Google and Excel just reviewed over

4,000 AS startups for their accelerator

program, but they rejected 70% easily

because they were just useless rappers,

thin layers on top of existing models

with nothing special proprietary

underneath. Out of 4,000 applications,

they ended up only accepting five into

the actual program. In this video, I

want to talk to you about why most AI

companies are lies and why you could

maybe even build a better product

yourself if you were just honest. You

might have seen videos from Anthony Syly

or Moar covering individual cases of AI

companies faking their technology

recently. Now, those videos are great,

but what I want to show you today is

that those individual cases are symptoms

of something much bigger. This is a

systemic problem in the software

industry backed by data, regulatory

filings, and peer-reviewed science. So,

if you want to learn the truth, well,

keep watching. This article from

TechBuzz breaks down the numbers from

Google and Excel's joint AI accelerator

called Atoms. The 70% rejection rate is

bad enough on its own, but the economics

behind these kinds of rapper companies

are even worse. Their gross margins sit

around 23% compared to 80% or higher for

traditional SAS. API costs often eat 45

to 60% of the revenue already because

you know AI models are expensive and

none of these companies actually own

their models. So the business model is

structurally broken from day one. Google

Cloud FP Maui said it directly. If

you're just counting on the backend

model to do all the work and you're

simply white labeling it, the industry

doesn't have patience for that anymore.

And we already saw this play out with

Jasper AI, which was one of the poster

childs for this model of in a way just

reselling the AI models. Jasper reached

90 million in annual revenue and a $1.5

billion valuation as a rising tool on

GBT. And to be fair, it does work well.

Those are rapper companies. So, these

are companies that aren't necessarily

lying, but have nothing real under the

hood. Now, sometimes they are still

lying because they might put something

on their front page saying that they've

trained an AI model on your data, which

is often not true. Nothing is trained at

all. They're just putting some of your

own data in their prompts. But some

companies here go further and actively

lie about their AI systems and what they

do. The SEC and FTC have started

criminally charging some of those

founders. And this article from Stone

lays out the full enforcement timeline.

Some of this started all the way back in

March 2024 when two investment advisory

firms, Calia and Global Predictions,

settled SEC charges for claiming they

used AI that didn't exist and the

combined penalty was $400,000.

Then in January 2025, Presto Automation

got hit, a publicly traded drive-through

AI company, no less, that claimed 95% of

orders needed no human help. Now, the

reality was that over 70% required human

intervention, and at some locations, it

just came close to 100%. The core voice

AI technology that they used wasn't even

theirs. It was owned by a third party.

And then there's Nate Incorporated from

April 2025, which was a shopping app

that raised $42 million and the actual

AI automation rate was effectively 0%.

The entire product was powered by a call

center of hundreds of workers in the

Philippines. Now, I'm not saying you

cannot make an AI system like this work,

but you have to do more than this. You

cannot just lie away out of it. The

founder was indicted on securities fraud

and wire fraud, each carrying a maximum

of 20 years. So, he's looking at 40

years in prison. The SEC described it as

an old school fraud using new school

buzzwords. And Nate is not unique in

this pattern. You might have heard of

Builder.ai, which raised $450 million

for Microsoft, Soft Bank, and more,

which claimed an AI called Natasha could

build apps autonomously. The company

filed for bankruptcy in June 2025 with

liabilities reaching $100 million

against assets under 10 million.

Amazon had a similar situation with

their just walk out technology in

cashless stores where roughly a thousand

workers in India were manually reviewing

70% of transactions and they quietly

face it out in April 2024. Again here I

fully believe that we will be able to

create AI systems that do this kind of

work but companies aren't always upfront

that they're still in a training process

and this really is the problem with the

current bubble of AI as well. If you're

interested to learn how AI works beyond

just prompting an LLM, well, that's what

I made this channel for. And I also have

a set of free open source local AI

projects linked in the description below

that can walk you through building

things from scratch. And this way, I

just hope to push the industry to a

better standard, honestly. So, why does

this keep happening? Now, it's not just

greed that people just want to make a

quick buck. Researchers at Alto

University in Finland ran many

experiments where people use Chachi Piti

to solve certain school admission

reasoning problems and every single

group of users overestimated their

performance after using AI regardless of

their actual ability level. Now normally

with this kind of Dunning Kruger effect

the least skilled people are the most

overconfident but with AI that flips

completely. Sometimes people with higher

AI literacy, the power users who use

these tools the most show the greatest

overconfidence. And we are seeing this

happening in the founder space. Young

people with not that much career

experience just think that they're

geniuses when they can create an MVP

with clawed code. Professor Robin Welch

put it clearly. When it comes to AI, the

Dunning Krueger effect vanishes. What's

really surprising is the higher AI

literacy brings more overconfidence.

Picture

this, a non-technical founder sitting

down with claude code five coding a demo

in three hours while the AI tells him

it's brilliant the entire time. It will

never push back. It will never say this

won't scale and it might not even say

that they need a real engineer for this.

After that experience, the founder is

just completely gassed up and genuinely

believes that they built something

amazing. Some of these people really

aren't cynical grifters, but the tools

themselves just made them delusional.

88% of Y Combinator's latest batch is AI

native and a quarter of the W25 batch

had code bases that were 95% AI

generated which makes sense because you

can just throw together a quick demo and

impress a couple of venture capitalists.

Now I've shipped production software for

years and the gap between a demo and a

real product is enormous but the tools

make that gap invisible to the person

building the demo. The whole startup

ecosystem is running on overconfidence.

And the AI itself is the engine that's

going to keep this going for years. So

next time you see an AI product launch

go viral with a really slick 60-second

demo, ask yourself whether that's a real

product with real engineering behind it

or a prompt rapper that someone vibe

coded and that will not be supported in

2 years from now. Because based on the

data we just looked at, the odds say

it's probably more of a scammy product

70% of the time. If you want to work

towards a world where we have really

useful AI products that go beyond just

prompting an LLM, check out my set of

free open source projects linked in the

description below that teach you how to

build real things from scratch. Now,

subscribe if you want more AI

engineering truth like this one, and

then I'll see you next time.