Replacing Humans with AI is Going Horribly Wrong

Hi, welcome to another episode of Cold

Fusion. Since 2023, the fast food chain

Taco Bell introduced artificial

intelligence at over 500 locations in

the United States. The aim was to reduce

mistakes and speed up orders. But in

some cases, AI delivered just the

opposite. Like this frustrated customer.

>> And what will you drink with that?

>> Oh my.

>> I want a large mouth soup.

>> And your drink?

McDonald's drive-throughs also tried AI,

but decided to scrap it because it was

too unreliable. One person had bacon

added to their ice cream in error, and

another had hundreds of dollars worth of

chicken nuggets mistakenly added to

their order. As for Taco Bell, the

fiasco has caused them to rethink their

use of AI. Their chief technology

officer, Dne Matthews, told the Wall

Street Journal in regards to deploying

the voice AI system, quote, "Sometimes

it lets me down, but sometimes it really

surprises me." End quote. And what he

said is the very crux of consumer

generative AI today. It works most of

the time, but about a few% of the time,

it just gets things wrong. But where

does this all lead? A recent MIT report

found that after surveying 150 business

leaders and 350 employees, just 5% of

integrated AI pilots are extracting

millions in value, while the vast

majority remain stuck with no measurable

profit and loss impact. In other words,

AI implementation fails in 95% of the

cases. The market was spooked by these

findings and shares in Nvidia, the $4

trillion company whose chips power the

AI boom, dropped by 3.5%. While

Palanteer fell by 9% off the back of the

news. It all sounds pretty heavy, but

this was just a reaction to the

headline. Because as you dig deeper into

what was actually said in the report,

the picture isn't as clear-cut as AI

simply failing everywhere. A few

episodes ago, I talked about the current

crisis with new graduate jobs. A large

part of that was the AI threat. It's

true and happening in some entry-level

jobs, especially in the creative fields.

I did caveat in that episode that

current AI systems still get a lot

wrong, but will improve in the future.

But today, let's explore that fact a bit

more deeply. What if generative consumer

AI continues to underperform in

businesses for years to come? In this

episode, we'll look at how AI is

actually performing once put into the

position of taking people's jobs. To

summarize the sentiment of this episode,

it's basically this. Consumer generative

AI will revolutionize global

productivity eventually, but as for now,

it's possibly in a bubble.

>> You are watching Till Fusion TV.

>> So, first thing, let's lay it out

straight. AI has some legitimate use

cases and works well in non-critical

areas that don't require 100% precision.

robots, live translation, and even some

prototype website builders like Lovable

are some examples. But the sentiment is

clear. Most people online find it

annoying. But that aside, there's a

fundamental problem with current

generative AI. Let me explain. You see,

everything we call artificial

intelligence today was built on the

findings of a 2017 paper from Google.

That paper provided a way for AI to

focus on different parts of an input

sequence simultaneously and determine

the relevance of each word to every

other word. This innovation called a

transform neural network allowed

researchers, computer scientists and

later companies to change the world. But

here in lies the fundamental problem

with this method of transforming neural

networks. It turns out by determining

the relevance of each word and

predicting the next word, it just makes

up stuff. Essentially, it doesn't know

what it's saying. We call this problem

hallucinations, and it has real world

consequences.

[Music]

Picture this. You're a business that

decided to replace its staff with AI to

write up patient documents, fill in

patient information, summarize meetings,

or do some basic scheduling. But as time

goes on, to your horror, you realize

that the AI system makes up 10% of

everything that it delivers. But the

real problem is you don't know what it's

made up and you don't know what's

accurate. So you or your staff manually

have to go back and check everything.

This ends up just being extra work for

the remaining staff and ultimately is a

waste of time. It sounds stupid, but

this is exactly what's happening now.

The following are Reddit comments taken

from those in the workforce who had to

take the brunt of upper management

thinking that blindly implementing AI

was a good idea. quote, "My company paid

for some AI scheduling software a few

months ago, thinking that it could free

up the accounts team so that they could

continue their hiring freeze in that

department. Now the accounts team is

having to do extra work making sure the

program isn't messing everything up. Not

to mention, our production team that

never had to worry about schedules is

double/ triple checking everything.

