[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: K-Pepe.png (1.3 MB, 1151x1367)
1.3 MB PNG
broke CS student here.

building a niche nonprofit SaaS because apparently my hobby is making software for organisations that also have no money.

2023 dev stack:

> ChatGPT Plus
> browser
> ctrl+c
> ctrl+v
> "why doesnt this work"
> context dies
> new chat
> explain project again
> repeat

thought $20/mo was expensive.

lol.
lmao even.

then I discovered agents.

now instead of wasting tokens manually I have engineered a sophisticated autonomous system for wasting tokens automatically.

gave agents repo access for debugging / diagnostics / triage.

moved to OpenRouter thinking PAYG would save money.

> agent:
>
> reads giant context
> thinks
> calls subagent
> reads it again
> cache miss
> checks same file for spiritual reasons
> "I have identified several possible approaches"

brother identify LESS

some weeks I was burning $30+ just testing loops.

this isn't production btw.

there are like 5 users in my imagination.

tried Codex / OpenCode Go etc. Fixed subscription felt much better.

until quotas change and yesterday's functioning agent suddenly becomes:

> rate limit :)

so now I'm using the ad-supported Freebuff client for Hermes/Prime testing because I have $0 to my name and V4 Flash + caching stays predictable.

I will watch text ads in my terminal. idc. just don't bill me $30 because an agent discovered recursion.

this whole experience radicalized me into model-provider agnosticism.

my requirements:

> 1. model works
> 2. cache works
> 3. still works tomorrow
> 4. I can still afford food

biggest lesson:

**cost/token is almost meaningless for agentic software. cost/completed-task matters.**

a "cheap" model looping 14 times can cost more than an expensive model understanding the task in 2 calls.

also never allow unlimited recursion.

> agents will reinvent philosophy.
>
> at your expense.

any other broke solo devs running real multi-step agents?

**how are you stopping API costs from becoming the largest employee at your pre-revenue startup?**
>>
Claude Max 20x costs $200/month and is still usage capped. Your brain is always free to use.
>>
mother of reddit spacing
>>
>>109608005
Agents are the most retarded concept ever. The logic behind them reminds me of NFTs for some reason
>>
Yeah, I'm not reading all that. You sound retarded.
>>
>>109608005
I work an actual job with fully funded access to all the models.
I spend $100+ PER DAY in tokens and I still have to fix shit manually, sometimes rewrite from scratch. My actual finalized output increased maybe 30% at best.
20$ a month subscription is a stack overflow replacement, not a viable agent backend.
>>
>>109608005
i liked your post, here's an upvote
what model did you generate it with?
>>
>>109608005
>any other broke solo devs running real multi-step agents?

uploaded deepseek with 10$ , millions of tokens run through. Task completed for 1$ or less.

uploaded 50$ on openrouter.
choosed kimi 3 because i hearded its best model and price was in midrange. Burned the 50$ in like 20 minutes and didnt even completed the task.

now deepseek has risen their prices.

chatgpt subscription 20$ give task, task runs 30 minutes and completes. And i can still use chatgpt more.

>me realizing subscription are super cheap and > once they cant improve their models from
> user interaction prices will rise drastically.

1000$+/month is to much for average joe,
but not for a company.
for a company its still cheaper than a employee.
>>
the near future must be socialism or we all will live in slums
while the AI centers fly unreachable in space
and robot police controls any uprising
>>
>>109608965
OP here. since this somehow became an actual discussion:

I kinda agree that bare agents are retarded.

when I moved from IDE/CLI stuff to Hermes, first experience was basically hiring a senior dev with dementia.

> knows architecture
> reads repo
> makes reasonable plan
> forgets why we are doing it
> violates decision from 40 minutes ago
> rediscovers same problem
> confidently proposes thing we already rejected

had to babysit it like a toddler.

what finally made it useful wasn't just a better model.

it was memory.

RAG + injected rules/doctrine + persistent local state.

basically:

> this server does X
> never put identity here
> we already decided this
> this is how we deploy
> don't fucking reinvent architecture again

session/state now lives on my machine too, not in some vague API memory.

that changed everything.

