Agentic is such a shitty fucking word to describe how this works that you can tell it was dreamed up by fucking business and marketing idiots.
All of the new “agentic” models talk to themselves and consider it “thinking”. It’s completely ridiculous. Of course they eat up a lot of tokens, they’re fucking chattering with themselves.
There’s no way to quantify the amount of “tokens” any given task will take because the “thinking” chatter is verbose and unpredictable as shit. You can send it in a question like “What is happening?” and it burns through a million tokens trying to describe why the latest bash command it tried to run is taking fucking forever because it didn’t write the command correctly and it is permanently paused waiting for stdin.
I see some young dude shilling this on tiktok where he had this claude war room setup that kind of looked like an idle miner game with them all working for him. He supposedly said they were making him more money then they cost him. But then he showed their “work” and it was all them just selling AI art on Etzy, which feels not sustainable and also like such a lazy scam. All I am thinking is how much energy and water it’s taking for someone to buy some disappointing AI art from someone they thought was a real person.
The business is not selling AI art on Etsy, it’s selling courses on how to sell AI art on Etsy.
Honestly I wouldn’t be surprised if he was paid directly or indirectly by Anthropic for his marketing. It’s not like he was selling knowledge, he was just telling everyone to try it.
I was looking for wall art on Etsy a few months ago but just gave up, the majority of listings are AI slop now.
Sounds like it’s rife for exploitation. If the AI company’s revenue needed a boost this quarter, just send out an update that does a few extra loops of talking to itself.
Did they literally learn nothing from “The Mythical Man Month?” Geesh.
No, management as an overall group rarely learns anything from failed concepts. Hope springs eternal, so whenever something new promises a productivity revolution, there are always clueless managers who think it will get them into yacht world. They’re treating AI like a bag of Magic Beans, while legions of equally clueless spectators demonize the tech for the stupid decisions those managers are making. AI is a very promising technology that’s simply not ready for how it’s being used.
I miss times where you copy pasted whatever you had to google and correct stack overflow link was first result so you copy pasted first answer back to your IDE and everything magically worked. Now it’s tokenmaxxing, ai guide spec files, skill files, agentic development, copilot code reviews and trainings because you’re doing things wrong, you need to be context engineer, vibe code engineer, prompt engineer to use AI efficiently but event if you do all of it you’re still doing things wrong.
A vibe Cuck. Please daddy, please implement this as a function.
it’s now apparent that using AI is more expensive than hiring people, especially since it offers only limited productivity gains at the moment.
And there it is: the wakeup call. The sound of a bubble at least deflating a bit (probably more to come). But we all know it will take months or years for that realization to really sink in. Corporate leadership will mess around trying to cut costs while denying they made a bad call by falling for the hype.
Also, this isn’t just hitting tech giants by the way. This is hitting everyone who jumped on the AI bandwagon. What were going to see is a frantic scramble in two directions:
-
hiring a subset of people back because the C-level now realizes they are under-staffed for purpose - emphasis on subset, because they will happily keep overloading those remaining as much as possible.
-
pivoting to open source in-house hosted models to cut costs, and initially struggling with all the janky tooling that will come with it. But even once the tooling has been improved to be somewhat usable, there will still be some initial frustration with these models. For all the problems with the bleeding edge ChatGPT and Claude models, in most cases they are more useable than the self-hosted ones, which lag by a couple of generations.
What I’m hoping to see is the push to rely on AI drop off sharply and a resurgence for online community collaboration and resources. They weren’t perfect, but at least they keep one’s critical thinking and research skills a little sharper.
I don’t know if a lot of tech needs to hire back. Outside of AI, tech adoption is on the right side of the adoption S-curve and most development time right now is going to revenue generation.
And if you do need to hire back, you don’t need to hire from HCOL areas. Make the position full remote and price the salary to a LCOL area. You don’t need to pay a Bay Area wage any more.
lol you think those fucking idiots will go back to remote work? they’ve spent every year since covid trying to force everyone back into the office.
They will for new hires when they can pay them 50%.
I don’t know if a lot of tech needs to hire back.
I think many CEOs will see it that way. Better to underhire and ride their slop code base into further customer loss, data breaches, lawsuits and bankruptcy, than put up a weak quarter while something can still be done to save the business.
Has any data breach ever significantly hurt a company financially?
