LLMs are Here for the Long Haul
A lot of people are deluding themselves into thinking AI is a passing fad. They're wrong and delaying us preparing for what's to come.
AI has proven to be an incredibly polarizing technology, with some welcoming it as a superpower and others reacting in disgust to anything AI-related at all. Regardless of what side you are on in that debate, it is important to understand LLMs are here to stay. Many talk as if this will all be over once the “bubble bursts,” but it is, simply put, delusional to think even a major market setback will eradicate this technology from regular use.
Is the current level of hype and VC investment in the field unsustainable? It’s highly likely, and we’ve seen similar trends before, such as the “dot-com bubble,” when the embrace of e-commerce led to a proliferation of online businesses, only for many of them to collapse shortly thereafter. As the bubble burst, the amount of money consumers spent on online businesses continued to rapidly grow — “internet consumer goods sales grew from $7.8 billion in 1998 to an estimated $108 billion in 2003.” There was just too much money poured into too many companies shooting for the stars at once for the hyperspeculative frenzy to remain sustainable.
We are seeing similar aggressive growth in spending on AI. While it is completely understandable to expect some of these aggressively capital-intensive AI labs to shutter, consolidate, or massively dial back spending, the idea they will vanish is incredibly unlikely. These companies are instead, assuming there is a bubble pop, simply vying for who will be the Amazon or eBay rather than the Pets.com.
However, let’s entertain the idea that something does take down all of these companies and their intellectual property — the models — are not sold off to the highest bidder to someone else. There is an ever growing wealth of open source AI available for anyone to run. DeepSeek and Qwen, two popular families of open source models developed by China, are able to compete with a lot of commercial models made by American companies and, at times, have even edged them out in various benchmarks. There are also American companies like Mistral that open source their models while generating revenue from selling subscriptions to enterprise solutions and consumer chatbots.
Many of those who think they can stop the tide of AI also broadly hate on “data centers,” unaware of how much they rely on them — keep in mind that streaming video remains one of the biggest drivers of data center demand! — and imagining they will simply be left to gather dust if demand collapses. However, those who put down big money for capital expenditures will try to recoup their investment however they can, even if it means driving down the price they can charge for compute.
Other businesses continuing to provide AI-as-a-service will be trivial in the lieu of alternatives, even if not with models they’ve created. Ideally, more AI will be hosted by the end users themselves as well. Data centers are not a necessity — a cluster of a few Mac minis can run cutting-edge models at full power. Models can also be “distilled” to reduce the memory necessary to run them, at the expense of some of their performance, but many models still do extremely well at distillations in the 30-70 billion parameter range. That only needs in the range of 32-64 GB of RAM, a range becoming increasingly accessible via Macs and gaming GPUs.
While numerous problems and concerns remain associated with “vibe coding” where AI is used to generate entire apps from scratch with little human guidance or review, AI use in software development is increasingly ubiquitous, as it is undeniably helpful with writing more basic code. This is already creating problems for junior engineers trying to break into the industry, who would usually have a lot of that work offloaded onto them, though it is empowering solo devs to not be bogged down with minutiae when trying to tackle ambitious projects without funding.
Executives, in a race to save as much money as possible, will undoubtedly try to replace some labor with AI prematurely, causing problems in the process, but as AI’s capabilities improve — and we also see commensurate improvements in robotics — it is undeniable that the ascent of AI will likely make it harder for a lot of people to find jobs. But rather than fight a futile battle, we should start seriously considering what the world looks like when automation has driven down costs but fewer and fewer people have significant money to spend. We will need UBI and social programs to ensure people can benefit from this impending abundance rather than be locked out of it.
Educators also must adapt. If you truly need proof that an essay was written by the student themselves without AI assistance, it will need to be in class — though that was, on some level, always the case. Nerds getting bullied into or paid to do homework for their classmates is a common film and TV trope for a reason! LLMs simply reduced the complexity and cost of outsourcing such work. However, students will need to be prepared for a world full of AI and we should be teaching them to use it responsibly.
Pandora’s Box is open, chaos is loosened upon the world, but within the box, something important remained: hope. Everyone now has a choice: embrace that hope or try in vain to put everything back in the box. Choose hope. Fighting progress, in the end, never yields lasting results, be it with Luddites or bigots. Both, in the end, are a form of reactionaryism — a desperate attempt to claw back the status quo ante, that hurts people in the process of accomplishing nothing. Instead, fight for a world that is full of acceptance and abundance. That future will undoubtedly come, but the more fiercely we fight for it now, the more people we can spare from suffering sooner.