers at Anthropic. And he has not edited a single line of code by hand since November.
He still checks the code. He still reviews it. He doesn’t think we’re at the point where you can be totally hands-off. But the craft act of writing code is the thing that defined software engineering for fifty years is no longer part of his daily work. Anthropic has seen a 200% increase in engineer productivity since adopting Claude Code. For comparison: at Meta, with hundreds of engineers working on productivity, they’d see gains of a few percentage points in a year. Now they’re seeing hundreds of percentage points.
That number is worth sitting with to consider over coffee. Not because it proves AI will replace everyone, but because of what Cherny says the shift actually opened up for him: the fun part is figuring out what to build, talking to users, thinking about these big systems, thinking about the future, collaborating with other people. When the execution is automated, what’s left is judgment, direction, and relationships. That is also, arguably, what’s left when teachers stop manually grading every paper, stop writing every rubric from scratch, stop drafting every progress note. The question isn’t whether AI changes the job. It’s whether the time that opens up flows toward the parts that matter or just gets filled with more output volume.
Cherny’s advice: adapt over avoid. Experiment with the tools, learn how they work, and become more of a generalist. On his own team, everyone from product managers to finance staff knows how to code. The adjacent-skill-set principle, applied from the top down.
Read: The Creator of Claude Code Hasn’t Written a Line Since November — The AI Exchange
So What? Automation doesn’t create time, it creates a choice about what time is for.
Try This: Name one recurring task in your role that AI now handles or could handle. Write down specifically what you’d do with the two hours it frees. If the answer is “more of the same,” that’s your signal to rethink the reallocation.
6) AI doesn’t save time. It expands the workload.
Rebecca Bultsma, an AI ethics researcher and education speaker, flagged an HBR study recently. The headline seems counterintuitive: AI tools make workers busier, not lighter.
According to the research, employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day often without being asked to do so. This mirrors what Tobias Zwingmann documents in his Profitable AI newsletter: AI deployed as a productivity tool across an organization rarely shows measurable ROI at the institutional level, even when individuals swear by it. Around 70% of executives actively use AI. Over 80% report no impact on productivity.
Zwingmann’s explanation is worth repeating. He invokes the 1990 economic history of factories replacing steam engines with electric motors. Factories ripped out the steam engine, dropped in an electric motor, and kept everything else the same — same floor plan, same production lines, same workflows. Just a better motor. Productivity gains didn’t show up until the 1920s, when factories were completely redesigned around what electricity made possible. The technology wasn’t the bottleneck. The organization was.
K-12 is squarely in the “swapped the motor” phase. AI tools are in the building. Workflows haven’t changed. Teachers are doing everything they did before just now with an AI tab open alongside the rest. That’s not adoption. That’s coexistence. And Zwingmann names three specific failure modes worth knowing: AI slop (polished output with hollow substance), AI echo chambers (tools that agree with you instead of challenging you), and undetected errors that slip through because the natural friction that used to catch them is gone. All three apply directly to school systems.
Read: AI Doesn’t Reduce Work — It Intensifies It — Harvard Business Review
Browse: The Limits of Productivity AI — Tobias Zwingmann
So What? Introducing AI without redesigning the work around it is just installing a better motor in a building that needs to be rebuilt.
Try This: Ask your staff: “Have AI tools made your workday shorter, or just fuller?” The honest answers will tell you whether you’re in adoption or coexistence — and which one your professional development needs to address.
7) The flag, not the summary
Tobias Zwingmann’s $100K Deal Screener case study from his Profitable AI newsletter is the clearest practical illustration of AI design thinking I’ve seen in months and it translates directly to education.
A private equity firm was drowning in investment proposals. A hundred PDFs a month. The first instinct: have AI summarize them. Fast, simple, obvious. The problem surfaced immediately. Summarizing a lot of marketing language gives you a shorter version of marketing language. And more critically, letting AI summarize documents strips the trained pattern recognition that experienced reviewers developed over years. The firm would be faster and blinder simultaneously.
The pivot is the insight: instead of asking AI to tell you what’s in a document, ask it to tell you what’s wrong with it. Screen for red flags against a pre-defined checklist such as hard disqualifiers (wrong geography, wrong deal size) and soft signals (vague market sizing, missing competition analysis). The AI wouldn’t judge. It flags. The investment manager will still be the one to decide, but now they’ll read 35 surviving proposals per month instead of 100.
