@DavidOndrej1 Are you using any of the messaging apps as your communication channel? ssh? something else?
English
AgenticRebirth
33 posts
AgenticRebirth
@AgenticRebirth
building with AI, sharing the journey in real time
Joined Ocak 2026
43 Following36 Followers
Pi has been my go-to agent over the past few days
I've been using all other AI Agents less and less
if you put in a few hours to learn it, Pi really is the shit
English
This is genuinely so cool. Thank you @pewdiepie.
Going to get stuck in and do a writeup over the next few days.
pewdiepie-archdaemon.github.io/odysseus/
English
@therealSteveCr I didn’t mention not sending data to third parties, but good point!
English
@AgenticRebirth You should have also added: work that is security sensitive, and should be kept off a third party's servers.
English
I spent a weekend running numbers on LLM subscriptions versus running models locally. I wanted to find out how to maximise the quality of my LLM usage per dollar. What I found genuinely surprised me.
At 5 million tokens a day, which one solid agent workflow with tool calls burns through in a morning:
🔵Claude Opus 4.8 costs about $1,500 a month.
🔵GPT-5.5 runs closer to $1,700.
That is $18,000 to $20,000 a year on token bills.
A local machine with an RTX 5090 costs about $4,000 to $5,000 and a used RTX 3090 runs $800 to $1,000. Electricity adds maybe $40 a month.
The break even point is three to six months depending on which API you choose to use. After year one you are $10,000 to $16,000 ahead. Those are not small numbers. In reality, the gap is even larger, since I use billions of tokens a month. Then there are the intangible benefits like no rate limits, no vendor lock-in, no sending your information to third parties.
The "buy a GPU" crowd (h/t @TheAhmadOsman) actually had it right.
What actually makes sense is running both, and routing based on the task:
🔵Local models for bulk work: evals, experimentation, batch processing, anything where zero marginal cost changes how freely you iterate.
🔵API models when quality is the actual constraint: customer facing output, complex reasoning, the decisions that cost more to get wrong than to pay for.
For maximum efficiency per dollar, you could use DeepSeek V4 Flash for things that do not need a frontier model, and use Claude Opus or GPT-5.5 for the 30 percent that genuinely do.
There are only two questions that actually matter: what is your real daily token volume, and what is your quality sensitivity for each category of work you do. Your best setup follows from those.
Looks like I'm buying some GPUs.
English
Super cool. I have no idea how to do this, but am tempted to try and break things...
Mario Zechner@badlogicgames
I wrote up how I built the shitty robot so you can too. This was a fun project that will keep on giving. Thanks to all the open weights folks out there, without whom this would not have been possible. mariozechner.at/posts/2026-05-…
English
@JorgeMvrfil Building more random things is part of being productive.
> Start a DnD campaign.
> Realise I don't like any of the existing map editors.
> Create my own for personal use, might productise it.
Do this for 15 different friction points you encounter, and you've got 15 products.
English
be honest:
are you actually more productive with AI coding tools…
or just building more random things?
English
@sudoingX is the cursor sub still ultra subsidised? Looking to switch now that my $100 ChatGPT sub got nerfed
English
if you've got a cursor sub and you're only running composer 2.5 inside the editor, you're leaving the best half on the table.
put it through hermes agent and composer's frontier brain picks up hermes's memory layer, skills, and cross session depth. same model, far more agent. and it's not chained to one editor, it runs on linux, mac, or windows, from the terminal or telegram, wherever you actually work.
live on my fork today while the official path takes its time. cursor sub or credits, plug it in and feel what frontier plus real memory does.
Sudo su@sudoingX
if you have a cursor sub or credits, you can test this now on my fork. github.com/sudoingX/herme…
English
@AgenticRebirth Now do the math on cloud inference of models you would run on few 3090/5090 and see if it still makes sense
English
@AgenticRebirth I honestly didn’t know you could run Opus 4.8 or GPT 5.5 locally. Apples to apples comparison would probably be more like buying Qwen3.6-27B on the cloud since that’s probably what you’re running locally. I think that would end up at $0.60/mil tokens. About $90/mo.
English
I'm of two minds about this.
1. You're right about this. There's a lot of work going into building productivity shaped tools instead of actually being productive.
2. But some small fraction of these tools will be genuinely useful. Why create IDEs when you can just use a text editor? Why make higher level languages when you can just code in lower level ones? Why create graphical interfaces when you can just use the CLI?
