For Proper GIF experience, please read these articles on :
https://goelh.substack.com/p/andrew-ng-made-five-predictions-the
and
https://medium.com/@hardik.goel214/andrew-ng-made-five-predictions-the-internet-heard-millionaires-both-missed-the-point-8a9b3928476a
The viral thread is mostly right about the technology. It’s precisely wrong about where the money is. And its own last paragraph, the one everyone scrolled past, is accidentally the whole story.
This one is a bit different article from the ones you must be reading as instead of the architectures, this one talks about the possible future from one of the greatest person in this field whose prediction is really worth writing :)
There’s a thread doing the rounds right now, and you’ve probably seen it, because these things travel faster than the ideas inside them. It goes: Andrew Ng called deep learning in 2008, called the online-education boom in 2011, called China’s AI rise in 2014, and now he’s revealed five opportunities that will create more millionaires than anything before.
Let me get something out of the way first. The credentials are real. Ng built Google Brain, co-founded Coursera, ran AI at Baidu, and has taught more people machine learning than possibly any human alive. When he talks, listening is a reasonable default. I listen.
So I’m not here to argue with the predictions. Most of them are directionally sound, and I’ll walk you through them.
I’m here to argue with the word opportunity. Because there’s a pattern running through all five predictions that the thread never names, and once you see it, the “millionaires” framing doesn’t just look optimistic. It looks backwards.
The five, in two minutes
One: agentic AI beats scaling. Ng has been banging this drum since early 2024, and he’s earned the right: his benchmark work showed that a smaller model wrapped in an agentic loop, reflection, tool use, planning, multi-agent collaboration, can outperform a much bigger model prompted once. This is the least controversial item on the list. It’s not a prediction anymore. It’s a job description.
Two: defense AI is real now. In February 2025, Ng publicly welcomed Google dropping its pledge against AI weapons work, which made a lot of Silicon Valley clutch its oat milk. Whatever your ethics on this, and mine are complicated, the observation stands: the taboo is gone, the budgets are enormous, and pretending otherwise is not analysis, it’s aesthetics.
Three: AGI is decades away. Ng’s test is my favorite piece of the thread, because it’s the rare AGI take that’s falsifiable: until companies can fire all their intellectual workers, you don’t have AGI. By that bar we are nowhere close, and his advice follows naturally, go build boring, profitable things while everyone else debates consciousness on a podcast.
Four: China can win through open source. Not by building the biggest model. By shipping faster, cheaper, and open, until the world’s default stack quietly becomes a Chinese one. If you’ve priced DeepSeek or Qwen against the frontier labs recently, you already know this isn’t a hypothetical. It’s an invoice.
Five: small specialized models beat giants. Token prices have collapsed, edge hardware is exploding, and a fine-tuned small model on a phone or a ₹8,000 device does most jobs a giant does, minus the cloud bill, the latency, and the privacy lawyer.
A quick architect’s footnote before we continue: the thread garnishes each prediction with hockey-stick market numbers, $5.1B to $69B, $930M to $5.45B, precise to the decimal about the year 2032. These come from the market-report mills, the ones that will happily project the 2035 revenue of an industry that doesn’t exist yet. Treat every such number as vibes wearing a spreadsheet. The directions are right. The decimals are theater.
Now the pattern the thread doesn’t name
Read the five again, but this time, ignore the topics and watch what each one does to price.
Agentic workflows: get frontier-level results from cheap models. Small models: get results from cheaper hardware. Chinese open source: get the models themselves at a tenth the cost. Even the AGI take is deflationary, it says the magic isn’t coming, so stop paying magic prices and buy tools instead. Four of five predictions, and honestly the defense one too if you squint at drone economics, are the same sentence wearing different hats:
AI capability is getting radically cheaper and more available to everyone.
08a0e9aa-4c59-4d89-86a5-350d58e39a03_960x540.jpg
Which is wonderful. Genuinely, civilizationally wonderful. But sit with what it means for the word opportunity, because here’s the thing two decades of watching technology waves has taught me, and it’s the part every “get rich off this trend” thread structurally cannot say out loud:
A capability that everyone can afford is a capability nobody can charge a premium for.
When only Google could do deep learning, deep learning was a moat. When anyone with a laptop and a Colab notebook can do it, it’s a utility. Ng’s five predictions, taken together, are a forecast that AI is completing its journey from moat to utility. Electricity created fortunes too, but not for the millions of people who used electricity. Using it just became the price of admission.
So when the thread says these five trends will mint more millionaires than ever, ask the only question that matters: if everyone can deploy the same cheap agentic stack on the same open models on the same $99 edge box, what exactly is scarce?
