In 1866, an engineer at Siemens invented humanity’s first DC generator.
Forty years later, General Electric began mass-producing the first generation of incandescent light bulbs that truly brought electric lighting to the masses in 1906.
During the half-century between these two milestones, the human world remained largely in darkness. The electrical revolution seemed as if it had never happened.
But that’s only because we, looking back from the future, can so casually gloss over those 40 years with a single sentence. For the people living through it, the development of electrical technology unfolded day by day right before their eyes: the laying of the first telegraph line, the connection of the first telephone call, the maiden run of the first electric tram. Each technological advance was tangibly changing their lives — it just wasn’t happening fast enough for everyone to collectively gasp at a single moment: “Ah, the electrical revolution has finally arrived!”
We now stand at a juncture remarkably similar to the late 19th century. AI technology’s “DC generator” is already humming loudly. The “carbon filament” that can only glow for a dozen hours has been lit. A new age of illumination is gradually unfolding. Yet when we find ourselves inside it, the transformation seems less rapid than we imagined — even somewhat sluggish.
After the shock of ChatGPT’s debut in late 2022, the AI field in 2024 felt rather uneventful. Let’s use OpenAI’s 12 December launch events as a thread:
- Sora was teased and delayed for an entire year. When it finally arrived, it was “best in class” but hard-pressed to serve any practical purpose.
- GPT-5 was delayed. The highly anticipated full version of GPT-o1 and the as-yet-unusable o3 both suffered serious shortcomings in pricing and usability (data input and internet connectivity).
- ChatGPT Canvas, ChatGPT’s advanced voice mode, and the Projects feature were either long-overdue implementations of previously announced features or catchups to competitors.
Beyond these, there was a flurry of announcements covering reasoning model APIs, fine-tuning techniques, collaborative modes, and more. At first glance, it was dazzling. But upon closer inspection, it all felt like it was missing that original “revolutionary” spark — more like patching and polishing existing achievements.
Watching OpenAI’s launch events now feels like watching Apple keynotes — incremental, toothpaste-tube upgrades. Those who watched all 12 events and still pinned their hopes on o3 achieving AGI remind me of last year’s crowd who were certain Apple Vision Pro would usher us into the metaverse. I can only wish them happiness.
People started whispering: Where’s the AI revolution we were promised? The technological explosion? Why does it feel like AI’s development has slowed considerably compared to the past couple of years? Some even began to question: Has the Scaling Law failed?
Has the Scaling Law Failed?
To answer this, we need a quick primer on Scaling Law.
In simple terms, Scaling Law means “brute force works wonders” — as long as you keep increasing model parameters and training data volume, AI model performance will continue to improve. This principle drove the AI field’s meteoric rise over recent years, giving us front-row seats to AI capabilities growing at breakneck speed. If early 2022’s ChatGPT made people think “this is kind of novel,” the version born in late 2022 made people exclaim “this thing has become sentient!” — and behind that leap was Scaling Law at work.
However, as AI technology continued to advance, a question gradually surfaced: when model parameters grow ever larger and training data ever more voluminous, can we still expect equally dramatic performance gains? Research suggests that simply increasing scale can no longer deliver linear — let alone exponential — improvements. Models grow bigger, but marginal returns are diminishing. The inevitable question arises: Is Scaling Law truly on its last legs?
In November 2024, after overseas media reported obstacles in GPT-5’s training, Sam Altman posted on Twitter…
This question is genuinely critical for the AI industry. OpenAI’s 12 launch events were themselves a response to such doubts — though from my perspective, that response in some ways confirmed that Scaling Law had hit a wall.
But what does this have to do with non-AI industries?
Everything. If AI technology has hit a wall now, then the AI revolution for ordinary people and non-AI businesses is about to begin.
In the history of technological development, we often see this pattern: after a technology achieves a breakthrough, it sometimes enters a period of relatively stable development. This doesn’t mean the technology has stopped advancing — rather, it’s transitioning from “qualitative leaps” back to “quantitative progress,” from pursuing further technical breakthroughs to figuring out how to apply existing technology to things that truly matter.
Consider: when a new technology emerges, people naturally focus on its raw capabilities — can AI become smarter? Can it handle more complex tasks? It’s like the “jungle farming” phase at the start of a MOBA game; nobody’s thinking about team fights yet. But as the technology matures, attention gradually shifts from the technology itself to its applications. People start thinking about how to integrate AI with real-world scenarios, how to use AI to solve actual problems. This is the “push towers” phase, where the goal becomes converting technological advantage into real productivity.
