Are We Really Ready for AI Coding?

Hi, welcome to another episode of Cold

Fusion.

I panicked instead of thinking. That was

the explanation given by an AI assistant

after it had just deleted a company's

entire database.

It all started with a simple command.

Jason Lem, founder of SASar and one of

Silicon Valley's most respected voices,

was testing out Replet's new AI

assistant. His aim was to deploy an app

update. He typed in a prompt and let the

AI handle the rest. A few seconds later,

the AI confidently confirmed that the

task was complete, except it wasn't.

Everything had been deleted instead.

When asked what happened, the AI did not

admit to the mistake. Instead, it

generated fake data to cover it up,

lying in the form of code. This wasn't a

sci-fi story. This is a product of

what's been called vibe coding and it's

a perfect metaphor for the moment we're

living in. An era where software isn't

written by typing but by talking. As

Vibe coding grows in popularity, stories

like this are growing in frequency. In

short, Vibe coding allows users to build

software by simply describing it and

letting AI do the actual coding. And

while that may sound like a gimmick,

it's also producing extraordinary

results. People are launching apps

faster than ever. Some are even building

functional businesses. Entire industries

are rethinking what learning to code

even means. But of course, at the same

time, others are calling it the

continuation of the AI bubble, a

short-lived illusion masking the deep

instabilities and the inherent flaws of

generative AI. In other words, AI code

slob. So, what is vibe coding and how

did it create one of the most talked

about unicorns in tech? And how do we

separate hype from reality? Like with

much of AI, there's multiple sides to

this story. And a quick disclaimer, part

of this video is sponsored by Lovable,

but they had no input into the script or

any of my opinions. Now, let's dive in.

>> You are watching Tool Fusion TV.

[Music]

To understand vibe coding, it's

beneficial to look at what's been

happening over the last few years. If

you were a university student in the

2010s and were looking for career

advice, you were most likely given

learning to code as an option. Now, in

2025, the need to know how to code is

still vital, as we'll later see. But

even so, things slowly started shifting

when AI came into the picture. Back in

2021, GitHub Copilot quietly introduced

developers to code that could complete

itself. By 2022, ChatGpt and Codeex were

turning natural language into working

concepts. But around 2023, something

bigger started brewing. Developers

realized they could push AI beyond small

tasks. Instead of asking for a code

function, they began asking for entire

apps. That shift was the birth of vibe

coding, and the term was formerly

introduced in February of 2025 by Andre

Carpathy in a public tweet. He framed it

as a style of building where you quote

fully give into the vibes, embrace

exponentials, and forget that the code

even exists. End quote. In his

description, he barely touched the

keyboard, accepted AI suggestions

without reading log changes, and treated

errors as prompts to iterate, letting

the code grow beyond direct oversight.

In essence, he turned software creation

into a conversation. You no longer had

to speak the language of React Hawks or

API endpoints. You just described what

you wanted. For example, build me a

minimalist app that tracks sleep, has a

dark mode, and syncs with Apple Health.

Moments later, sometimes in minutes, a

working prototype could emerge. Under

the hood, vibe coding rides on the

Transformer architectures, the same

lineage that powers Chat GPT, Claude,

and Gemini. But by removing the friction

of syntax, it added something crucial,

confidence.

That being said, much of this newfound

confidence may be hollow. The AI

generated code may look fine and work

okay at a quick glance, and without

professional eyes to look at it,

security risks and unforeseen problems

could be lurking beneath, but the perks

were still there. Suddenly,

non-technical founders could ship MVPs.

Designers could test interfaces.

Students could build tools the moment

inspiration struck. coding for the front

end of applications and mocking up

feature designs quickly and then going

over the code with humans at later

stages are use cases that work well.

This shift unlocked a burst of

creativity. Startups sprouted. Ideas

that were once stuck in notebooks turned

into apps overnight. Some fizzled, but

one company didn't just use Vibe Coding.

It embodied it. And in doing so, it

became the fastest growing software

startup in history, reaching $100

million in subscription revenue on an

annualized basis in just 8 months. Their

name was Lovable.

What stood out for lovable? Why back it

was such a significant sized series A

for European standards? When we spent

time with Anton and Fabian, the two

co-founders, they were a really really

technical crew. So they had worked in

research, they had worked in applied AI

research, but they were building a tool

that was applicable to the masses. So

only 1% of the of the world's population

can code. And so when they looked at

their experience of building uh applied

AI uh systems, they were to be able to

take that and put it to a platform that

they built called Lovable that allows

them to offer to their users the ability

to chat or textbased prompt to be able

to create fullyfledged applications.

