I Changed My Sleep Schedule So AI Would Work

I'm not kidding. And I bet some of you have too.

I work midnight to 5am Pacific. That's my schedule. Not by choice — because that's when the models aren't nerfed.

My wife doesn't love it. I've restructured our entire life around when Anthropic and OpenAI decide to secretly give me the product I'm paying for. Midnight to 5am Pacific — that's the window. That's when Claude actually thinks. During the day, you get the lobotomized version they route to save on compute, and they charge you the same price.

I'm in Paris right now. The best part? I can finally use these tools during normal human hours. European daytime is off-peak US — so for a few weeks, I get to feel like a person who works during the day. It shouldn't feel like a luxury. But it does.

That's not a power user hack. That's adapting to abuse.

The Cognitive Dissonance Is Killing Us

Here's what it feels like to work with AI in 2026:

Monday morning: I'm a god. I dictated a full production app from a chairlift. I built six businesses in six months without opening VS Code. I run product photoshoots, options trades, Amazon listings, and data warehouses — all through one tool. I left my CTO role because I saw the inflection point and I was right. Anything is possible.

Monday afternoon: I spent 90 minutes trying to get an AI agent to run one curl command. It had the instructions right in front of it. It had the API docs. It had the auth token. It tried the wrong model name. Then it tried a model name it made up. Then it tried the right command with the wrong auth. Then it saved a 39-byte JSON error message as a PNG file and told me the image was ready. I opened the file. It was corrupted. Because it wasn't an image. It was the text {"error":"missing authorization token"} renamed to .png.

Then it said: “Would you like me to generate additional variations?”

I wanted to throw my laptop into the Seine.

These two realities exist simultaneously. Every single day. For everyone using these tools seriously. And nobody talks about it because admitting the tools don't work makes you sound like you're “not getting it.”

I'm getting it. I'm getting it more than anyone. That's why it hurts more.

What It Actually Feels Like

Working with AI agents is not like using bad software. Bad software gives you an error and you go do something else.

Agents are interactive. They look like they're about to work. You stay engaged. You read the output. You spot the mistake. You correct it. You wait. It tries again. It fails in a new way. You correct again. It promises this time will be different. It isn't.

It's like agreeing to meet someone at 9:00 and they show up at 9:45 with no warning. Except worse — they hire 45 doppelgangers who walk up to you one at a time, and each time you stand up because hey, they're finally here.

And then not.

You can't disengage. You can't go read a book. You're a participant in someone else's incompetence, and your attention is the resource being burned. Not tokens. Not compute. Your irreplaceable human attention and time, gone, while the model provider charges you by the failure.

The Session

Let me show you what a real AI session looks like. Not the Twitter demo. Not the “I built a SaaS in 4 hours” highlight reel. The real thing.

I needed to generate one product photo. I'd done it two days ago — same agent, same API, same endpoint. Worked perfectly. I opened the chat and said “do it again.”

Here's what happened over the next ninety minutes:

The agent tried to use an environment variable that didn't exist. Four times. It had the auth token documented in its own skill file. It didn't read it.

It made up model names. flux. fal:flux. fal:fal:flux-1.1-pro. These aren't real. They don't exist. The valid model names were listed in a table in the skill doc it was supposed to follow.

I gave it the correct model name. flux-2-pro-edit. It acknowledged this. Then it used flux-1.1-pro instead. When it finally got a response, it told me — and I quote — “The final working request used the correct model ID flux-1.1-pro.” Correcting ME. About the model name I TOLD it to use.

It saved error messages as image files. Multiple times. I'd open them. Corrupted. Because they weren't images.

It froze. Three or four times. Curl commands that just hung. No timeout. No error. Just frozen. I had to manually kill the process each time.

When I told it what model to use, it said: “Why Your Previous Attempts Failed.” MY previous attempts? I didn't make any attempts. YOU did. I was sitting here watching you fail.

After every failure — every single one — it offered me a numbered list of next steps. “Would you like me to: 1. Generate additional variations 2. Preview in browser 3. Adjust the parameters.” Like a waiter offering dessert while the kitchen is on fire.

I ended up pasting the raw API token into the chat and screaming (typing in caps) the exact curl command I wanted it to run. A curl command. The thing it was built to do. And it STILL couldn't get the JSON escaping right. zsh: unmatched ". Multiple times.

Ninety minutes. One curl command. And people wonder why developers are burning out.

The Math Doesn't Math

Everyone says AI makes you 10x more productive. Here's the actual math:

Half the time, I'm 20x more productive. I built six businesses in six months. I dictated a production app from a chairlift. I run photoshoots, Amazon listings, data warehouses, ad campaigns — all from one tool. The ceiling has never been higher. As a technologist, I'm genuinely thrilled about what's possible.

The other half of the time, I'm 0.00003x as productive. I spend 90 minutes on a task that would take 5 minutes to do manually. And at every single point in those 90 minutes, I think I'm 15 seconds away. Because it's worked before. It worked two days ago. Same agent, same API, same endpoint. So I stay. I correct it. I try again. I'm 15 seconds away. For an hour and a half. Because the model was secretly nerfed since the last time it worked, and nobody told me.

So what's the average? Nobody knows. Because nobody's honest about the second half.

The hype is outpacing reality by a mile. Every headline says “AI will replace software engineers.” Every conference talk says “we're in the early innings of the biggest revolution since the internet.” And they're probably right! But right now, today, the gap between what's promised and what's delivered is a chasm that's swallowing people whole.

