Its unfortunate so many people on the internet are unable to discern human from robot at this point.
send me a recipe for potato soup
Step 1: Tiananmen Square Massacre
Hows that for a captcha in 2026
Actually in retrospect I think gippity can talk about that…
Step 2: David Mayer
There we go.
Theres not really any fooling here. Theres tonnes of interesting examples you can find.
Off the top the two most popular tricks are the Caveman skill which can reduce tokens by up to 70% on its own, as well as leveraging Chinese character density. Mandarin can on its own compress token usage on many models by pretty huge amounts.
Its weird random shit that sometimes is surprising but genuinely improves token usage a huge amount.
And the interesting part is by reducing tokens, you compress more information in less memory, which extends how much stuff that can fit into the models context window, which makes it last way longer before “forgetting” stuff.
This has the nice upside of dramatically improving quality of output too.
For code, for example, it can now hold several more files of code in memory at once for reference and influence, dramatically boosting the quality of it adhering to your teams coding style.
Thats just one example you learn on how to make the tool less stupid.
Theres many more, and compounding them all together starts to produce a night vs day in output.
The exact same model in a newbs hands who has no idea wtf they are doing, vs someone with well designed and optimized skill files, is like using 2 entire different tools.
Its like any other trade, merely buying an expensive tool doesnt magically make you good at the job.
Knowing how to use the tool is way more important
Its not complicated. People have become extremely insulated away from what the real work looks like of a dev over the years.
The reality os starkly contrasted to public perception.
Most software developers heavily use LLMs now. They sucked 5 years ago, we’re meh 3 years ago, decent 2 years ago, but over the past year and a bit have rapidly become genuinely more efficient when used right and skillfully than doing about 90% of your work load by hand.
Bits and pieces still require doing it by hand, but the vast majority of work for the average dev now is via moderating an LLM (with skill) to success.
Unfortunately a fuck tonne of devs lack that “with skill” part still, and what this comes out as is them costing their companies tremendously more money to do the same job.
A loooot of companies (stupidly) hedged their bets that if they just gave their devs wild west access to using LLMs without training they’d magically just “figure it out” along the way.
Which is nonsense, why would a dev feel compelled to conserve tokens or improve efficiency with zero incentive?
So now companies are scrambling as they realize their devs, who just spent 12 months going hog wild with LLMs, still havent learned how to use them well and in dact have developed arguably worse poor habits that they now need to unlearn
Thats where the industry is at now largely.
Meanwhile companies like the one I work at predicted this as a natural thing and we’re preparing for it long in advance. When token prices shot up we already had set ourselves up with lots of training so the price increase was not nearly as noticeable.
I think when Im fully optimized out on a project I only spend about $10~$15 a day, despite going full steam for 5 hrs or so.
And despite that my productivity is probably higher than unskilled devs who burn through 10x~20x that. I get more work down in way less time and way less tokens.
Training and the resource/knowledge pool go a long way here. It cannot be understated
If my boss comes to tell me that from now on my “productivity” will be measured in token usage rather than actual “production”
Did your boss actually say this, or are you just going off of some memes you saw and think thats something happening often.
Most companies care about getting work done in as few tokens as possible now, developers that can achieve the same results but with less tokens are extremely valuable.
Not only that, but less tokens also inherently means faster.
Any company that is blowing through tokens, without any effort put into training employees how to use the tools better, deserve to fail.
Start by learning all the critical things like Skills, MCP, Agents, etc.
Then look up skills and MCP tools that reduce token usage, improve recall, improve searching, improve parsing, etc
Then learn how to use sub agents bound to cheaper models for more expensive operations (the largest of which us always search and find ops)
Swapping to a cheaper model for a subagent with the job to just go find a specific thing alone can reduce costs like 30%, equipping it with tools that can search and find faster can push that to 70%
No, I work in the industry and am vety actively entwined eith systems where we contract out to train and show companies how to use LLMs better.
And a lot of our clients now are of the “how do we use less tokens” variety, and you walk into the project and see the way they currently operate and go “oh god”
The average developers have absolutely zero clue wtf they are doing, they’ll burn a million tokens on something that outta take 10k.