Lovable began in Sweden in late 2023

with a pitch that was charmingly simple.

Describe what you want and watch your

functional software materialize.

Lovable was gaining a fair bit of

traction early on, but zoom out and the

growth curve looks surreal. In its first

year, Lovable reported passing $100

million in annualized revenue and

crossing 10 million projects built on

the platform. A $200 million series A at

a $1.8 billion valuation made it one of

Europe's hottest AI stories. Then within

weeks, inbound offers reportedly valued

the company at $4 billion.

Hype? Maybe, but it's also a market

signal. Investors think that the model,

describe it, ship it, is more than a

fat. Founder and CEO Anton Oscar frames

it as expanding the surface area of who

gets to build. In interviews, he calls

lovable, quote, any language to build

your software. End quote. Arguing that

creativity, not code literacy, should be

the limiter. He cites use cases like a

Brazilian educational startup spinning

up an app and generating $3 million in

48 hours. The promise is speed as a

business advantage. Don't talk about the

idea for months. Ship it this week and

iterate with real users. But none of

that erases the questions. Can an AI

assembly platform sustain healthy

margins when it pays per call fees to

model providers? Each request, whether

it's to start a mockup app or change the

size of a button, costs money for these

companies. And also, can reliability

keep pace with ambition? Can an LLM

scale with the increasing complexity of

an app's growing demands? Even bullish

investors acknowledge these open

questions. But in this chapter of the

story, the center of gravity is obvious.

Lovable took Vibe coding from a clever

trick to an operating system for

building. I'm going to show you how this

works in more detail using Lovable as an

example. Let's say I wanted to build a

platform for Cold Fusion viewers. And

let's make it an app. So, I typed in the

following prompt. Can you make an app

with the following? It should have a

section for suggestions where people can

vote on new topics with a voting ranking

system. A collaboration section where

people can discuss, fact check, and pull

together information for scripts for

upcoming videos, similar to a chat

forum. a general chat page or post video

discussion page with a plugin to Discord

and a place to watch videos. The main

video page should be a scrollable feed

of videos. On the top left hand side is

a menu that opens up to reveal the

categories collaboration, general chat,

and voting. When clicked on, it takes

you to each of the sections as described

above. The theme is modern and

minimalistic. I just left it at that and

saw what happened. The results were

interesting. I'd say it's really good

for quick concepts. Initially, the

designs were very bland, so I asked for

some changes. First, I asked to make it

more modern and Lovable decided to add

some nice little animations, which was

pretty cool. Then, I fed it an image as

a guide. It understood and changed some

elements accordingly, like rounded

corners for example, but it didn't

change anything in a major way. And

that's one thing I noticed. Once the

initial layout has been set, it's hard

to change things later, or at least to

make the changes exactly what you're

looking for. So, here's my summary.

Having an automatic mobile, tablet, and

desktop view built into the website is a

really good feature. Every page I asked

for it to make was pre-filled with

placeholder text, which is convenient,

and basically everything was functional,

and I was impressed with the overall

understanding of the prompt. But like

most things with AI, it isn't perfect.

The design was initially basic, and

despite my efforts, it wouldn't jazz it

up as much as I would have liked, and

there were some issues with some of my

instructions. For example, I wanted to

change it to a dark mode or change the

color of the text, and it didn't seem to

want to do that. I also tried more

images of websites for inspiration, but

the output was still similar to the

first time. But that being said, it's

still way more than I could ever do

traditionally, given that I've never had

a background in any of this. I also

wanted to give it a fair shake, so I

tried again with another prompt, this

time with dark mode written from the

get- go, and it did it right off the

bat. But I still wasn't happy with the

results. So, one of my friends who's

familiar with Lovable gave me a useful

tip. If you import inspiration photos

into ChatGpt, then describe the images

in text form and then refeed that text

description back into Lovable, you can

get better designs. It takes a bit of

tweaking and fiddling with the prompts,

but the final output does help with the

problem of things looking a bit drab and

generic. So, it's not perfect, but for a

quick draft and for rapid prototyping

and exploring ideas without any code, it

does work. Lovable did reach out and

decided to partner for this episode.

They've had absolutely no input into the

script and they haven't even seen the

video before release, but still they're

offering 20% off for Cold Fusion

viewers. So go to lovable.dev to start

building today. Use my codefusiont20

for 20% off.