Hermes still occasionally wakes up with no family and no memories, but now it can actually continue real work across sessions.

at this point I think "agent" is almost the wrong word.

model + tools + retrieval + persistent state + enforced rules is the actual system.

the LLM is just the senior dev with brain damage in the middle.
>>
>>109609058
>>109608005
why are you using an ai to post slop
>>
>>109608005
Just use your brain faggot, oh wait you're a vibetard so you don't have one.
>>
>>109609058
ai is more like someone with autism.
you need exactly say what you want and dont exspect any common sense.
>>
>>109609069
Do you really have to ask these faggots why?
>>
>>109609069
idk bro I already live with real brain damage + memory problems, I don't need AI to simulate those
>>
nigga I do $200 a month claude max and I run out like 12 hours before the weekly reset you silly nigga
>>
>>109609074
if an AI was writing my posts it would've fixed the reddit spacing by now
>>
>>109609093
yes. just doesn't seem fun to ctrl-c ctrl-v so you can have a simulation of really interacting with people
>>
>>109609102
bro pays $200/mo to get rate limited professionally
>>
File: 1765277494418.jpg (83 KB, 1080x1275)
83 KB JPG
>>109608090
Reddit spacing is expecting single line breaks to have additional padding.
Like this.
They always begin new paragraphs like this.

This is ideal readability spacing used when making longer posts.

You goddamn larping faggot.
>>
File: codexauth.png (414 KB, 1807x841)
414 KB PNG
This is how I manage it, built this to rotate auth payloads to amongst users.
Means available tokens are always purposed and power users can keep working without credit top ups. Nothing gets left untouched and lets me milk mucho maximum value out of AI SIMPSAM
>>
File: 1777868432929.jpg (144 KB, 1600x1600)
144 KB JPG
>>109608005
I've been using AI to help me create social engineering attack vectors for a singular goal: Making AI expensive.

>Be me
>Use AI to write massive script
>Put it on auto play rubbersucky
>Leave rubbersucky in parking lot of large tech firm
>Random employee finds it, plugs it into a work machine
>In the blink of an eye 1200 lines of code execute
>Every Desktop AI, Every Bookmarked AI domain, etc immediately open and get 100 high cost tasks a piece
>Power shell curl
>Push updates to shared folders as executable
>Send hundreds more AI requests
>>
>>109609274
This but it's a few pictures of my ass in panties and then a full crotch shot with my penis hanging out of the panties
>>
>>109608005
go back to india
>>
>>109608005
my favorite streamer looks a little different today
>>
>>109608005
>shit that burned through billions over the better part of last decade ain't free
no shit sherlock.
>>
>>109609189
>he knows what actual reddit spacing is
you have to go back
>>
File: file.png (515 KB, 786x719)
515 KB PNG
what would your parents say if they knew you were a dirty, good for nothing prooooooooooooooooooooooompter, with no skills on his own?
>>
>>109609189
There's a difference between linebreaks for readibility and linebreaks between every. Goddamn. Sentence.
>>
>>109610794
"at least he's not an unemployed pedotrans CS grad"
>>
>>109610835
at least he's not a tripfag
you sicken me
>>
>>109609058
>wasn't just a better model.
>it was memory.
>unironically NotEcksButtWhy
You need your brain fixing
>>
>>109610864
then kys you stupid nigger
>>
sub-agents and automatically created ones especially seem like complete trash and a scam to me:
>put sota model in ULTRACODE(tm) mode
>creates a dozen subagents for research etc
>all dumb/cheap-shit models
>they obviously fail at their job
>main model takes shitty model's 'research' at face value
>this fucks everything
gee, thanks! whoever could have guessed it'll 'work' like that??
>>
File: 1767200392290067.jpg (39 KB, 540x405)
39 KB JPG
>>109608005
>>
>>109608005
>Paying for tokens
>>
>>109611258
I do the opposite. I use qwen 3.8 q4 via pi harness and llama cpp and i have it spawn sub agents which are the frontier models. I have a strict procedure and a sandbox in place so they cannot get full filesystem access and they see no secrets. It's working well and peak comfy.
>>
>>109608005
you should be smart enough to run qwen 3.8 on your own hardware
>>
>>109608005
Never buy extra credits. That shit is expensive as fuck.
>>
>>109610864
newfag alert



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.