Microsoft accidentally told the truth about AI

There's a dude on YouTube making a video about how miserable the job market is for software developers right now.

And I catch his video because in the thumbnail, his left eye is super red. And you know, humans, we catch on to these things quickly.

So I click the video, and it is the most unhinged beginning to a YouTube video that I have ever seen.

He spends the first five seconds just staring at the camera deadpan while taking a sip of his tea without breaking eye contact.

Then he suddenly leans in, points at his eye, and he's like, yeah, this. This is from AI. This is literally from the stress of job hunting.

He's been laid off for nine months now and has been applying nonstop to jobs, and in some cases, going really far into the interview process with a couple of companies.

One company in particular he went nine rounds with, and he's 100% sure this job is his. He doesn't get the job.

And his heart drops. He couldn't believe it. He asked him, he's like, why? What happened? And they share with him the winning job entry.

The winning guy used 100% AI. It has 30 more features than the poor bastard Garrett, who used 50% AI and 50% manual.

Garrett's basically the guy who, in 1849, heard about gold in California and went, you know what? I'll walk to Nevada and see how I feel.

I mean, the winning dude wasn't even present when his application was submitted. For all we know, he was taking a shit while Claude was flummoxilating.

And. And that dude now has Equity and a 401K. And meanwhile, Garrett is at home writing unit tests and making his code dry and shit, and his eyes filling with blood.

One guy is shitting, the other guy's hemorrhaging. Guess which one the market chose. And the jobs exist, by the way.

Software engineering postings are up 11%. But the market has spoken. And the market said, we want more stuff and we want it faster and we want it cheaper.

And you can sit there and say, but what about quality? What about maintainability? And the market goes, did I stutter, bitch?

The thing that got you hired in 2019. 2019 is the thing getting you fired in 2026. Nobody reads your own code anymore, man, let alone your code.

It's a dead skill. It's like reading hieroglyphics. It's time for you to adapt and face the slop. Don't argue.

Imagine more slop. Do the slop. Be the slop. Slop is peace. Slop is strength. You will Eat ze bugs and you will make Z slop.

Learn to enjoy it. For over two years now, AI Tools were the greatest sale in the history of software.

20 bucks a month and you get a half brilliant, half retarded senior engineer in a box. Your CEO almost had a heart attack, it was such a good deal, they called it unlimited.

And back then, unlimited used to mean something unlimited meant. And I looked this up without limit. Now unlimited means until we find out how much you actually want it and then we'll tell you the real number.

It was basically a survey. And in 2026, the results are finally in. And the results are you'll pay anything.

And now the price is anything. Microsoft just leaked that copilot is moving to token based billing because its weekly cost of running the thing nearly doubled since January.

What's wild is that the bad answer costs as much as the good answer. And the tokens you're paying for include the tokens the model wastes on itself because the model thinks out loud.

It like argues with itself. It writes a draft, then second guesses. So it starts over. Then it goes, actually, let me reconsider.

And every one of those anxious utterances is on your tab. And you are paying for this thing to have a nervous breakdown in front of you.

And you're billed to the millionth of a syllable. And at the end of the breakdown, it produces some answer.

And the answer is wrong 25% of the time, according to this new Microsoft paper. A new paper dropped on Arxiv last week with a title where you could tell.

They try to make it sound less bad than it is, but they just couldn't figure out how to pull it off. It's called LLMs.

Corrupt your documents when you delegate. The authors work at Microsoft Research, the division of Microsoft, where for obscure historical reasons, the employees are allowed to say things that are true.

And across 52 domains, the frontier models corrupt 25% of document content by the end of a long workflow.

The paper describes the errors as sparse but severe, which, you know, they're trying to make it sound like, charming, like it's glitter or something.

Sparse but severe is how you describe a sniper, bro. And they gave the model the ability to use tools, and they found giving the bot tools made it 6% worse.

This is the first time in documented history that we've managed to create a tool that becomes less useful when you give it more capability.

One of the experiments was essentially edit a document, then undo it. And the best AIs couldn't do it without insane Corruption.