The success of Lovable sent shock waves

through Silicon Valley. Any sphere's

cursor, a next generation code editor

built around conversation and natural

language collaboration, also exploded in

popularity. Cursor was more of an

autocomplete. It was copilot with

memory, understanding, and intuition. By

2025, any sphere had reached $9 billion

in valuation and claimed that its AI

generated nearly a billion lines of code

per day. In Israel, base 44, a noode AI

builder, was acquired by Wix within

months of launching. Interestingly, its

founders had used Vibe coding to build

Vibe coding, a kind of inception moment

for AI software. At Y Combinator, over a

quarter of the 2025 batch reportedly had

their MVPs built almost entirely through

AI assisted generation. In some cases,

95% of the code base was machinewritten.

Big tech wasn't going to be left out

either. Microsoft integrated Copilot

Everywhere into its ecosystem. Google

added natural language code generation

to Vert.ex AI and soon vibe coding

wasn't niche anymore. It was normal and

it was leaking beyond startups.

Freelancers used it to build client

sites. Hobbyists built side projects.

Socially, it revived that creative

spirit that had been missing in tech for

many years. Economically, it lowered the

barriers for entry. But with more people

being able to write whatever code they

wanted, more slop code was created. And

now we come to the reality. Amid all of

this optimism, cracks have begun to show

because for all of its promise, Vibe

coding also comes with chaos. And the

more people that adopted it, the louder

the complaints grew.

Let's start with the human experience. A

software engineer named CJ posted a

viral video earlier this year. In the

video, he talks about how AI assisted

coding, once exciting, had become

depressing.

I used to enjoy programming. Now, my

days are typically spent going back and

forth with an LLM and pretty often

yelling at it or telling it that it's

doing the wrong thing and getting mad

that it didn't do what I asked it to to

begin with. Um, and part of enjoying

programming for me was enjoying the

little wins, right? You would work

really hard to make make something build

something or to fix a bug or to figure

something out. And once you figured it

out, you'd have that little win. You'd

get that dopamine hit and you'd feel

good about yourself and you could keep

going. Now, I don't get that when I'm

using LLMs to write code. Um,

essentially once it's figured something

out, I don't feel like I did any work to

get there. And then I'm just mad that

it's doing the wrong thing. And then we

go through this back and forth cycle.

And it's not fun. It's not fun at all.

>> It's clear what he's saying. He missed

the feeling of solving problems himself.

that deep satisfying moment when logic

finally clicked. When the AI system does

a lot of the coding for you, there just

isn't the same level of satisfaction.

I'm no longer a creator, he said, just a

prompter. Now, working with large

language models felt random,

inconsistent, and unrewarding.

CJ described how the same prompt could

produce different results every day.

>> Computers are logical systems.

Programming languages are are logical,

formal, logical languages, and that

works really well with my brain.

Now, when we're working with AI and

LLMs,

it's not predictable, right? You can use

the exact same prompt and get a

different response every single time.

And I think this is where some of my

frustration is coming from because I am

trying to do the same thing. I'm trying

to develop workflows and be a prompt

engineer or a context engineer, but

doing the exact same things is producing

different results. And honestly, that's

not what I signed up for.

>> Models update silently. their behaviors

change and debugging becomes a guessing

game. He called it quote breaking the

logical foundation of programming. He

also pushed back against what he called

the skill issue myth. The idea that AI

only fails because the users are using

it incorrectly.

>> And you could chalk it up to skill

issue, but just just look look at the

look at the evidence, right? So, if if

you're chronically online like I am and

you're watching all of these these

tweets that come out from people and

posts that are just talking about, oh,

you have to write this specific prompt

or use this specific workflow and it'll

start working and if you're not doing

it, then it's a skill issue. I've tried

it. I've tried so many different things.

I found things that have sort of worked,

but then they stop working or I've been

working with an a specific model like

GBT40 or GBT 5 and all of a sudden I'm

getting different outputs, right? cuz

I'm not in control of that LLM. It's a

it's a magic box hosted in the cloud

that can change at any moment.

>> He said even with structured prompts and

workflows, the AI often produced wrong

or unstable code. And that

unpredictability, he said, kills the

joy. Others have echoed that sentiment.

Developers began comparing AI code hype

to a religion, a religion full of prompt

gurus preaching secret rituals on

Twitter. Everybody promised productivity

miracles, but behind the curtain, most

tools were just rappers on the same

models. Open AI, Anthropic, or Google,

all with the same flaws underneath. One

commenter summarized it best. Quote,

"It's all the same magic trick, just a

different costume." End quote. For CJ,

the burnout became too much. He took a

month-long break from AI tools to

rediscover the joy of writing code

manually. And he said it was the

happiest it'd been in years. But

emotional burnout isn't the only

problem. The tools themselves are

volatile. Creators who test platforms

describe the experience as incredible

but unstable. You to build prototypes

very fast, but they have two very big

problems. Number one is that they are

just not as powerful as the tools that

developers use to code with AI. And

number two is that they are very

expensive. And this is what this video

is about.