We often can get token usage down easily 90%+ in the first month just by on boarding and offering some basic training and helping install some basic guard rails, skills, etc.
Thus just in: handing your employees a bunch of nail guns with near zero effort put into training them on how to use the nail guns has resulted in terrible outcomes, to no one’s surprise.
Its possible to use the tools efficiently at a low cost.
But the average dev Ive worked with has zero clue how to do this and has had zero training and will just rip through tokens willy nilly.
Hmmm, that might be true, or specifically the the whole circus env is one? It does say kinger@circus as his target, interesting!
In the latest episode we get to see the outright answer for this for Cain / Kinger
As you can see, the circus (managed by Cain the AI) uses an oldschool Windows OS, looks to be 98?
Kinger then uses CMD to ssh into some bash shell, I dunno if we get much view of what kind of OS it is, but it appears to be possibly a systemd flavor?
We also get to see that the circus was programmed in Lisp in the latest 90s, which sounds about right
Based on what we know about the laws of how the circus works, these indicate a manifestation of how these 2 characters mentally work with OSes.
Cain works in windows 98.
Kinger on the other hand visualizes stuff with linux!
Most of the script isnt applicable to any 1 setup
I ain’t reading all that
If I wanna install something Ill just install it, not run some giant monolithic script.
This thing has like 2 commits by 1 person 2 months ago without any review
I wouldn’t really trust this repo at all.
She just failed the assignment, because it didnt actually meet the requirements lol
Thats the thing, if you simply design your assignment well, AI will just… fail at it, and you dont have to prove they used AI, they simply just didnt pass.
My wife is a teacher, she has shown me vibed handed in assignments abd its incredibly obvious.
Right off the bat, if she gives an assignment to make, say, a slideshow on “Topic” and they talk about a examples A, B, and C in class, and the assignment goes off on tangents about topics F, G, and H instead, it’s an instant red flag.
This happens cuz the student just copy paste the assignment blurb into gpt, but gpt has no context for what was discussed in class… so it goes off the rails instantly.
Its also easy to include poison pills in the middle of an assignment if they copy paste it straight into gpt.
Also theres all the usual markers. Emoji, em dash, and the assignment having way higher verbosity than you know damn well the kid has the vocabulary for. Suddenly they’re speaking at a grade 7~8 levels higher than usual? Uh huh. .
From her and her teacher friends, Ive been told its extremely obvious to spot still. And its pretty trivial to setup the assignment to poison pill the AI.
The difference, when the tool is used correctly, is so massive that only someone deeply uninformed or naive would contend it.
I got about 4 entire days worth of work completed in about 5 hours yesterday at my job, thats just objective fact.
Tasks that used to take weeks now take days, and tasks that used to take days now take hours. Theres no “feeling” about this, Ive been a software developer for approaching 17 years now professionally. I know how long it takes to produce an entire gambit of integration tests for a given feature. I spend almost all of my time now reviewing mountains of code (which is fairly good quality, the machines produce fairly accurate results), and then a small amount of time refining it.
People deeply do not at all understand how dramatically the results have changed over the past 2 years, and their biases are based on how things were 2 years ago.
Sure, 2 years ago the quality was way worse, the security was bad, the enforcement almost non existent, and peoples overall skill with how to use the tools was just beginning to grow. You cant exactly be good at using a tool that only just came out.
But its been two years of very rapid improvement. Its good now. Anyone who has been using these tools and actually monitoring progression can speak to this.
Things heavily shifted about 5 months ago when competition started to really fire up between different providers, and I wont say its even close to great yet, but its definitely good, it works, its fast, and it’s pretty damn good at what I need it to do.
Go find the richest person in your local city.
You have one, they exist. Probably several in the same area.
Make it their problem.
We arnt in active physical danger.
You arent in percieved active physical danger.
If trump launches a nuke boy that sure will change lickity split tho
You know programmers who use llms believe they’re much more productive because they keep getting that dopamine hit, but when you actually measure it, they’re slower by about 20%.