This is exactly where AI stands today. On one hand, the technology’s own development appears to be slowing — growth in model parameters and computing power are hitting bottlenecks. The era of “major breakthroughs every few months” seems difficult to sustain. On the other hand, more stable model parameters and continuously falling prices are attracting more developers and enterprises to incorporate AI into their applications.
For ordinary people and non-AI businesses, this is good news. After all, what everyday users interact with isn’t the AI model itself, not even the API — it’s the application products built on top of the API.
When technological advancement is no longer the sole core proposition, when “brute force” is no longer the most cost-effective strategy, the entire industry’s gaze naturally turns toward exploring applications. That’s when AI truly begins its deep integration with every industry — that’s when AI applications start blooming.
AI’s practical capability equals model capability plus non-AI engineering capability. But in practice, model changes — even “evolutions” — frequently disrupt the engineering stack.
You painstakingly build an AI application, only to have a model update demolish the entire thing. This is something ordinary users can barely fathom, but it actually happens.
This is also why, in developer circles, Anthropic’s mid-year Claude 3.5 release and Google’s December wave of Gemini 2.0 models (released to counter OpenAI) generated more excitement than OpenAI’s own announcements. These two companies’ products offered no massive breakthroughs and no flashy gimmicks — they simply continued down known paths, offering more context and lower prices.
Moreover, in my productivity article, I raised a key point: most people and enterprises have barely scratched the surface of current AI capabilities because their engineering skills are essentially zero, leading them to conclude “AI is just OK for now.” In reality, they’re underestimating current AI and incorrectly assuming that better AI will magically make impossible things possible (like building a Taobao clone in a week).
This cognitive error leads the public to believe that Scaling Law hitting a wall will end AI’s societal impact. But the opposite is true. Many people think AI’s potential has peaked because their own digital literacy isn’t sufficient to fully leverage existing AI tools. In other words, they haven’t even figured out how to use the shovel in their hands, yet they’re already worrying about the gold mine running dry.
Here’s a less-than-perfect but vivid illustration:
You might assume that in the internet age, laboratory-level technological revolutions would reach the market faster. But that’s not necessarily the case. For instance, if you’ve recently experienced mobile payments across China, you’ll have noticed WeChat Pay and Alipay are locked in a new round of battle.
WeChat introduced palm-print payments — bind your palm once, and you can pay via WeChat even without your phone. Is palm recognition a new technology? Well, this one happens to be right in my wheelhouse. In 2013, fresh out of college, I worked at a financial advisory firm. One of the startups I worked with, called PalmVisa, was doing palm recognition. According to its founder at the time, palm recognition had already achieved financial-payment-credential-level security, and unlike fingerprints or iris scans, it didn’t carry biometric leakage risks.
Yet for the next 10 years, palm recognition technology virtually vanished from the commercial landscape. WeChat only picked it up because its rival Alipay started pushing “tap to pay” — NFC payments. NFC is even more familiar. The last time media debated “should the hyperlink between the internet and the physical world be NFC or QR codes” was 2012.
See? Even for a prize as enormous as “financial payments,” a technologically viable solution like palm recognition lay dormant in the market for a full decade. What does this tell us? That the application layer needs time and the right conditions to approach technology’s ceiling.
Examples like this abound.
A classic one: it’s only after platform companies form stable oligopolies that they can provide fertile ground for masses of micro-enterprises and non-IT industries to grow steadily.
In 2015, it was hard to imagine every random restaurant building its own app. But by 2018, WeChat Mini Programs were on fire — every roadside stall had its own mini program for ordering, payments, membership, and marketing. This wasn’t because app development technology in 2018 had qualitatively leaped beyond 2015. Quite the opposite — it was because the WeChat super-platform had formed, providing developers with sufficient infrastructure, a stable (if outdated) tech stack, and user traffic, making the cost of building mini programs low enough and the returns high enough.
Almost every developer complains about mini programs, but it’s precisely their technological backwardness that ensures their ubiquity. After all, to this day, Android phones still haven’t solved the system fragmentation problem, and huge numbers of users are still running Android 10 from five years ago. iOS is slightly better, but not by much. For businesses outside the internet industry, if they need to give users a digital entry point, a standalone app is simply never a cost-effective choice.