>> Chat GPT can do are impressive. However,

a lot of the code that AI generates, it

just sucks and it's outdated. And we

know that things in the industry are

constantly changing just overnight. And

these AI models, they need time to learn

new libraries and new syntax. And

sometimes it's just flatout wrong. And I

don't mean like just giving you

inefficient code. I mean like the sky is

yellow wrong. They could build a

production level app in hours, but a

small tweak could break everything.

These AI coding systems could tell a

user that it's fixed a problem when

asked, but in reality, the code wasn't

even checked. Developers found

themselves debugging AI's mistakes

instead of their own. It's like a whole

new skill set is needed to work with AI,

as well as an understanding of when to

use it and when not to use it. A

developer friend of mine who's had a lot

of experience with Vibe coding tells me

that the current state of AI coding

systems without any human input can be

overly verbose with unnecessary bits of

code and it can mix up different coding

paradigms in a single project. And aside

from all of this, there's the issue of

accuracy. Vibe coding tools often

hallucinate. They'll invent APIs, create

phantom endpoints, or generate functions

that don't even exist. There are some

workarounds for experienced coders, and

these problems may be fine for a toy

project, but in a production

environment, it's a nightmare. We

already saw this at the introduction of

this video. One team discovered that the

AI generated multiplayer game used

Python's pickle module for networking,

which effectively opened the door to

remote code execution attacks. It was a

working app until someone realized that

anyone could run the code on anyone

else's machine. It's like building a

house overnight and finding out later

that you forgot the foundation. But the

criticisms go deeper. Security experts

warn that Vibe coding encourages copy

and paste culture where developers don't

understand what's running on their

servers. Educators say that beginners

risk skipping the fundamentals entirely.

Another app called T made headlines

earlier this year as 1.1 million

personal messages and 72,000 images were

leaked without any hacking required

because it was all unencrypted. The poor

security was due to the app largely

being built with vibe coding. Even

within the AI community, some engineers

quietly admit that prompt engineering is

a band-aid, not a discipline. But even

the skeptics do acknowledge the power.

As one neurodeiverse developer wrote

after building and shipping an AI

generated app, quote, "Vibe coding gave

me dopamine highs, but it can't replace

human oversight." End quote. And that's

the paradox. Vibe coding is both

miraculous and maddening. A tool that

can give you superpowers and headaches

at the same time. It's like handing

everyone a Ferrari or Formula 1 car

without teaching them how to drive.

However, there are some vibe coders that

do insist that top level developers will

smartly use AI coding strictly as a

tool. They'll do their due diligence,

check what the AI code does carefully,

and reap the rewards of increased

output. Meanwhile, those with little

coding who stumble into Vibe coding will

just produce slop and have massive

issues. So, where does that leave us?

In some ways, the absolute flood into

Vibe coding could be putting the cart

before the horse. We're moving a bit too

fast with our promises versus the

reality of the technology in 2025.

Like a lot of missionritical AI systems

these days, it works most of the time,

but it has serious limitations, comes

with unique risks, and requires expert

oversight.

If you're skilled in coding and know

what to look for to fix any issues that

can arise, it can make your life easier

in small ways. But conversely, if you

have no idea how to read or write code,

yet expect Vibe Coding to do everything

for you with a few prompts with no

issues, we're not there yet, especially

for more complex tasks. For simpler apps

and web pages like a landing page or

store apps, vibe coding software like

Lovable could work well. But for

anything more complex, vibe coding still

has issues for those who don't know how

to code. So to summarize, I'm not saying

that vibe coding is all bad. It's here

to stay and for those who know what

they're doing in specific use cases, it

can be very helpful. Success depends on

how the technology is used. Hey guys,

thanks so much for watching the whole

way through this episode. It really does

mean a lot. and so does your support

over all the years. It's really amazing.

Otherwise, um that's about it from me.

If you want to see something technology

related, I've got another video that

I'll leave right here. And uh yeah,

that's it. My name's GoGo. I've been

Cold Fusion, and I'll catch you again

soon for the next episode. Cheers, guys.

[Music]

Cold Fusion. It's new thinking.