Common Arguments About the AI Bubble

1. AI is in a bubble

Yes, but there is a trap here. It leads many people to think that if there is a bubble in the market, the technology must be a scam.

Bubbles are a feature of human nature, not a property of technology. Every revolutionary technology goes through the same script: early progress exceeds expectations, expectations are raised to even more absurd heights, and eventually valuations become completely detached from fundamentals. You can mock the internet bubble of 1999, but it is hard to deny that its overinvestment built an entire set of infrastructure ahead of time: fiber optics, internet data centers (IDCs), content delivery networks (CDNs), payments, logistics, ad delivery, and more. After the bubble burst, Amazon survived and Pets.com died. The technology did not die; the companies that “raised money on stories and stayed alive on funding” did. Economist William Janeway calls this phenomenon “productive bubbles”: through waste, bubbles pay in advance for society’s infrastructure and room for trial and error.

And trial and error is unavoidable in the development of any technology. You cannot wait until the technology matures and then look in the rearview mirror to judge how absurd and wasteful the earlier bubble was.

2. There are lots of AI scam companies

As above, new technologies give rise to many “scam” companies, as well as companies that were once wildly popular but were quickly eliminated by later technological advances. These things are inevitable, but none of them shows that the technology itself is a scam.

The timing of technological breakthroughs is unpredictable. Of course you can look at the current situation and conclude that “valuations are too high, expectations too full, and supply is expanding too fast.” But you do not know which team, approach, architecture, or product form will suddenly push the boundary a little further at some point.

For example, you think tulip valuations are outrageously high, until a scientist extracts a cancer cure from tulips. Then a short squeeze sends prices to even more outrageous heights. As the saying goes, take fandom far enough and you become a hater; take shorting far enough and you become a buyer.

3. Today’s AI is not AGI; the entire technical approach is wrong

Well…

Yes, the human brain does not need this much electricity or such vast amounts of training data to do these tasks. Just as a horse does not need this much coal or any tracks laid down: it eats grass and handles all sorts of terrain with ease. So the technical approach behind trains must be wrong.

4. AI scaling laws are unsustainable

Parameter growth is often entangled with other innovations: architecture, training objectives, data recipes, inference strategies, tool use, product form, distribution channels, the hardware stack, and even the restructuring of organizational workflows. Give an LSTM as many parameters as a Transformer and it still will not become a new GPT.

“Predicting AI’s economic value from a parameter curve” is about as reliable as economists using theoretical models to predict a country’s future GDP.

5. A report from some university or research institute says AI has limited productivity benefits at large companies; someone claims they deployed Copilot to 4,000 employees and found it useless

If you look closely, the central obstacle these reports repeatedly identify tends to be how AI is put into use, rather than model capability. Organizations have not redesigned their workflows for AI; they just treat it as an “add-on chat box.” Or management puts on a show for KPIs, calling the launch of a ChatGPT-style bot “digital transformation.” The engineers below them pile on a layer of “agents,” but these so-called agents just swap out a system prompt, connect a different vector store, and add a tool-call wrapper. Of course there is no productivity gain: you never brought AI into the complete production workflow in the first place. What the reports measure is how bad outdated management can be, not the upper limit of AI.

I strongly agree with the point Andrew Ng made at Davos in 2025. Earlier technological revolutions spread “top down,” from governments to businesses to individuals. Generative AI, with its low barrier to use, reverses that pattern and spreads “bottom up”: individuals adopt it at scale first, while businesses and governments work to catch up and set rules.

So I trust my own experience more than reports from institutions or assessments by “experts” with all sorts of titles. Next come the experiences shared on social media by top engineers such as Andrej Karpathy, followed by the experiences of ordinary users and independent creators. Within days of OpenClaw taking off, social media was flooded with posts: reviews, installation guides, usage tips, criticism, and attacks. Many may have been clickbait or attempts to attract traffic, but the pace of information updates and feedback far exceeds that of sluggish large companies and governments.

This article was translated from the Chinese original posted on X.