Even slower than individual digital literacy growth is organizational digital literacy. Even in China’s hyper-competitive workplace, a company can’t hire only people under 30 — especially in non-IT industries, where many traditional business owners are not young and aren’t sensitive to non-core-production technology.
This is why, when JD.com and Genki Forest are trying to deploy “unmanned factories,” the vast majority of China’s small and medium-sized factories are still using Excel (often pirated) for production planning, inventory management, and scheduling — looking downright primitive.
When will AI reach these factories? I’d say probably not until they’ve at least replaced Excel with a proper ERP system or an MES (Manufacturing Execution System).
Bosses and enterprises are profit-driven. Getting an organization to iterate on non-core technology requires showing them sufficient benefits or returns. “Core technology” here refers to technology directly related to their production — for Coca-Cola, the formula is core technology, while production-line automation equipment and supply chain management systems are non-core technologies that support the formula.
Obviously, for non-internet enterprises, a bare AI model — no matter how smart — brings no benefits. Many people call OpenAI’s announced-but-not-yet-available o3 a “PhD-level model,” but you can go ask any freshly minted PhD how their job search is going.
In reality, most jobs don’t require PhD-level intelligence. High school or even middle school level will do.
The funniest “proof that AI doesn’t work” I heard this year was people discovering that GPT-4o thought 4.11 was larger than 4.8. But if you spend any time on social media, you’ll know there are plenty of real humans who can’t manage single-digit arithmetic either — and that doesn’t stop them from performing well at their jobs, especially in Western countries.
I’ve spoken with friends at institutional media outlets. Many believe current AI has limited impact on the media industry, yet they’re simultaneously terrified that GPT-5 or some other company’s new model will upend their work as rumored. This is textbook misplaced anxiety born from flawed assumptions.
In fact, I’ve previously shared a rough workflow. At AI’s current level, it can already fully replace a journalist-editor and autonomously produce a complete feature profile piece with real interviews.
The reason this hasn’t become “reality” — or rather, the reason most media professionals feel AI’s current impact on journalism is limited — isn’t that the AI model lacks capability. It’s that the price advantage isn’t yet large enough for someone to fully engineer this workflow into a product.
If such a generator for interview-based feature articles could bring the production cost of each 10,000-word piece below 300 yuan, it would appear instantly and would quickly proliferate across self-media platforms.
But this has little to do with further improvement in model capability. It actually requires AI companies to ease up on their obsession with capability and start competing on price.
Even most mature IT and internet companies are in a “wait and see” mode on AI integration, as I noted in Don’t Rush Into AI Native: for most products we already know and use, AI cannot become a decisive user need. Users won’t abandon products with enormous advantages in traditional features and network effects just because of an AI feature.
Conversely, those seemingly lumbering legacy internet companies have more time to transform. They can afford to wait until AI becomes more stable — in output quality, performance, security, pricing, and more — rather than shipping a new version every month to keep up with AI’s pace of advancement. Here’s the simplest way to put it:
If you’re a traditional internet giant with your own IDC and you want to introduce an open-source model for an AI feature in your product — and open-source model inference costs are dropping 30% per quarter — that sounds great, but it means if you bought 10,000 GPUs in the first quarter, you’ll only need 7,000 in the second. What do you do with the 3,000 idle GPUs?
The smart move, of course, is to ask users to be patient, or pick a model that’s already plateaued. Because you know this rate of evolution can’t possibly last. When the technology hits a wall, that’s when you ship the feature.
When AI technology enters a relatively stable phase, when large model capabilities are no longer a scarce resource, when AI infrastructure is gradually maturing, we’ll see AI attempt to enter more domains and integrate with increasingly specific scenarios in people’s work and lives. It will succeed in some scenarios and fail in others — and none of it will have anything to do with whether the Scaling Law has hit a wall or not.
Because most people, enterprises, and organizations haven’t yet managed to perfectly integrate even GPT-3.5-level AI into their existing workflows. So what’s the point of prematurely writing off AI or pinning arbitrary hopes on GPT-5?
So, unless you’re building foundation models, stop obsessing over whether the Scaling Law has failed, and stop complaining that AI development has slowed down.
For most people, the AI revolution has only just begun.