In general, abstracting and creating new primitives can add a lot of power. We're still early in the LLM coding saga, so it makes sense for people to be building out the tooling.
x.com/WillManidis/st…
English
"AI people are increasingly building tools that build tools that build tools."
-endless agent workflows
-recursive productivity systems
-Llttle real-world value
-lots of self-referential outputs
Brad Mills 🔑⚡️@bradmillscan
One of my friends @RowanNorth recently described the AI workflow space as "a human centipede of eating each other’s shit" What do you think 😂
English
@royjossfolk This goes without saying! Optimising for token usage on its own is pointless.
English
@AgenticRebirth It is more important that the thing you are spending tokens on turns into something that has value beyond just breaking even on API costs.
English
AgenticRebirth retweeted
Buy a GPU, The Movement, was always right
AgenticRebirth@AgenticRebirth
I spent a weekend running numbers on LLM subscriptions versus running models locally. I wanted to find out how to maximise the quality of my LLM usage per dollar. What I found genuinely surprised me. At 5 million tokens a day, which one solid agent workflow with tool calls burns through in a morning: 🔵Claude Opus 4.8 costs about $1,500 a month. 🔵GPT-5.5 runs closer to $1,700. That is $18,000 to $20,000 a year on token bills. A local machine with an RTX 5090 costs about $4,000 to $5,000 and a used RTX 3090 runs $800 to $1,000. Electricity adds maybe $40 a month. The break even point is three to six months depending on which API you choose to use. After year one you are $10,000 to $16,000 ahead. Those are not small numbers. In reality, the gap is even larger, since I use billions of tokens a month. Then there are the intangible benefits like no rate limits, no vendor lock-in, no sending your information to third parties. The "buy a GPU" crowd (h/t @TheAhmadOsman) actually had it right. What actually makes sense is running both, and routing based on the task: 🔵Local models for bulk work: evals, experimentation, batch processing, anything where zero marginal cost changes how freely you iterate. 🔵API models when quality is the actual constraint: customer facing output, complex reasoning, the decisions that cost more to get wrong than to pay for. For maximum efficiency per dollar, you could use DeepSeek V4 Flash for things that do not need a frontier model, and use Claude Opus or GPT-5.5 for the 30 percent that genuinely do. There are only two questions that actually matter: what is your real daily token volume, and what is your quality sensitivity for each category of work you do. Your best setup follows from those. Looks like I'm buying some GPUs.
English
@AgenticRebirth did you manage to set it up with premium subscription or do you have premium+ ?
English
Last week I had to decide between Openclaw and Hermes. After installing, configuring and briefly test driving both, the choice wasn't difficult.
Hermes was fully capable within 5 minutes of being installed. Browser automation, OS control, X searching. It just works out of the box. Openclaw could only do the same after a fairly tedious onboarding and setup, and even then not as well.
Hermes configuration is easy. Two CLI commands and you can switch off anything you don't want. It takes 5 minutes, I don't understand all the fuss. Openclaw configuration... don't even get me started. Incredibly opaque and unpleasant.
Obviously, I stuck with Hermes.
Teknium 🪽@Teknium
Just want to make this clear: We didn't make Hermes Agent to be a "starts with nothing, you work it all out" agent. This is not the minimalist, start from nothing, agent. We want Hermes to work out of the box for most people. So you aren't spending weeks just getting the agent to work, or have the capabilities you need. This means that yes, there are more built in things then something like nanoclaw or pi, which start with nothing, and you just have to figure it out. That is an intentional design decision. You can from the modest baseline that has capabilities that are likely broader than you need, but not egregious, take it from there if you want to tinker with it. Run `hermes skills config` or `hermes tools` to disable whatever you want. We even have a way to upload your whole "Agent" as a github repo, so you can install hermes fresh with your exact setup again later or share them. We have a massive interface for extensions so you can tinker with it to infinity. But if you don't want to become an agent engineer - with Hermes, you don't have to.
English
@badlogicgames He's just pointing out the philosophical difference between an opinionated harness and a minimal one. Makes sense to me.
English
@NeurAlch Yep. I use a separate coding harness, so all I really need my agent to do is this kind of work. Hermes just works.
English
@AgenticRebirth Hermes setup is amazing in a few minutes i have reminders, crons for hacker news daily, talk to it thru Whatsapp, have it update it's own personality, coding things on a kanban board, etc
English