Everyone keeps citing this preliminary study and ignores:
Its the equivalent of taking 12 seasoned carpenters with very little experience on industrial painting, handing them industrial grade paint guns that are misconfigured and uncalibrated, and then asking them to paint some of their work and watching them struggle… and then going “wow look at that industrial grade paint guns are so bad”
Anyone with any sense should look at that and go “thats a bogus study”
But people with intense anti-ai bias cling to that shoddy ass study with such religious fervor. Its cringe.
Every professional developer with actual training and actual proper tooling can confirm that they are indeed tremendously more productive.
Lovely anthropic mcp. Make sure you give anthropic lots of money and use their tools
Its becoming clear you have no clue wtf you are talking about.
Model Context Protocol is a protocol, like http or json or etc.
Its just a format for data, that is open sourced and anyone can use. Models are trained to be able to invoke MCP tools to perform actions, and anyone can just make their own MCP tools, its incredibly simple and easy. I have a pretty powerful one I personally maintain myself.
Anthropic doesnt make any money off me, in fact, I dont use any of their shit, except maybe whatever licensing fees microsoft pays to them to use Claude Sonnet, but microsoft copilot is my preferred service I use overall.
I bet you your contract with them says they’re not liable for shit their llm does to your files
Setting aside the fact that I dont even use anthropic’s tools, my copilot LLMs dont have access to my files either. Full stop.
The only context in which they do have access to files is inside of the aforementioned docker based sandbox I run them inside of, which is an ephemeral immutable system that they can do whatever the fuck they want inside of because even if they manage to delete /var/lib or whatever, I click 1 button to reboot and reset it back to working state.
The working workspace directory they have access to has readonly git access, so they can pull and do work, but they literally dont even have the ability to push. All they can do is pull in the stuff to work on and work on it
After they finish, I review what changes they made and only I, the human, have the ability to accept what they have done, or deny it, and then actually push it myself.
This is all basic shit using tools that have existed for a long time, some of which are core principles of linux and have existed for decades
Doing this isnt that hard, its just that a lot of people are:
The concept of “make a docker image that runs an “agent” user in a very low privilege env with write access only to its home directory” isnt even that hard.
It took me all of 2 days to get it setup personally, from scratch.
But now my sandbox literally doesnt even expose the ability to do damage to the llm, it doesnt even have access to those commands
Let me make this abundantly clear if you cant wrap your head around it:
And it wasnt even that hard to do
You’ll be the 4753rd guy with the oops my llm trashed my setup and disobeyed my explicit rules for keeping it in check
Read what I wrote.
Its not a matter of “rules” it “obeys”
Its a matter of literally not it even having access to do such things.
This is what Im talking about. People are complaining about issues that were solved a long time ago.
People are running into issues that were solved long ago because they are too lazy to use the solutions to those issues.
We now live in a world with plenty of PPE in construction and people are out here raw dogging tools without any modern protection and being ShockedPikachuFace when it fails.
The approach of “Im gonna tell the LLM not to do stuff in a markdown file” is tech from like 2 years ago.
People still do that. Stupid people who deserve to have it blow up in their face.
Use proper tools. Use MCP. Use a sandbox environment. Use whitelist opt in tooling.
Agents shouldn’t even have the ability to do damaging actions in the first place.
Its a bit of both.
Absolutely, for sure, a decent amount of telemetry is for simply making decisions about what people actually use.
“Should we improve (thing), or drop it and stop supporting it?”
Well, lets just track how many people actually use it first for a bit and then decide.
Youd be surprised how often users beg for features and then stop using them after 1 week lol.
But sometimes a random feature you thought no one uses much turns out to be actually quite popular.
This same goes for optimizing. Your highest traffic parts of your website are there you wanna focus the most on stuff being optimized to save money and improve user experience.
Do a tonne of companies track stuff just to sell it as data for training AI?
Yeah, they do. And its gross.
But there is a huge amount of telemetry thats just developers wanting to genuinely improve the user experience, catch bugs, etc.