Hard to say, for sure, in each case. There’s often plenty of factors that could have been key. Fuck-ups tend to fuck up in a variety of interesting ways.
An obvious seeming one is Roomba’s infamous leaked person-on-toilet picture. I didn’t stop buying Roomba because I doubted their engineering. It was their (lack of a) Privacy team.
-
People who repeatedly make horrible decisions, with no sign of learning, should not be allowed to continue making decisions. All of these frothy pro-ai people need to go. Let them spend the rest of their days picking up litter or something useful.
Rich people make the decisions though, and they earned the right to decide how the world economy runs because their ancestors worked hard (owned slaves)!
I don’t know any of the surrounding context but I think about this poem (?) a lot
"Get off this estate." "What for?" "Because it's mine." "Where did you get it?" "From my father." "Where did he get it?" "From his father." "And where did he get it?" "He fought for it." "Well, I'll fight you for it.Carl Sandburg, Selected Poems
Agentic really moved the needle on this. I mean, if your LLM is suffering from a limit, why not let another LLM watch it. While that’s going on, why not let yet another orchestrating LLM watch all the LLMs and tell them what to do. It starts to get confusing fast so better use yaLLM (yet another LLM) to watch the whole circus.
I know some people at a company where they pushed AI hard. Told employees to use it as much as possible, we’ll pay for the tokens, and people were running out of tokens in the first week and the company didnt want to pay for more so they had to wait for counters to reset.
Good, fucking crash and burn bastards!
tokenmaxxing
I’m doing my part. 👍
At most, Ill prompt a LLM single chat now sub agents no running around. Justvl a very small focused task, I hate the agentic stuff. Its always crap.
The autocomplete is ok, it’s nice for highly repetitive stuff.
But in my view if your doing something highly repetitive you probably shouldn’t be doing it by hand lol. So most the patterns I use avoid that even
Ive said it before and Ill say it again. These are tools that require skill and training. If you throw an untrained intern at it without prep they’ll eat through tokens like crazy, and you wont get much gain.
Theres lots of ways to reduce tokens used by huge amounts, the difference can be on the scale of 90% or better reductions in usage.
But it takes setup, skill, tools, etc. All which require training and learning for the user to know what to use, how to use it, and when.
I’m so good at asking the bullshit machine to work for me. I can get it to pump out bullshit for 10x less than some idiot.
Remember, telling your model to talk like a caveman will save you about 15% on output tokens!
Indeed, and with tuning, you pass the threshold and it is cheaper than a junior dev, faster, and better quality.
Im still all for hiring juniors, but they should be getting trained to use these tools.
Indeed, and with tuning, you pass the threshold and it is cheaper than a junior dev, faster, and better quality.
Lol. My office chair is cheaper and more useful than a junior dev, too.
Pretending junior devs are worth anything other than as an investment toward later having a loyal senior dev who knows our product is the weirdest part of the whole AI farce.
There’s this weird myth being shared by managers that 3 junior devs in a trenchcoat is equivalent to one senior dev, or something. It’s a little bit like believing that 5 toddlers can do the work of one college student.
It’s like saying “we won’t let anyone under 25 rent our cars, except if there are at least three children in the car, then it is okay by us.”
Developers become expontially more valuable as they gain experience. We invest in them early because some of them cannot be arsed to take a new job every two years, once they are happy and comfortable, and we end up getting an amazing bargain for a few years, once they are senior devs.
Good thing we are going to kill that pipeline to prevent more senior developers thus ensuring we have to use AI. And of course betting on replacing the senior developers as well before they all retire. It is quite the gambit to end employment as we know it.
How magnanimous of you.
Well you see, if you dont hire junior devs, and senior devs eventually retire… then you have no devs at all and thats usually considered bad.
So you hire junior devs… and train them to use the tools, so that they become competent with them… like literally any other job in the world.
Its like trying to argue that because your site workers on a job can use a CAT to move material way faster, now you dont need to hire new guys… which only works up until your CAT operator retires and you never bothered to train their replacement…
Any half competent manager would consider that to be a super stupid thing to have done…
When people tell me AI makes their job faster I assume they’re either shit at their job or their job is useless.
A very very large part of programming is boilerplate code you cant get past, it must be written.
A huge part is integration testing.