They're finally scrapping it for some

basic scheduling software." And same

experience here. Products sold and

demonstrated to be a lifesaver.

Ultimately, the staff that used to do

the work spend all their time making

sure the AI doesn't screw up and help

train it. Also, working in a

medical/clinical setting, we really

don't trust the new file sorting and

labeling system. Names, date of birth,

insurance data has to be perfect. AI is

less than that. It fails at gathering

proper demographic information and

assigning relevant tasks. Sometimes it

thinks the doctor is the patient or

doesn't know where the document was

faxed from, etc. And yet another user

states that AI was good at taking notes

from Zoom meetings, but would make up 5

to 20% of the content, even when handfed

the transcripts of the meeting. Those

who looked over the AI generated summary

realized that some of what was written

wasn't even said.

All of these disasters make sense. LLM

predict the next word statistically, but

they'll never tell you when it doesn't

know the answer or if it can't

understand something. There's many such

stories of company regrets after rushing

into halfbaked AI solutions. A report

suggests that 55% of companies regret

replacing people with AI.

This bank for example, who fired staff

to install an AI chatbot that was so bad

they begged for their old humans back.

And take the example of Cler. 2 years

ago, they implemented a hiring freeze

and began to replace their human staff

with AI. By 2024, their headcount had

fallen from 3,800 to 2,000. But lo and

behold, their customers wanted to speak

to actual humans. Cler said that its AI

chatbots perform the work of 800

employees. But the company admits that

their service quality and customer

satisfaction have dropped and they

lament that human interaction is still

needed.

on replacing people with AI. The

publication Fortune notes, quote, "Not

only is it shortsighted, it's

fundamentally bad business. The

companies cutting people today in the

name of AI will be the ones playing

catch-up tomorrow. There's no doubt that

AI is excellent at doing more with less.

It speeds up processes, cuts down

repetitive work, and buys back time. But

AI on its own cannot create the next

generation of products and services."

End quote.

So on Cold Fusion, we tried to be

thorough here. So it has to be said that

all of these failures aren't the full

story. There are indeed companies that

are extremely successful at using

artificial intelligence.

[Music]

Now, even though the MIT paper said that

95% of companies who implemented

generative AI failed in said

implementation, the same MIT paper

states, quote, "Some large companies,

pilots, and younger startups are really

excelling with generative AI." They mean

startups led by 19 or 20 year olds. They

quote have seen revenues jump from 0 to

20 million in a year. It's because they

pick one pain point, execute it well,

and partner smartly with companies who

use their tools." End quote.

So, how companies adopt AI is crucial.

For example, purchasing AI tools from

specialized vendors and building

partnerships succeed 67% of the time,

while internal builds succeed only

onethird as often.

This goes to show that you can't just

slap AI everywhere and expect it to

work. It needs thought in its

implementation and also depends on the

specificity of the AI tools in question.

But in the grand scheme of things, it's

still early days for AI and it'll be

naive to think that it will stay the

same forever. All it would take is

another groundbreaking paper, a new

underlying neural network architecture

and everything could change again. This

could mean another giant leap forward

that none of us are expecting. But in

the meantime, it's uncomfortable

territory. So what happens?

>> Like the steam engine which sparked the

industrial revolution of the late 1700s,

the internet is changing everything it

touches. And at the cutting edge of the

revolution is Wall Street. In the dotcom

bubble during the mid '90s, everyone who

simply put a.com at the end of their

company name saw massive valuations

because investors who didn't understand

the technology saw them as the future.

In reality, these companies had no solid

way of making a profit or didn't even

have a business plan. When the broader

market realized that it was all a

smokeokc screen, the sector crashed.

Most doc companies went out of business

and only a handful made it out and are

giants today. So, let's compare that to

the AI wave of today. The long-awaited

release of chat GPT5 was a

disappointment, and many users even

thought the previous version was better,

and this caused Open AI to scramble.

In addition, the company was also caught

blatantly fudging the performance

numbers in some very strange graphs.

Incremental rather than revolutionary

was the tone.

Soon Meta would announce that it's

downsizing its AI division and this

comes amongst a growing chorus of

analysts saying that AI is heading

towards if not is already in a bubble.