2 (1).jpg
The funny thing is, the thread answers its own question. It just does it in the last paragraph, in a tone that suggests the author was already reaching for the “post” button.
The buried lede
After five predictions and a lot of exclamation points, the thread ends with what reads like a compliance-vendor’s LinkedIn caption: customers need to trust your AI, regulators demand explainable models, the winners deploy validated, governed, monitored systems, not just more of them.
It’s framed as a footnote. It’s actually the entire thesis, and I’d bet the author doesn’t fully realize what they wrote.
Because in a world where the models are commodities, the workflows are open-sourced design patterns, and the hardware costs less than a decent dinner, trust is the one thing on the list that doesn’t deflate. You can’t pip install a track record. You can’t download a regulator’s sign-off from Hugging Face. You can’t fine-tune your way into a hospital’s, a bank’s, or a defense ministry’s confidence. Trust is accumulated slowly, verified painfully, and destroyed instantly, which is precisely the economic profile of a thing worth owning.
Look at where the money actually pooled in previous waves. The web made publishing free, and the fortunes went to whoever organized the resulting chaos into something people trusted, a search box, a blue feed. Cloud made compute cheap, and the fortunes went to whoever enterprises trusted to run it without losing their data. The pattern is boring and undefeated: when creation gets cheap, the premium migrates to verification.
3.jpg
Ng’s five predictions describe creation getting cheap. The last paragraph describes where the premium goes. The thread thinks it wrote five predictions and a disclaimer. It wrote one prediction and its consequence.
So what do you actually do with this
Architect hat on, hype off. If I were advising a founder, or the version of me that reads threads like this at 11pm, here’s the translation from prophecy to Monday morning.
Don’t sell the capability, sell the accountability. Anyone can wire up an agentic workflow now, the four design patterns are public, that’s the point of them. The buyer’s real question has already moved on from “can it work?” to “who do I call when it doesn’t, and can they prove it usually does?” Build for the second question. It has fewer competitors and better margins.
Pick a domain where being wrong is expensive. The whole reason small models plus edge deployment is exciting, medical AI on phones, factory-floor AI, retail decisions in real time, is that these are places where mistakes cost real money or real safety. Which means these are exactly the places where the vendor who shows up with evaluation results, audit trails, and a straight answer about failure modes beats the vendor who shows up with a bigger demo. Cheap models made entry easy. Consequences keep it profitable.
4.jpg
Treat Ng’s AGI test as a business filter. “Would this still be a good business if AGI never arrives?” is a shockingly effective way to sort real companies from prompt-wrappers with a pitch deck. His own portfolio strategy, boring, profitable, specific, is the tell. The man who could bet on anything is betting on plumbing.
And one uncomfortable meta-point, offered with affection to everyone screenshotting the thread: a prediction that reaches you via a viral post has, by definition, already reached everyone else. Widely shared maps don’t lead to unclaimed gold. They lead to crowds. The value was never in knowing the five trends. It’s in noticing the sixth thing, the one the thread mentioned last, in its smallest voice.
The internet read that thread and heard five ways to get rich. I read it and heard something quieter and more useful: the models are about to be free, the workflows are about to be folklore, and the last scarce thing in AI will be the oldest scarce thing in business.
Being the one they believe.
-Hardik
Buy me a coffee:
buymeacoffee.com/HardikGoel
I’ve spent about two decades building data and AI systems, long enough to have watched three separate technologies complete the journey from “moat” to “checkbox” while the threads were still promising millionaires. I write about AI in the real world, the economics, the hype cycles, and the last paragraphs everyone scrolls past. If this changed how you’ll read the next viral prediction thread, subscribe. There’s always another gold rush being announced, usually by someone selling maps.
Sources / further reading: Andrew Ng on agentic design patterns (reflection, tool use, planning, multi-agent), The Batch / deeplearning.ai and his 2024 Sequoia AI Ascent talk. Ng’s February 2025 comments on Google’s revised AI-weapons stance (Military Veteran Startup Conference, as reported by TechCrunch). Ng on AGI being overhyped and “many decades away,” various interviews 2024–25. Ng on China’s open-source path to AI leadership, 2025 interviews and The Batch commentary. Market-size projections cited in the viral thread trace to commercial research reports (MarketsandMarkets and similar) and should be treated as directional at best. Note: this piece responds to a widely circulated social media summary of Ng’s views; where the summary and the primary sources diverge, trust the primary sources, and ideally, the man’s actual newsletter.