You have like hundreds or thousands of tests that are very repetitive. “Open this, click that, do this, type that, wait… assert this happens”
You have all these in a human readable format/doc provided to you by your requirements team
You uabe to convert all of that into code so a program can do the steps, such that it runs every day to confirm “website isnt broken boss!”
Converting all that human text into code is very easy but very monotonous
It can be tens of thousands of lines of code, easily.
This type of task is extremely easy for AI though, especially if you start it off with 10~15 examples to show it how it outta be done
Then you just give it the doc and go “okay draw the rest of the owl please”
This is called “nth shot prompting”, where it has samples to start with and then it just copies your work 1000 times given input.
AI is extremely good at this.
It can turn weeks of work into hours.
The ability to turn human contextual English into code is one of its top skills, and turns out for a serious prod scale app at a company, thats like… 90% of software dev
And since we can work in parallel, I can focus more of my effort on that remaining actually challenging 10% where stuff matters more.
Its like swapping from pulling a wagon by hand of dirt to a job site, to instead pulling it with machine.
I get that part done 10x faster/easier, or better, and now I can actually put my time/energy into way more important shit than just hauling dirt around.
Summarized: when ppl talk about it making them do their job faster/better, its usually cuz it got rid of a bunch of boring parts and now they can focus on the actually important parts of their job more, and thus produce way higher quality results net.
Innovating the boring stuff is how you do your job better. AI doesn’t do that for you.
Okay I get it now. You’re in the shit at your job group. Thanks for clarifying.
It’s honestly so depressing that you’ve taken the time to explain how there is in fact a use case for AI, and what that use case is, in plain language, and the response has been to act like chimpanzees, hurl insults, and downvote you.
Shit like this makes me think Lemmy is doomed.
They aren’t the first industry to do it.
A lot of other engineering fields in the USA are seeing a manpower issue because the fields stopped hiring new engineers for a few decades then the Great Recession reduced the number of engineers they were hiring.
I expect this to happen for programming.
What people need to wrap their heads around is almost all of these cases are/were companies that were gonna lay people off anyways.
They just used “cuz AI” as a PR spin to avoid making investors panic.
“We laid off 25% of our workers cuz we are failing” makes your stocks tank.
“We replaced 25% of our workforce with AI” makes stocks go up.
And they can just lie about it. They probably arent even fuckin using AI… or they are half assing it to make it look like they are using AI to investors.
The reality is, the company is just shitting the bed and upper management is doing everything they can to stop the stocks ftom ranking for just 6 more months so they can sell out first before the company goes bottom up.
Companies actually using AI intelligently arent “replacing” anyone with AI.
But saying
“We laid off 25% of our workers cuz we can”
Usually makes the stock go up.
“I’m trying like crazy to distance myself from the actual ramifications of ai because it’s devastating for my arguments for ai.”
What does training look like? This is new tech and it isn’t like tech has figured out the best way to design and build large projects. Is there even a set of best practices developed for AI yet?
We are way past that, there are several very solid ways to approach it, and lots of debate over which is “better”
What matters is picking one and being consistent company wide and getting all your employees on the same standard.
But yes, the ecosystem/practice now has plenty of powerful tools and methodologies.
And the difference between “Idunnolol” winging it vs using any of the tried and tested architectures is huge.
You go from “I burnt 100k tokens and got a mangled mess” to “I used 8k tokens and got actually half decent results that needed a bit of human tweaking after”
Skill files, MCP protocol, RAG memory… theres a buncha stuff that elevates the tools from “plundering ox in a china shop” to “actually useful”
But you mentioned three technologies instead of methodologies of how to use AI in a workflow.
What tasks make economic sense to devolve to AI? What is the workflow? Who is checking to make sure the tokens are being used economically?
What tasks make economic sense to devolve to AI?
So, the main one I use it for, as its my job, is software development.
I offload about 90% to 95% of my workload to AI, almost all of which is “boilerplate” code that sits in the realm of “very easy to do, but very repetitive and time consuming” which is… most of it. Thats just the reality of software dev, especially web app dev. A lot of our stuff is just plumbing “this api endpoint calls this backend logic which just maps to this basic database operation”
IE, the POST endpoint to
/users/{id}invokes theUpdateUserHandlerwhich takes in anUpdateUserRequestwhich maps to anUPDATE [dbo].[Users] ...sql statement… not exactly super complicated stuff, but you do have to actually write the stuff that defines this.This type of work is trivial for AI, but any given project has its own set of business rules, code rules, syntax rules, formatting rules, etc etc etc.