We had the Control Z button, man. It was free. A dad walks in on his nine year old daughter on her laptop and she immediately slams the laptop shut and starts crying.

She's like, dad, it's not what you think, I swear. But he knew exactly what she was doing. She was using Google AI.

And he wasn't specifically pissed that she was using Gemini, although that is pretty upsetting. He was worried for her future and the way she was using it was honestly so adorable.

She's asking the thing questions on how to get along with her sisters better and how to swim faster after a swim meet and for creating fan fiction plot lines for her favorite book series.

And her dad freaks out and posts this on Reddit like he just discovered his daughter selling feet pics on the Internet.

He's like, she's not in trouble. But we had a long conversation and she's devastated. She now fully understands how sycophantic and insidious it is, which firstly, like, good job, you've managed to teach your kid 24 syllable words that are insults directed at a bot.

But I mean, the poor kid, her spelling test next week is gonna be insane. Her teacher's gonna go, use sycophantic in a sentence.

And she's gonna go, the large language model was sycophantic. And the teacher's gonna go, huh? And then she's gonna go, it is also insidious.

And the teacher's just gonna walk out, man. She's like, I'm done. I'm so done with this generation. But.

But I feel for the dad, Honestly, I do. He has to make sense of a technology for his kid that he himself doesn't even understand.

There's no precedent for an omniscient. Always on ghost in your kid's pocket. Like, what do we want for our kids?

You want her to be good at something, right? You want her to earn her skills. But the problem is that her friend's parents let her use AI and her friend at like 9 is shipping AAA games to Steam by 13, your daughter is making basic 3D printed fidget spinners, and her best friend is shipping Enterprise Sass.

The SAS has 3% uptime, which sounds bad, but honest, honestly, anthropic would kill for that number. And I'm so conflicted because if my 4 year old draws a house with a little square and a triangle on top with a little sun in the corner, I would literally cry.

Whereas if she showed me an oil on canvas that she generated that is objectively, stunningly beautiful, and may even be something that represents her.

I'm still ultimately unimpressed and maybe even bothered. And the question is why? Why does the ugly thing make me cry while the pretty thing makes me cringe?

And I've been thinking about this a lot and really this brings to mind the fact that while AI has whopping width, like a sheer breadthness that is impossible for humans to even fathom, humans have depth.

That is impossible for an AI to fathom. The reason I cry at my daughter's sloppy drawing is because of the infinite depth it possesses here.

Represented in this drawing is a manifestation of love. It is a process of nourishment and nurturing that has been four years in the making.

From not knowing how to say dada to crawling, to laughing, walking, eating with a spoon, and now drawing a picture of the house in which she lives with the family that loves her more than anything.

That crayon painting is more Mona Lisa to me than the actual Mona Lisa. So when my daughter prompts for an art piece, I'm not moved.

Because there's no story. And without a story, humans are just aimless floating amalgamations of molecules and cells.

Art is precisely depth. That is precisely what it is. It is precisely story. To me, it seems like the usage of AI has no benefit other than for commercial purposes and even then has limited usage.

If you're creating a hobby project, you're either creating it for the enjoyment the process brings you, or you're creating it to learn about the process itself and to climb the ladder of skill.

Using AI robs you of both opportunities. It is neither enjoyable nor educational. And so I am of the mind that the dad in this situation situation, while it is a tough call that is impossible to model out, he did the right thing.

He gave his daughter the opportunity to struggle in a time where people are gambling with their child's upbringings and values and livelihood on the oversold promise of AI that is exclusively being propagated by OpenAI and Anthropic.

Here is a man who chose to raise his daughter with his own values and beliefs and not those of Sam Altman and Dario Amodei.

If he turns out to be wrong, well, guess what? Opus 20, by the time she's in college, will surely be a thing that requires no mind at all.

And that's already arguably the case today. And by then she's gained a voice, a style, perspective, suffering, struggling, experience, depth, longing, understanding, and ultimately a story.

A story about what life, beauty and its lack is. Whereas the kid prompting all day learned how to delegate.

End of list. Nothing else was learned. As with all things, the truth is likely somewhere in the middle.

But I'm very curious what you think. Thanks for watching.