No, the AI Genie Won't Go Back Into the Bottle
Why the Potential “AI Bubble Burst” Won’t Matter at All
This post follows my standard early access schedule: paid subscribers today, free for everyone on November 25.
I’ve noticed something curious in recent conversations with fellow educators: a kind of wishful waiting. Some colleagues speak hopefully about the eventual bursting of the “AI bubble,” imagining a return to familiar pedagogical ground once venture capital loses interest and major AI companies fold or scale back. The underlying assumption feels almost comforting: that this technological disruption might prove temporary, a passing frenzy like so many tech booms before it.
But this hope rests on a fundamental misunderstanding of where AI capabilities currently exist. The common perception frames artificial intelligence as something housed primarily in corporate data centers, accessible only through subscription services and dependent on the continued operation of companies like OpenAI, Anthropic, and Google. If these companies disappeared tomorrow, the thinking goes, their AI systems would disappear with them.
The reality looks different. While most educational discussions center on ChatGPT, Claude, and Gemini, a parallel ecosystem has quietly matured to where it matches, and in some ways exceeds, the capabilities of these commercial services. This open-source, locally executable AI infrastructure cannot be uninvented or recalled. The tools exist, the models are distributed, and the knowledge of how to use them spreads daily. More importantly, these systems run not on remote servers but on consumer hardware that many professionals already own.
What “Local AI” Actually Means
When I refer to local AI, I mean models that execute entirely on a personal computer without requiring internet connectivity or cloud services. These aren’t simplified versions of “real” AI. Current open-source language models with hundreds of billions parameters rival the capabilities of commercial offerings, while image, video, and audio generation tools that once required expensive API access now run on capable consumer machines.
The hardware barrier has largely dissolved. Consider my own setup: a MacBook Pro with Apple’s M4 Max processor and 128GB of unified memory. This MacBook costs between $4,500 and $5,500, depending on configuration. And while this is certainly expensive, it is within reach of professionals and comparable to what many schools spend on teacher workstations. A similarly capable desktop PC with an NVIDIA RTX 4090 and sufficient RAM would cost approximately $3,500 to $4,500.
These aren’t exotic research machines. They’re consumer products available at any electronics retailer. The Mac’s unified memory architecture proves particularly significant here. Where a traditional PC separates system RAM from graphics memory (with the high-end RTX 4090 limited to 24GB of VRAM), Apple’s design provides a single, large memory pool accessible to both CPU and GPU. This architectural choice means a 128GB Mac offers approximately 96GB of usable video memory. This is enough to run language models that would require multi-GPU server configurations on traditional hardware.
The performance characteristics matter because they determine usability. Running Meta’s Llama 3.3 model with 70 billion parameters requires about 75GB of VRAM on my MacBook and generates text at roughly 350-400 words per minute. This is faster than most people read and more than sufficient for interactive work. I am using the 8-bit quantized version, which offers a strong balance of speed and fidelity. It runs at 6.5 tokens per second, while a more compressed 4-bit version can run even faster with slightly reduced output quality. These speeds feel responsive. The system doesn’t lag or stutter. It just works.
The Language Model Landscape
The center of this ecosystem revolves around open source user interface software such as LM Studio, a free application that manages local language model execution. Its interface resembles ChatGPT with a conversation window, a text input box, and a sidebar for managing chats. The difference lies entirely in what runs underneath. Instead of sending queries to OpenAI’s servers, LM Studio loads models directly into the computer’s memory and processes everything locally.
The available open-source models are no longer simply playing catch-up. For instance, Meta’s Llama 4 family (specifically the Scout model) is natively multimodal, handling text and vision tasks with a massive 10 million token context window. And for pure reasoning and technical skill, DeepSeek-R1, a 671-billion parameter Mixture-of-Experts (MoE) model, is fully open-source and competes directly with the best proprietary models in coding and reasoning benchmarks.
Perhaps most impressively, however, the recently released Kimi K2 Thinking model is a 1-trillion parameter open-source “thinking agent” (with 32B active parameters, making it viable for high-RAM local systems) designed for complex, autonomous tasks. And even the “ChatGPT” name isn’t just for the proprietary model anymore, with OpenAI’s own GPT-OSS 120B freely available for download and accessible within LM Studio. These models represent a profound leap in capability, moving far beyond the GPT-3.5 era to challenge even the most advanced commercial-grade services.
What strikes me most about working with these models is their practical indistinguishability from commercial alternatives for most educational tasks. Ask for a lesson plan, request feedback on student writing, generate example problems, or explore a complex topic through dialogue—the local models handle these tasks with comparable sophistication. The responses show similar reasoning capabilities, similar factual knowledge, and similar linguistic fluency.
The key difference lies not in capability but in deployment. Local models run without content filters or usage restrictions. They never experience downtime or rate limiting. They process sensitive information without transmitting it anywhere. They don’t require ongoing subscription fees. And they continue working regardless of whether an AI company remains in business.