And next we have the massive valuations

and spending. Nvidia H100 the GPUs that

power the artificial intelligence boom

are about 30 to $40,000 each. Google has

26,000 of them and they've managed to

create Alpha Fold, Gemini, V3 and more.

But Meta on the other hand has 600,000

Nvidia H100s. And while they do have the

open- source Llama LLM, it isn't

discovering new science like Google's

Alpha Fold for 23 times the compute. And

to give you an idea of how extreme this

is all getting, AI itself has caused a

4% increase in electricity use in the

US. Morgan Stanley states that data

center investment will reach $3 trillion

over the next 3 years in preparation for

AI use. And of course, that's heavily

fueled by debt.

The belief is that AI will cut costs by

40% and that should add 16 trillion to

the S&P. But as we've just seen at the

beginning of this episode, according to

that MIT study, that could be very

unrealistic.

So, this could be the future if AI

doesn't improve massively soon. Number

one, business executives and business

owners will get frustrated with

hallucinations. useless solutions, bad

code, and a very poor return on

investment.

Number two, a lot of the AI gurus like

Sam Olman will have to admit that

artificial general intelligence isn't

going to be achieved by LLMs, and

they're essentially a dead end. Number

three, the general populace begins to

get sick of LLMs, and we're already

seeing it. the absolute flood of AI

slop, the hallucinations, AI agreeing

with what users say, sometimes driving

them insane.

Number four, in seeing this, the venture

capital finally dries up and LLMs become

just too expensive to justify unless

there's a massive increase in

efficiency. The cost to run OpenAI's

data centers, all the pipes and guts and

things that like keep AI running is

about $40 billion a year. Their revenues

right now are only like 15 to 20

billion. And finally, number five, after

a long winter, new implementations come

around that truly live up to the hype

promised by the first wave.

So, just like the com bubble, there's

going to be a few winners that rise from

the ashes. But if or when it crashes,

from here we see the true artificial

intelligence companies that last the

distance.

So to finish off this episode, let me

end with this. I've talked about this

many times in my older episodes, but

take a look at this diagram. It's called

the Gartner hype cycle, and it describes

the typical progression of the

infiltration of new technologies into

society.

So where do you think we are? The

technology trigger, the peak of inflated

expectations, the trough of

disillusionment, the slope of

enlightenment, or the plateau of

productivity? Feel free to comment down

below. So where to from here? Well, Sam

Oldman and all the other AI leaders

should focus their efforts on fixing the

hallucinations.

Once again, a different neural network

architecture could be discovered, one

that fixes hallucinations.

Or perhaps it could just be fixed

manually. Either way, ushering in a new

boom. But that all being said, and

that's the funny thing about AI. The

point is, with AI, nobody knows the

future. So, what do you guys think? Do

you think we're in a bubble and a crash

is imminent, or do you think the next AI

innovation is just around the corner?

Now, if you've been recently hired by a

company or just simply need to brush up

on your knowledge, today's sponsor is

perfect for you. Brilliant is the best

way to learn subjects like computer

science, data analysis, and AI, but in a

way that's interactive and hands-on.

Brilliant gives you puzzles and

step-by-step challenges where you learn

by doing. It's proven to be six times

more effective than watching lecture

videos. Their courses cover everything

from the fundamentals of neural networks

and probability to the mathematics that

underpins AI and even everyday topics

like algorithms and data security. And

because all their content is crafted by

experts from places like Stanford, MIT,

Caltech, Microsoft, Google, and more,

you can trust that what you're learning

is not just accurate, but also useful.

Learn at your own pace to brush up on a

project for work or just for your own

self-development. Whether it's leveling

up or learning new skills, I highly

recommend checking it out. To get

started for free, head to

brilliant.org/coldfusion

or scan the QR code on screen or click

the link in the description. Brilliant

also given our viewers 20% off an annual

premium subscription, which gives you

unlimited daily access to everything on

Brilliant. Hey guys, it's me. Hi. I'm

not an AI, so you're finally seeing my

face. But anyway, thanks for watching.

If you did enjoy it, there's plenty of

interesting stuff here on Cold Fusion,

so feel free to subscribe. Otherwise,

that's about it from me. My name is

Togo, and you've been watching Cold

Fusion. And I'll catch you again soon

for the next one.

>> Cold Fusion, it's new thinking.

[Music]