It’s the summer of 2026. Is AI really cheaper than humans or is it an outright lie? People said AI models be improving, becoming ten times more efficient and cheaper, but what is the reality? What can we really expect in 2027, for example?
If my experience with it at work is anything to go by, no. Increasing amounts of my time are spent fixing the sloppy output of coworkers who use AI. Far from being “left behind” I’ve become one of the few people who is actually able to solve complicated problems.
There was a great example of the issue with AI when compared to skilled workers.
There was a company that needed a special function in their software, they wrote it in four weeks and it worked fine.
Then they hired an AI pilled CTO, and needed to write a very similar function to the previous one, only this time, the CTO demanded to save time with AI.
The second function took months of testing to be production ready, and was still far more buggy than the first one, even a year or two later they still had issues with it.
I work at a company with AI-pilled management, they push hard for copilot, and I have noticed that the admin center for Copilot in Microsoft 365 has an overview screen where it has stats, and one of the stats is “hours saved with copilot”.
I would LOVE to see how those stats are calculated, they are probably extremely vauge and highly inflated.
The hours saved are likely calculated with Copilot. Just more Microslop.
I wouldn’t be surprised if they just make the assumption that any time spent in Copilot means that the user saved at least that amount of time, if not more.
This exactly. We have one or two people, who are supposed to be more technician level people who have started submitting thousands and thousands of lines of code for “features” they dreamed up and which aren’t on any backlog or roadmap, and the actual software team is just like “fuck off we don’t have time for this.” They aren’t actually contributing, they are just causing drama at this point.
speaking of which, it seems to me that over time one person will check the work of 3 to 5+ AI agents, when most people are fired, do you think it will be damn hard for the remaining people to work?
I don’t think my colleagues are going to be replaced by ai. I think they are going to continue to use llms to generate “output” that needs someone like me to constantly fix until that becomes too expensive. At which point they will go back to doing their work the way they used to.
These llms are impressive word guessing machines, but the are nowhere near as capable as their companies say they are.
The price of using these LLMs is already outpacing what it costs to hire a developer.
And their output is generally trash.
It’s quite good for any problem that nobody cares about. For example, if you have a boss who wants hourly status-updates but does not actually read the updates. Or if you need to fill out security forms, but you were going to lie on all the forms and trust that nobody reads them. Or if you get a bonus for writing >100k lines of code, but the code doesn’t need to do anything that people want. Or if you need to have someone answer customer-support questions, but it doesn’t matter if the customers get helped or are happy.
TL;DR it’s excellent at making things that look like what they’re supposed to be at first glance. If anyone looks more closely, good luck.
Best description ever.
The thing about knowledge work is it isn’t about knowing things, it’s about knowing how to put things together. AI lets me accomplish certain tasks faster than I could alone. I can sit in a meeting and spend 30 seconds asking AI to run an analysis and then I can focus back on the meeting. After the meeting, the AI response might be right, but it’s probably not. But it has gathered up all the raw data and now I can add calculations to a spreadsheet, or look at the data in one place rather than having to query 8 different systems.
AI is good at rote tasks and grunt work. It’ll write you a bit of code slicker than shit, but you have to very specifically tell it exactly what to do and how to do it. That’s knowledge work. Actually following the instructions without fucking up is the mundane part. And AI can do that really fast, and if your instructions are very clear, fairly complete, and you understand the likely failure modes that will trip it up, you can really increase your productivity.
But “vibing” anything with AI is bullshit and doomed to failure. If you don’t know exactly what you want and exactly how you want it built, you are going to get some rancid garbage. Knowledge work isn’t in any danger.
What I do fear is where the next generation of knowledge workers will come from when AI is faster and easier than building up juniors to seniors. I believe they will continue to find their way, but it might be harder and there might be fewer.
Fuck no, it is a lie. It is like an actor pretending to be a knowledge worker which fools people who don’t know better or who want to believe the lie.
Remember, it just regurgitates what it is fed plus some randomization and it can’t come up with novel ideas based on experience. That is what knowledge workers are for, applying knowledge to novel situations.
I like to call AI the next gen search engine, but be careful of some things:
-
you’re much easier to track and profile
-
you may be fed false information easier without context of source information
-
you’re killing the planet
-
hallucinations
That makes it something other than a search engine, not a new version of one.
I disagree.
I can go into chatgpt and search for landscaping companies near me. It will link them.
I can search for a recipe.
I can search for anything I would normally use a search engine for with much faster speeds.
If that’s not searching, I don’t know what you mean.
Google and Bing return the results instantly for me.
Is chatgpt twice as instant?
Ya with their AI things. For example, of you dont HSE their AI things, searching for a recipe yields you results of said recipe. Click on it. Then you have to scroll though a bunch of bullshit about their family recipe and how it tastes, then they should you the ingredients. Then the sizing. Then the instructions.