Beyond Text: Image, Video, and Audio Generation
Language models capture most attention in educational discussions, but the local AI ecosystem extends far beyond text. ComfyUI, an open-source framework for AI-powered media creation, has recently undergone a significant development that makes sophisticated image, video, and audio generation accessible through pre-built templates rather than requiring extensive technical knowledge.
For image generation, models like Stable Diffusion 3.5 and FLUX.1-dev produce results comparable to commercial services. These systems generate novel visual content from text descriptions, convert sketches to finished artwork, and manipulate existing images with remarkable precision. The integration of SUPIR (a model for upscaling and restoration) and ControlNet (for precise compositional control) within ComfyUI workflows means an educator can generate custom diagram illustrations, create visual aids for historical periods, or produce culturally specific imagery for diverse classrooms without artistic training.
Video generation remains the most computationally demanding domain, but open-source models have recently made unexpected progress. Wan 2.2 Animate, a 14-billion-parameter model, transforms static images into animated sequences or applies consistent style transfers across video frames. And while commercial services like Sora or Kling show superior capabilities, the gap narrows monthly. More significantly, however, running video models locally means unlimited experimentation without per-generation costs or content restrictions.
Audio generation recently also saw a major development with the release of LeVo, an open-source music generation model that rivals commercial services like Suno or Udio. Initial demonstrations suggest output quality approaching professional production levels. And ACE-Step, another recent release, generates music directly within ComfyUI workflows. Both models run on consumer hardware, albeit slowly on current machines. But they are functional.
The practical significance for education centers on accessibility and experimentation. Commercial AI services charge per generation or limit monthly usage. Local systems have no such constraints. A teacher developing curriculum materials can iterate freely, generating dozens of variations until finding what works. A student exploring creative applications can experiment without worrying about exhausting credits or incurring costs. The removal of financial friction changes how these tools integrate into educational practice.
The Training Capability
Here’s where the irreversibility becomes most clear: these models can be modified and specialized through techniques that require modest computational resources. LoRA (Low-Rank Adaptation) fine-tuning allows individuals to train custom model behaviors directly on personal hardware.
Consider a practical example. Suppose an art teacher wants students to analyze paintings in the style of a particular art movement, but available language models lack the specialized vocabulary and analytical frameworks specific to that movement. Using LoRA training with a few dozen text examples demonstrating the desired analysis style, that teacher can create a customized model variant that costs nothing and runs on the same hardware as the base model.
The same principle applies to image generation. If Stable Diffusion cannot accurately render architectural styles from a specific historical period, training a LoRA adapter on relevant examples takes hours rather than days and requires no specialized expertise. The Kohya_ss training suite and similar tools provide graphical interfaces that handle the technical complexity. The resulting trained adapters measure only a few hundred megabytes and load on top of base models without replacing them.
This training capability means that the open-source ecosystem improves through distributed contribution. When someone develops a useful specialized behavior, they can share the LoRA adapter with others. Communities form around specific use cases, collectively building capabilities that no single company would prioritize. The models become more capable over time through this distributed intelligence, independent of corporate development roadmaps.
From an educational perspective, this accessibility matters profoundly. Students can train models on their own data, learning not just to use AI but to shape it. Teachers can customize tools for specific pedagogical needs without waiting for companies to serve their niche requirements. The technology becomes genuinely malleable rather than remaining a fixed external service.
The Energy Consideration
Discussions of AI’s environmental impact typically center on training large models, a process that requires enormous computational resources and generates substantial carbon emissions. The figures often cited involve thousands of GPUs running for weeks or months, consuming mega- or even gigawatt-hours of electricity.
But running inference—actually using the already trained models—requires far less energy than most people assume. My MacBook running a 70B language model draws approximately 60-70 watts at peak usage, comparable to a conventional laptop under heavy load. A gaming PC with an RTX 4090 draws 200-300 watts during model inference, similar to running modern AAA video games. These power draws remain well within residential electrical capacity and generate minimal environmental impact for individual users.
The per-query energy cost for local inference typically measures between 0.5-2 watt-hours, depending on model size and response length. For comparison, streaming an hour of HD video consumes approximately 120-240 watt-hours. The environmental concern around AI centers primarily on centralized data center operations serving millions of users, not on distributed local inference.
This efficiency matters because it undermines a common assumption about AI’s fundamental requirements. The technology doesn’t inherently demand massive infrastructure. The current centralized model exists for commercial rather than technical reasons. Local execution proves entirely viable for individual and small-group use cases, which includes most educational applications.
What This Means for Educators
I regularly encounter educators who focus only on commercial AI services, carefully tracking the latest ChatGPT updates or testing Claude’s newest capabilities. This attention makes sense, and I often find myself doing that as well. After all, these services are accessible, well-documented, and widely discussed. But this narrow focus creates a dangerous blindness to what students actually have access to.