A naive approach is just yolo throw an AI agent at it, cross your fingers and pray that it randomly chooses to read the right stuff and happen to succeed at following your code quality and methodologies (it wont)
The naive agent also will demolish its way through tokens as it reads way more files than it has to, because every single time you do work its basically starting all over again from ground zero with no context of wtf its doing. This wastes… so many tokens, because its gonna sit and read like 20 files just to figure out “what am I supposed to be doing here?” and this in turn pollutes its context window up so damn full it will start forgetting shit anyways.
This is where actual tools come into play that make this stop being an issue…
-
RAG Memory, instead of blindly searching your codebase, you can tokenize your codebase into a much easier to semantic search system, the agent can WAY faster do a simple search and get returned pointers to “Look over here”, its like creating an index of your codebase itself so the agent has a useable optimized “map” of the project.
-
MCP Tools, which are basically tools that the agent can invoke to do… anything. Normally by default the agent is just given willy nilly access to the terminal and it’ll just fumble through trying to use that to do anything it needs to. This is a great way to fuck stuff up, especially if it also has access to shit it shouldnt (que “our agent accidently deleted our database” type shit). MCP tools allow you to build a curated set of stuff it can invoke, so instead of just doing random shit it has prefab commands to run. If you flesh out your MCP Tools well enough you can outright disable its access to the terminal entirely, because it doesnt even need it anymore. No more accidental database deletions, and it uses waaaay less tokens fumbling around in the terminal.
-
Skills, which are special files that allow it to “lazy load” guides on “how to do x/y/z”, which you can break up into bite sized pieces. So instead of a giant AGENTS.md file that uses up half its context, even though the agent doesnt need 90% of whats in it for a given job, it can instead have a big list of “how to do this, how to do that” and it’ll load a skill for a given task its working on ad hoc, only loading in the instruction relevant to the task at hand. These are huge and critical to further reduce token usage a lot
-
Token reduction skills (caveman being the most popular), theres certain skills people have made that outright change the way the agent behaves, namely the caveman skill. Make agent talk like this. Why more word when less word good. Less word, less tokens. Less tokens, less money. Also faster. Ungabunga. (Caveman mode can give you like 50% to 70% token reductions alone)
-
General prompting skills, know how to prompt agents genuinely makes a big fuckin difference. Baseline thing you learn asap is NEVER correct an agent, this does not work well. Instead you should be going back in the timeline and editing your prior prompt to preemptively correct the mistake before it even happens
Example, responding with “X is wrong” is bad. Instead, going backwards and editing your prior prompt with “And dont do X btw” is far better
This is just a handful of stuff, this is literally the basics, but hopefully gives you an idea of how deep the rabbit hole can go, I didnt even touch on stuff like agentic workflows and agent orchestration, which is where shit really starts to pop off…
Based on what you’re describing, it almost sounds like creating this development environment is a full time job outside of the typical development structure, someone who has specialty knowledge of best practices to use AI that isn’t a senior dev or project manager.
I wouldn’t be surprised if the job of AI manager becomes a job in mid-size firms.
-
Replace tokens with compute and that sounds a lot like programming.
…no?
We arent talking about performance of the app, we are talking about how much money it takes to make the product itself.
Its like arguing that we should stop having workers on a site job use power drills cuz they keep putting holes in everything.
Like… no… you should train your workers how to use a power drill correctly and to stop putting holes in everything.
And if you then go “oh cool so I should stop hiring new guys?” The answer is “the fuck are you gonna do when your senior guys retire, dumbass?”
You obviously still hire new guys… Just… fuckin train your workers to use the tools.
If you buy super expensive tools for your workers and then dont train them how to use them right, you dont get to be shocked when you find out the tools are being used poorly…
You can use AI efficiently and effectively but its not easy, these tools require training and skill to make them work well.
Companies keep telling their workers to use AI but arent spending the money to train them to use AI
You need like minimum 50+ hours of training baseline for these skills to even remotely make the tools worth using.
Its not even a lot but damn these companies wont even do that much.
Then its ShockedPikachu.jpeg when their untrained employees blow through 100k tokens in a day cuz they dunno wtf they are doing.
Bunch of idiots with power tools is what it is, lol.