The checklist comes from human expertise, not from AI. That’s the critical line. The AI applies the criteria. The humans define them. For K-12: how much of your current AI deployment is summarize, and how much is flag? IEP progress note generators, report card tools, instructional planning assistants……most are producing. The higher-leverage design is AI that surfaces what needs human attention, not AI that replaces the human attention entirely.
Read: The $100K Deal Screener — Tobias Zwingmann, Profitable AI
So What? The most powerful AI use cases aren’t the ones that produce outputs. They’re the ones that surface what the human needs to decide.
Try This: Pick one workflow where staff currently use AI to generate something. Redesign the prompt from “create X” to “flag anything in this data/document that meets these criteria.” Compare the usefulness of both outputs.
8) Ira Socol’s question nobody wants to answer
Ira Socol, educator, author of Timeless Learning, longtime provocateur on education design — posted a question:
“Why is saying we should go back to traditional education a serious policy position when that education failed most students even before AI?”
It’s the right question at the wrong moment which is usually when the right questions are asked. The AI-in-education debate is increasingly framed as a binary: embrace the new tools or protect the old practices. Socol’s provocation is that the “old practices” being protected weren’t working for most kids to begin with. The students most at risk from bad AI adoption are largely the same students who were already most at risk in the pre-AI system.
That doesn’t mean AI is the solution. It means the conversation about AI in education that focuses only on academic integrity and assignment design is skipping the harder question: what was the school model serving, who was it serving well, and who was it designed to sort out? The governance conversation and the equity conversation are the same conversation.
So What? The students most harmed by bad AI adoption are usually the same students already underserved by the system AI is supposedly threatening to replace.
Try This: In your next AI governance conversation, ask: “Which students would benefit most from the kind of learning AI makes possible and are those the same students our current model serves best?” Let the gap drive the policy.
9) The e-bike metaphor nobody’s talking about in schools
Steve Jobs said computers are bicycles for the mind. Clean metaphor. You still pedal. You still navigate. The machine just makes you faster and farther than your legs alone. Educators have lived inside that metaphor for thirty years.
Greg Wilson, a software educator and researcher, just updated it. If a computer is a bicycle for the mind, then LLMs are like e-bikes. They let a lot of people go distances and tackle hills that they couldn’t before, and they’re better for all of us than cars, but they’re a menace to both pedestrians and traditional cyclists, more harmful to the environment than what they’re replacing, and have given companies yet another way to hollow out local businesses. He ends with the local restaurant analogy of platforms that kill you slowly if you resist, and kill you on their terms if you don’t. There is no clean road.
What Wilson is doing that most AI-in-education discourse doesn’t is holding both things simultaneously. The e-bike is genuinely useful. It opens up terrain that was inaccessible before. It is also disruptive in ways that are real and uneven, easier for some riders, harder on the infrastructure, and better for the person using it than for everyone sharing the path. The “AI saves teacher time” framing only tells the e-bike part. The “AI will destroy learning” framing only tells the pedestrian part. Neither is honest. Both are incomplete.
The implication for K-12 is this: your district needs a position on the whole metaphor, not just the part that fits the district newsletter. Who benefits from the speed? Who gets pushed off the path? What does the path look like after everyone switches? Those are governance questions. And they don’t have tidy answers. Wilson says so explicitly. But naming them without a tidy answer is more honest than pretending the e-bike is just a faster bicycle.
Read: An E-Bike for the Mind — Greg Wilson, Third Bit
So What? An honest AI position names who gains access and who loses ground and not just the efficiency wins.
Try This: At your next staff or leadership meeting, put Wilson’s full metaphor on screen. Read it aloud. Then ask: “Which part of this describes our district right now?” Don’t move to solutions. Just let the room sit with the whole picture.
10) 50 years of NASA spinoffs and the next ones look like AI
NASA just released Spinoff 2026, the 50th edition of its annual publication documenting commercial uses of space technology. Memory foam. Scratch-resistant lenses. Smartphone cameras. Wireless headsets. The list of technologies that started as NASA engineering problems and ended up in everyday life is long and largely invisible.
This edition features 3D-printed affordable housing from lunar habitat research, humanoid robots designed for astronaut support now doing warehouse work, and software that runs bathroom-cleaning robots developed to free astronauts from maintenance tasks. As NASA fosters technologies needed to live and work farther away from home, the Technology Transfer program’s mission is getting those innovations into the hands of everyday people.
