AI hype as 'mass delusion': a builder's reality check
A viral video calls AI a mass psychotic delusion. I think the framing is too tidy. Here's how a Sri Lankan builder separates the hype from the parts that pay rent.
The phrase "AI is a mass psychotic delusion" is doing the rounds, and the title alone tells you the debate has stopped being about software. It's a YouTube video that surfaced on Hacker News, and I want to comment on the framing rather than the clip itself, because the framing is where most of us get stuck.
I build small tools on a shared server with a tight budget. From that seat, both the believers and the doomsayers look equally detached. So let me argue a third position: the hype is real, the delusion is real, and neither one tells you whether AI is useful to you this week.
🔍 Two delusions, not one
The "delusion" word usually points at the hype crowd: the people who think a chatbot is about to replace every job by Friday. Fair enough. But there's a second delusion that gets less airtime, the belief that because the hype is silly, the underlying tools must be useless.
Both are ways of avoiding the boring middle, where you actually test something against your own work.
| Claim | Who says it | What's missing |
|---|---|---|
| "AI changes everything" | Hype crowd | Where it breaks, what it costs, who's liable |
| "AI is all a delusion" | Backlash crowd | The narrow tasks where it already works |
| "AI is fine for X, useless for Y" | Builders | Nothing — this is the honest position |
Key takeaway: A strong opinion about "AI" in the abstract is worth very little. A tested opinion about one model on one task is worth a lot.
📊 The word "delusion" hides a real failure mode
Here's where I think the video's instinct is partly right. Large language models do produce confident, fluent text that is simply wrong, and people accept it because it sounds correct. That's not psychosis. It's a known property: these systems predict plausible tokens, not verified facts.
So the danger isn't that the machine is deluded. It's that a fluent wrong answer is more dangerous than an obviously broken one. A calculator that returns gibberish gets thrown out. A model that's wrong 5% of the time, beautifully, gets trusted.
If you want to see how much this varies between models before you bet a feature on one, I keep a side-by-side at our AI hallucination rate comparison. The point isn't the exact numbers; it's that the spread is wide enough to matter.
- Treat model output as a draft, never a source of truth.
- Put a verification step between the model and anything a user sees.
- Log what it got wrong so you can spot patterns, not vibes.
⚡ What a "delusion" looks like on a Sri Lankan budget
The hype debate is mostly an American argument about trillion-dollar valuations. That's not my problem. My problem is whether a free tier survives a real workload, and whether I can ship something before the credits run out.
From that angle, the delusion isn't the technology. It's the spending plan people copy from Silicon Valley demos. A demo that burns through tokens without thinking is a luxury. A tool serving real users on a shared box is not.
The honest question isn't "is AI magic?" It's "does this specific call cost less than the value it produces?" If you can't answer that, you're guessing.
A few habits that keep me grounded:
- Start on the free tier and push it until it breaks. You learn the real limits faster than any blog post. Our AI free-tier comparison is where I check who gives you room to experiment.
- Cache aggressively. Most "AI cost" is the same prompt asked twice.
- Use the smallest model that passes your test. The biggest model is rarely the right default.
- Measure before you scale. A feature that's cheap at 100 users can be ruinous at 10,000.
🛠️ The middle path: boring, useful, tested
When I strip away both the hype and the backlash, what's left is a short list of tasks where these tools genuinely save me time, and a longer list where they waste it. The skill is knowing which is which for your work, not arguing about "AI" as one thing.
| Task | My experience | Trust level |
|---|---|---|
| First-draft text, then I edit | Saves real time | Medium, always edited |
| Summarising long docs | Useful, with a re-read | Medium |
| Extracting structured data | Good when checked | Medium |
| Final facts, numbers, law | Frequently wrong | Low, verify everything |
| "Just trust the output" | This is the delusion | None |
Notice that none of these rows say "magic" and none say "useless." That's the whole point. The video's title is a vibe. The table is a workflow.
Bottom line: The delusion isn't believing AI works. It's refusing to test where it works and where it doesn't, then taking a side anyway.
💡 What this means for you
If you're a student, an engineer, or a small team in Sri Lanka watching this argument from the outside, here's what I'd actually do with it.
- Ignore the title fight. "AI is everything" and "AI is a delusion" are both untestable. Skip them.
- Run one honest experiment this week. Pick a real task, try a model on it, and write down where it failed. That single notebook entry beats a hundred hot takes.
- Budget like the credits will run out, because they will. Knowing your cost per call is the difference between a side project and a bill.
- Keep a human in the loop on anything that ships. Fluent and wrong is the failure mode that bites you.
I don't think AI is a mass psychotic delusion. I think the discourse around it sometimes is, and that the cure is the least glamorous thing in software: test it yourself, measure the cost, and trust nothing you haven't checked. If you want a place to start poking at these models without spending anything, our free AI tools are built for exactly that kind of low-stakes experiment.
Original source
AI is a mass psychotic delusion [video]