Do the same on an AI assistant and its just text of the recipe and steps no bullshit.
One of these is improved over the other.
Recipes are something I rarely take the first result for, since they can vary widely for the same dish.
Do you care what website it came from?
Have you ever checked to see if it actually matches the source page recipe?
Recipes are something that I always check a couple sites for to look for consistency or variation and some cooking sites are better than others.
Yes I care what the source site is, that’s why you can check sources in chatgpt. Anyways its only one example, and I’m not defending AI here, my argument is that they are basically improved search engines. What is a search engine, you might need to ask yourself. You query, something responds.
-
what knowledge workers are for, applying knowledge to novel situations.
how often does that really happen in your daily life, and how much money do all people in society make with it?
i get a feeling that this mostly applies to scientists figuring out new stuff. which makes about 1 in 10000 people in society. the amount of money earned through wages that way is rather little. most knowledge workers really have an office job where they fill paperwork, send it to someone else, do calls, etc.
Every single working day where a suggestion or decision is made.
Sometimes it is a big deal and obvious but sometimes it is just keeping people doing the stuff that won’t break things that are already working.
LLMs cannot do anything new.
AI sucks ass for every serious task. I’ve used it dozens of times to work on research projects, and consistently it gives me responses that sound good, but have holes, like citing works that don’t exist, mis-applying arguments, etc.
these problems never appear if you ask it standard questions, e.g. simple exercise questions from the classes. it always solves those correctly, probably because they exist in the training data. but try to ask it anything where logical inference is non-trivial and it starts splicing facts together in ways that don’t end up straight.
anyways, all being said, it’s still great for routine tasks, and for informing you about basic knowledge in a new field. since basic knowledge is all written down in books somewhere, AI knows about it quite well. also, coworkers make mistakes, and lots of them. if i sum up all the steps of progress that coworkers make and that AI make while working on a new project, i’d say it’s about on the same level. coworkers tend to think about new topics more seriously, but also they often just don’t respond, give up, never call back etc. meanwhile AI tries to output something, even if it’s wrong.
all that being said, there will still be a decline in knowledge worker jobs, but not so much because of AI being excellent at actively exploring new areas and kinda spreading out to take over jobs, but instead because like 99.999% of our jobs are kinda routine jobs anyways. people always tell themselves they’re special, and they are. but at the same time it’s mostly routine jobs. these are not mutually exclusive. and that’s why a lot of jobs are still going to vanish.
The fact that companies are now telling their employees to scale back and be mindful of their token usage shows that it’s not cheaper. It can be efficient yes, usually for automating the mundane stuff, but it’s not cheaper.
Cheaper, prolly
Efficient? Debatable.
Reliable? Nowhere close, and humans aren’t even all that reliable.
That doesn’t stop the parasite class from viewing AI as a hammer, and every task as a nail… or spending trillions failing upwards.
As a programmer: HELL NO!
Nope.
“Do the thing. I have a script that does the thing. I made it by listing the things in excel and asking Claude to write a script that does all the things”
The script not only doesn’t run, but it also quietly doesn’t do parts, and breaks stuff too.
So now I’ve got a pile of code that doesn’t work and it would be faster if I’d written it myself than troubleshooting the spaghetti. It’s like having a dumb intern, except the intern is incapable of improvement.
Considering they’re raising the prices of it significantly, I’m gonna say no it’s not cheaper.
Right now, the state of the art appears to be that LLM’s are far more expensive to run than humans for many tasks. So, yeah, it is a lie.
There are uses for specialty AI that seem to work well, but they are usually bespoke for a certain task.
I expect that, if AI is used in 2027, it is going to be used with a lot more intent in targeted uses. I also expect some companies are going to realize it is better to fully control their own AI on their own hardware than to use a more state of the art AI which will use all their data as training data.
There seems to be a bit of an odd relation between its value and cost. Building a model and setting up a data center is horrifically expensive, so LLMs as a service have to be just as expensive. But the stuff that can be done by LLMs is low stakes, low thought work, like copy writing, chatbot customer service answering the same 30 FAQs but crashing out on anything else, summarizing this morning’s headlines, etc. The price on these things is still being hashed out but it’s looking a lot like using AI is like hiring Anthropic for $100,000/yr. to replace a person who only gets paid $45,000/yr.
In 2027 we can probably expect vastly larger bills or reduced usage limits. The price at which some models are being offered right now is in no way sustainable. For example Anthropics subscription plan gives you tons of quota for the most expensive models, while costing relatively little. Github Copilot recently changed its billing to token-based (before it was request-based) and it already produces incredible bills within the company I work for.