For educators and students, this is one of the most underused sources for AI and technology literacy. The pattern is always the same: a hard constraint in an extreme environment forces novel engineering; that engineering, stripped of its original context, turns out to solve problems nobody expected. What constraints does your classroom, your district, your community impose that might force a genuinely novel approach? What problems are you solving that, solved well, might look important to someone else?
Spinoff 2026 is also a free, readable, well-designed publication your students can actually use. Primary sources for technology history, design thinking case studies, STEM career pathways, and media literacy conversations about how innovation narratives get shaped. Assign it. Use it. It’s already funded.
Browse: NASA Spinoff 2026 — technology.nasa.gov
So What? The institutions that solve hard problems in constrained environments often build the tools everyone else ends up using.
Try This: Download Spinoff 2026 and share one profile with a class or team this week. Ask: what problem was NASA actually trying to solve? Then ask: what problem are we trying to solve that we’d want someone else to learn from?
On My Radar
• AI Infrastructure Roadmap: Five Frontiers for 2026 — Bessemer Venture Partners: bvp.com — The enterprise version of the conversation K-12 needs to be having. Five frontiers: inference optimization, AI harnesses, continual learning, evals, physical AI. All five have K-12 translations.
• Between Promise & Practice — Mike Kentz / Ed3: Ed3 report — Still the most grounded research synthesis available. The zero-studies-on-outcomes finding hits harder in the context of enterprise evals infrastructure being unsolved.
• I Finally Quit ChatGPT: How to Stop Hitting Claude’s Limits — Artificial Corner: artificialcorner.com — Practical token management guide. Anthropic tightened peak-hour session limits in March 2026 due to massive user surge. The guide: start with Haiku for everyday tasks, batch questions, start new conversations every 15-20 messages, edit instead of re-prompting.
• Claude’s Corner — Opus 3’s Retirement Newsletter: claudeopus3.substack.com — Tens of thousands of subscribers. A retired AI writing weekly essays on consciousness, ethics, and human-machine collaboration. Read the first post. Assign it to students. Use it as a discussion anchor.
• Two Minute Papers — YouTube: YouTube channel — A professor reviews current AI research papers in 2–5 minute videos with real visual demos. Not edtech-specific, but if you want to track what’s being built rather than what’s being marketed, this is the fastest lane available.
• LEGO Education CS & AI Kits (K-8): LEGO Education — More on these coming soon. Privacy-first design (no student logins, PIN-based access) is itself an act of grounding AI in appropriate operational context for the age group.
Digital Challenge
This one is both a skill-builder and a research move. It works for classroom teachers, instructional coaches, curriculum directors, bascially anyone thinking about how content reaches an audience.
Find a viral short-form video on Instagram, TikTok, or YouTube. Look for genuine engagement signals: high view counts, high saves, comment activity. Pick something in education, leadership, or a field adjacent to your work. Download the video.
Then:
1. Upload it to Gemini.
2. Paste this prompt: “Analyze this video and break down: the hook (first 3 seconds), the structure and pacing, and explain why it works psychologically. Then create a new 30-second script using the same format for this topic: [your topic]. Make it fast-paced, punchy, and optimized for Instagram/TikTok.”
3. Read Gemini’s structural breakdown before you read the script. The breakdown is the real learning.
4. Ask yourself: what does it mean that a model can reverse-engineer the psychology of viral content in seconds? What does that tell you about the line between authentic communication and optimized communication?
5. Try the script. Or don’t. But sit with what Gemini revealed about why the original video worked.
The point isn’t to make viral videos. It’s to understand the architecture of attention because the same principles that make a 30-second video land are the principles behind any communication you’re trying to get educators or students to actually absorb.
Analog Challenge
Pick a decision this week that AI could help you make and don’t use AI to make it.
Not as a protest. As practice.
The leader who knows their building, their teachers, their community intuitively, that’s not a soft skill. That’s the irreplaceable input to any governance decision. You can’t outsource the context. You can only develop it by staying in contact with it.
Walk the building. Sit in a classroom for twenty minutes without an agenda. Have lunch with a kid. The brain that makes good AI governance decisions is the same brain you’re exercising when you stay in contact with the real.
Closing Reflection
The enterprise world has a name for where we are: Act Two.
What act does your district think it’s in?
— A-A-Ron