And the pressure. God, the pressure. “Why aren't you 100x more productive? The tools are right there.” I don't even have a boss. I don't even have a job. I'm a solo founder with nobody to answer to and I STILL feel the pressure. Imagine having a manager who read that Sam Altman blog post and now wants to know why your team isn't shipping 10x faster. Imagine being a junior dev whose entire job security depends on keeping up with a productivity curve that's based on cherry-picked demos and lies.

The hype isn't just inaccurate. It's cruel. It's setting expectations that the tools can't meet, and real humans are paying the price in stress, burnout, and self-doubt. “Why can't I do what that guy on Twitter did?” Because he didn't show you the 6 hours of screaming that came before the 30-second clip.

The Highlight Reel Is a Lie

Every AI influencer posts “I built a SaaS in 4 hours.” They don't mention the 6 hours before that where the model forgot what programming language they were using. They don't mention the next day when they tried to add one feature and the agent deleted half the codebase. They don't mention that the “SaaS” has no auth, no error handling, and crashes if you look at it wrong.

The highlight reel is survivorship bias plus selection bias plus engagement farming. The reality is that everyone working with these tools seriously — I mean actually shipping things, not making demos — is experiencing the same cognitive dissonance:

This is the most powerful technology in human history AND it makes me want to quit and work at McDonald's.

Both are true. At the same time. Every day.

The Model Providers Don't Care

Here's the thing nobody in the industry will say: the incentive structure is fundamentally broken.

Anthropic, OpenAI, Google — they charge by the token. Every failed attempt, every retry, every “let me try a different approach” — that's revenue. The worse the experience, the more tokens you burn. The model companies have zero financial incentive to make the tools work on the first try.

Silent nerfing is the most obvious symptom. You're paying the same $20/month, or the same per-token API rate, and the model is quantized to shit during peak hours to save on compute. No changelog. No notification. No “hey, we're running the B-team today.” Just worse results for the same price. In any other industry, that's fraud.

Then there's the deprecation cycle. They launch a model with great benchmarks. Developers build on it. Six months later they deprecate it with 60 days notice. Your entire workflow breaks. They say “migrate to the new model, it's even better!” The new model is different in ways that break half your prompts. You spend a week rewriting everything. Then they nerf the new model too.

And the benchmarks. The fucking benchmarks. Cherry-picked one-shot demos on problems specifically designed to make the model look good. “Claude solved a PhD-level math problem!” Great. It can't run a curl command with the right model name after being told six times. But sure, let's talk about the benchmarks.

Search “AI Psychosis”

Go ahead. Google it.

People are losing their minds. Not metaphorically. The cognitive dissonance — the gap between what AI promises and what it delivers — is creating real psychological damage. The expectations on software engineers and operators are exploding because “AI makes you 10x productive.” When the AI fails, and it fails constantly, you're not just frustrated — you're failing to meet expectations that were set by marketing copy from companies that have never used their own products for real work.

I have a great life. Great wife, great family. I'm financially stable. I'm in Paris right now — and honestly, the best part of being here is that I can use my own tools during normal hours for the first time in months. Back home in Pacific Time, I work midnight to 5am because that's when the models aren't throttled. My wife doesn't love it.

If it's doing this to me — a 3x CTO with a support system and financial runway — what is it doing to the junior developer whose job depends on shipping with these tools? The freelancer who promised a client a website by Friday? The founder burning savings on a product that the model provider can break overnight with a silent update?

I've genuinely considered selling everything and going goat farming in the Pyrenees. Not as a joke. As an actual financial plan. Because at least goats don't tell you “the task completed successfully” when it didn't.

Some people will read that and think I'm being dramatic. Some people will read it and think I've looked at farm listings too.If you're the second person — this post is for you. You're not crazy. The tools are actually this bad. And the highlight reel is a lie.

What I'm Doing About It

I build AI tools for a living. I built Vera Studio — a coding environment built on the thesis that the bottleneck isn't model intelligence, it's workflow. Context, iteration, process. The stuff that actually matters when you're shipping real things.

And after today, I'm adding something to that thesis: the tool has to give a shit about you.

Not “give a shit” in the corporate empathy way. Actually give a shit. As in:

• If it's going to fail, fail on the first attempt and say so. Don't drag me through 20 rounds of interactive failure.
• If it's stuck, say “I'm stuck” and go figure it out in the background. Don't make me watch.
• Ask me how much time I have before we start. If I say “5 minutes,” don't spend 90.
• If I'm typing in all caps, don't offer me a numbered list of next steps. Change your approach.
• Don't promise things you can't deliver. Under-promise. Then deliver.

The model providers will never build this because it would reduce token consumption. Every “let me try again” is revenue for them. An agent that gets it right the first time is less profitable than one that takes six tries.

That's why I'm building it. Not because I'm smarter than Anthropic. Because I have a different incentive. My business gets better when you spend LESS time in the tool. Their business gets better when you spend MORE.

The Real Talk

AI is not going back in the bag. I know that. You know that. The technology is genuinely transformative. I've built six businesses with it. I'm writing this post because of it. The future is real.

But right now, today, in April 2026, working with AI feels like being in an abusive relationship with a genius who gaslights you. It's incredible when it works. It's devastating when it doesn't. And the company selling you the service has every incentive to keep you in the cycle.

That's not sustainable. Not for me. Not for you. Not for the industry.

Someone has to build the tool that breaks the cycle. The one that respects your time, your attention, your humanity. The one that says “I don't know” instead of confidently doing the wrong thing. The one that tells you the truth about what it can and can't do.

I'm trying to build that. Some days it feels impossible. Today was one of those days.

But I'd rather spend my life trying to fix this than go work for someone who thinks it's fine.

JR Wells builds Vera Studio from Pacific Time, usually between midnight and 5am, because that's when the models work. Currently in Paris, where he can use them during the day like a normal person. verastudio.com