#Web5 is more open and decentralized. Where you own the hardware, software and data, with no user agreements or contracts. #Bitcoin #lunarpunk #ArchipelagoNode
elsewhere
Joined April 2022
- Tweets4.8K
- Following3K
- Followers1.9K
- Likes9.3K
Pinned Tweet
Hey Anons,
Lab 484 is bringing the Lunarpunks to The Solarpunk Summit!
#Lunarpunk is an emerging aesthetic that was birthed from the Solarpunks.
We use defensive tech that empower users through decentralization, encryption, anonymity, faith and knowledge of self.
More👇
🤖 Made with AI
Polycarp Nakamoto priv/acc retweeted
OpenAI starts refusing normal work for a real firearms industry data business.
Wouldn’t look at an ammo dealer’s site.
Couldn’t pull range policies.
Wanted to lecture the owner.
So the founder built 3 servers with 12 AMD R9700s for $32.2k.
Same workload that six Mac Studios did at 12 jobs/min now runs at 175 jobs/min.
15× throughput. No API changes. No model drift.
Models they chose on hardware they own.
Love this qoute: "If the model can change at any time, we can’t build business processes on it."
This isn’t a basement experiment. It’s infrastructure.
Sometimes local isn’t a side project. Local AI stops being a hobby the moment the cloud starts breaking your actual business.
A few weeks ago, OpenAI's models started refusing normal work for my business.
So I built 3 servers with 12 AMD R9700 GPUs and moved the workload in house. $32.2k total.
Here's the whole story in 2 minutes. #LocalAI
Polycarp Nakamoto priv/acc retweeted
I'll be playing songs on the noderunnersradio.com until the death of the internet, or the death of me. Hoping those things are both a loooong space-time off.
Polycarp Nakamoto priv/acc retweeted
i keep watching people walk right past the door and the door to local ai is a 12, 16 or 24gb gaming gpu
> the rtx 3060 is still the #1 desktop gpu on steam
> 16gb is now the most common vram on steam, 26.9% of everyone
> 65.7% of steam sits between 8 and 16gb
that's the hardware. here's what it runs today, measured:
> 8gb: 27b with a 48k ctx
> 12gb: that same 27b at 50.8 tok/s with 131k ctx
> 16gb: 262k ctx, and it sees
> 24gb: the king qwen 3.8 27b q4
a lot of the people arguing would never spend $500 on the card that ends the argument. yes, gpu prices are dumb right now. it's still the cheapest seat you'll ever buy in this race
because the second the model runs on your own card, the rules flip
i watched opus 5.5 get cut off for reading anthropic's own blog. my local model has never cut me off once. it runs with the wifi off, it does the tasks the cloud won't touch, and for the first time you get to taste what owning your thinking actually feels like
local ai is the most underrated, under adopted thing, and the door is already sitting in two thirds of the gaming pcs on steam
what card is in your rig right now?
Polycarp Nakamoto priv/acc retweeted
NVIDIA'S JENSEN HUANG JUST MASS-DEPRECATED THE $200 CLAUDE MAX PLAN, THE $200 CHATGPT PRO AND THE $300 SUPERGROK HEAVY LIVE AT GTC 2026
sat in row 3 → watched a $500 used gpu from 2020 quietly kill my $700 a month ai bill → now the only thing i pay for ai is ~$10 of electricity
one rtx 3090 → 24gb vram → running every model they lock behind $200/mo paywalls → on a card you buy used for the price of one month of subscriptions
here's what i run on it now:
claude code → qwen3-coder-30b through ollama
18gb vram → one env var → claude code thinks it's talking to anthropic → it's talking to a 4-year-old gpu under my desk
chatgpt → qwen3.8-27b
90 on livecode bench → 18gb → the model openai doesn't want you to know exists
grok → devstral small 2
12gb → handles entire repos → the one job i kept supergrok heavy for at $300/mo
what i was paying:
claude max 20x ............ $200
chatgpt pro ............... $200
supergrok heavy ........... $300
total ..................... $700/mo
what i pay now:
used rtx 3090 ............ $500 once
power ..................... ~$10/mo
jensen held up the 3090 and said "why are you still renting intelligence when you can own it"
the gpu pays for itself in 22 days
save this before your next renewal → models in the comments
Polycarp Nakamoto priv/acc retweeted
i still can't believe this fits
uncensored⚠️ 27 billion parameter model on a 12gb rtx 3060, the whole model in a 6gb file, answering at 50 tok/s with the mtp head i grafted on. five years ago a 27b on a gaming card wasn't even a conversation, that class of model lived on servers with datacenter gpus, now it runs next to your steam library
ternary weights are how, every weight is -1, 0 or +1, about 1.6 bits each, and it still does real work, on a 16gb card it ran four agents building four games at the same time
8gb cards aren't left out either, the lean version of that head keeps a 64k window on an 8gb card at 46 tok/s
and the models keep getting smaller and smarter, something new drops almost every week, and every drop raises the floor of what a normal gaming pc can run
what card is sitting in your rig right now?
i'm running a 27b uncensored🚨model at 50 tok/s on a 12gb rtx 3060 and it does not say no to anything ⚠️
model just finished full harmbench red-team suite against it, 200 behaviors, and it refused zero. not a few. zero. aligned models sit in the 90s.
the model is ternary bonsai 2 with the heretic uncensoring pass, the weights are ternary, just -1, 0 and +1, so the whole 27b is a file small enough that a gaming card from 2020 holds it with room to spare, and it's a reasoning model. it thinks before it answers, and it still keeps the pace
the speed is the head i grafted on, a draft model that predicts the next token before the big model checks it. speculative decoding. the card fires a guess, the 27b confirms, and you get 50 tok/s instead of 39
the whole thing, weights, head and a 128k context window, sits in 10.7 of the 12 gb. kv cache at q4. nothing exotic, one card, one power plug
why uncensored. because it's mine. no api keys, no terms of service, no model deciding what an adult is allowed to ask. the refusals left in it are ones it chose, and there are almost none
that's the whole point of running local. the policy lives in your house, not someone else's
Polycarp Nakamoto priv/acc retweeted
GrapheneOS is the most honest project in mobile privacy about what it cannot fix. From its own docs on the modem and the SIM:
"Yes, the baseband is isolated on all of the officially supported devices."