Students with modest technical interests can download and run these same local models today. They don’t need permission, don’t require payment, and don’t leave traces in usage logs. The tools work offline, bypass content filters, and respond to any query. A student who wants to generate an essay or produce assignment content that evades AI detection tools can do so using local models that teachers may not even know exist.
This reality doesn’t argue for resignation or panic. It argues for informed engagement. When teachers understand what local AI can do, they can design assignments that account for these capabilities. They can teach students not just about AI ethics but about AI mechanics. They can model responsible use of powerful tools rather than pretending those tools don’t exist or remain somehow out of reach.
The economic argument reinforces this point. Schools may hesitate to adopt AI tools that require ongoing subscription costs, especially when budgets remain tight and competing priorities abound. But the one-time hardware cost of a capable machine, distributed across multiple classrooms or used for professional development, becomes more defensible. The total cost of ownership for local AI infrastructure can be lower than multi-year commitments to commercial services, especially at scale.
More fundamentally, however, the existence of this ecosystem means that decisions about AI in education cannot be framed as binary choices between adoption and rejection. The technology exists, it works, and it’s accessible. The relevant questions therefore need to focus on how we integrate it thoughtfully, what literacies we cultivate around it, and what pedagogical practices we adapt to acknowledge its presence.
The Irreversibility Thesis
Let’s return now to the initial premise: that the bursting of the AI bubble might take us back to pre-AI conditions. The evidence suggests otherwise. Even if every major AI company ceased operations tomorrow, the models would remain. They’re open source, widely distributed, and require no centralized infrastructure to function. The knowledge of how to use them spreads through documentation, tutorials, and community forums that exist independently of corporate support.
This pattern differs from previous technology booms that did indeed collapse. The dot-com bubble burst left many websites and services defunct because they depended on ongoing operational funding and centralized infrastructure. When companies failed, their services disappeared. But open-source software persists independently of its original creators’ financial status. Linux didn’t disappear when various companies built around it struggled. Firefox continued when Netscape collapsed. And the Apache web server remains ubiquitous despite having no parent company.
AI models follow this pattern. Once released, they propagate across torrent networks, academic repositories, and community archives. They become infrastructure rather than services—foundational tools that others build upon. The Llama models will remain available and functional regardless of Meta’s future decisions about AI development. Stable Diffusion will continue being developed through community contributions long after its original corporate sponsor, Stability AI, faced financial difficulties.
The genie truly won’t return to the bottle. The knowledge exists, the tools exist, and the community exists. Waiting for a bubble burst to resolve AI’s educational challenges represents a form of wishful inaction that misunderstands what has already occurred. We’ve crossed a threshold where certain capabilities have become widely accessible through means that no market correction can reverse.
There Was Never Really a Choice
This recognition demands a different approach to AI in education. We cannot rely on the technology becoming unavailable or prohibitively difficult to access. We cannot depend on content filtering or detection tools to identify AI-generated work when students can run unfiltered models locally. And we particularly cannot assume that limiting school access to commercial AI services meaningfully restricts students’ capabilities.
What we can do is engage honestly with the technology’s actual capabilities and limitations. We can design learning experiences that value process over product, making AI-generated final outputs insufficient for demonstrating learning. We can teach critical evaluation skills that apply equally to AI and human-generated content. And we can model thoughtful tool use that augments rather than replaces human thinking.
Most importantly, we can stay informed about the full landscape of available tools. The focus on commercial services makes sense from a convenience perspective—they’re easier to access and use. But comprehensive AI literacy requires understanding the broader ecosystem, including local tools that students increasingly employ. This means downloading LM Studio or ComfyUI, occasionally installing and running a local model, and experiencing firsthand what these systems can and cannot do.
The irreversibility of AI’s integration into daily life doesn’t require enthusiastic embrace or resigned acceptance. It requires clear-eyed acknowledgment and a thoughtful response. The technology exists. Students have access to it. Educators need to understand it. These facts remain true regardless of what happens to any company or whether the current market valuations prove sustainable.
The bubble may burst. The hype may fade. The valuations may collapse. But the models will still run, the tools will still work, and the capabilities will still exist. That permanence shapes the ground on which we must build educational practice for the coming decades. There’s no going back, but there was never really a choice about that. The question has always been how we move forward with wisdom, not whether we can remain where we were.
What are you seeing in your own environment? Have you encountered students using local AI tools you weren’t aware of? For educators and administrators: what assignment designs or institutional policies have you developed that account for the full spectrum of accessible AI tools? Are you experimenting with local models yourself, or do barriers still hold you back? Share your experiences and emerging practices in the comments, the conversation needs to move beyond what we wish were true to what actually is.
P.S. I believe transparency builds the trust that AI detection systems fail to enforce. That’s why I’ve published an ethics and AI disclosure statement, which outlines how I integrate AI tools into my intellectual work.