"A component being on a separate chip is orthogonal to whether it's isolated."
"There's a lot of attack surface between the baseband and the kernel/userspace software stack connected to it."
"Connecting to your carrier's network inherently depends on you identifying yourself to it and anyone able to obtain administrative access."
"eSIM management on GrapheneOS doesn't require any dependency on Google Play, and it never shares data to Google Play even when installed."
"It won't connect to a Google service unless the carrier uses one themselves."
The OS can wall off the modem. It cannot change what the network knows about the line. That part is decided when you pick the SIM.
@GrapheneOS
Polycarp Nakamoto priv/acc retweeted
i think i got hooked on the wrong kind of model
recently i started running small uncensored models on my gpus, and after talking to them for a while i'm completely hooked. these little worm sized models are so underrated, they do things opus 5.5 will never do
the one i've been talking to this now is an uncensored bonsai 2 27b, a 6 gb file running on a 12gb rtx 3060 at 51 tok/s with my mtp head grafted on, and it refused 0 of 100 prompts where the base model refused 95
in the cases opus 5.5 refuses, a model this small is an order of magnitude more useful, and i'm not throwing a bluff in the air, i've run them and seen it
tell me your vram and i'll point you to one that fits
Polycarp Nakamoto priv/acc retweeted
Reminder: Omarchy doesn't use AUR for install or any of its recommended options. We moved off all that a long time ago.
But maybe it's even time to drop offering AUR as an option out of the box? Our own package repository can just expand to capture whatever most people need.
Polycarp Nakamoto priv/acc retweeted
JUST IN: Vibecoder reverse-engineers Microsoft Office using AI, releasing free, open-source alternatives to Word, Excel, & PowerPoint.
Polycarp Nakamoto priv/acc retweeted
Linus Torvalds appears to be all-in on using AI to code the Linux Kernel, working on a new branch to speed up page cache reads.
Torvalds:
"The code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction."
git.kernel.org/pub/scm/linux…
Polycarp Nakamoto priv/acc retweeted
AI will soon be so good we'll be able to analyze exactly what our software is doing behind the scenes.
“Trust us” software is dying.
Privacy revolution incoming.
We tried out the open source clone of Photoshop from @get_artcraft
youtu.be/I-Es-R7VUwM
odysee.com/@NaomiBrockwell:4…
Polycarp Nakamoto priv/acc retweeted
We are in the midst of what is certainly the biggest and most significant technological revolution in my lifetime.
And honestly that is saying a lot.
yeah, software is kinda solved. i reverse engineered gen 1 apple tv remote in ~45 minutes?
it runs on linux over bluetooth, open source
github.com/a-funk/open-apple…
Polycarp Nakamoto priv/acc retweeted
All I was using OBS Studio for was recording monologues. So just made a dedicated app for that for Omarchy. Will ship in next version, and we will drop OBS as a default as a result. github.com/omacom/monologue
Polycarp Nakamoto priv/acc retweeted
PRO TIP: If you are getting into serious AI development and command line work, don't even try it under Windows, because Windows and Linux handle slashes and quotes differently, and nearly all the LLM training is done in Linux formats. So when the LLM, through a command line interface, issues commands to delete or move files, if you are under windows some of those commands can be mistranslated into WIPING ENTIRE HARD DRIVES, or deleting entire folders, etc., and then the llm will apologize to you about how it handled the slashes incorrectly.
Use Linux instead. Ubuntu is the default but there are a hundred other flavors. It works great. I run it on almost everything.
On top of all this, Windows is spyware and malware anyway, and it sucks, is super slow and insecure, and it reboots all the time to install pointless updates that often break your entire machine. There's that.
Polycarp Nakamoto priv/acc retweeted
ho
ly
crap
Todos los productos de Adobe reimplementados desde cero, gratuitos y de código abierto
→ getartcraft.com/apps
ALT Interfaz de herramientas creativas para edición de imágenes, ilustración, video y desarrollo de bibliotecas.
Polycarp Nakamoto priv/acc retweeted
Exactly
"Expect more software to retreat to the server and resist distillation."
Translation: Everything is going to be forced into the cloud to defend against decompilation. Studios will adapt to AI by gatekeeping games behind cloud subscription services
Resist. Local 4 Lyfe
In the future no one will trust an AI they do not own.
The ability to generate in private, without internet accesd and training your own model is the benefit.
Also big AI will eventually raise their prices. Making running your own model economical if you buy the hardware now.
It seems the New Thing is to sell integrated 'AI PCs' with a small old-gen GPU, 64GBs of DDR (or less), and 1 TB SSD to run small local models for 'personalized intelligence'. I have literally seen like 5+ new local AI hardware startups hit the TL in the past week.
I do not get the market for this? Who is buying this besides hobbyists? You think normies are gonna drop $5k on a PC to run Qwen 27B just because it has 'personal context' (that openAI and Anthropic also already have)?
I think this new industry is so misplaced I have admittedly changed my mind and have a very hard time seeing LLMs be served locally in any meaningful capacity. Most people don't really care about OpenAI training on their data if it means they only have to pay $20/month for the best model in the world and even those who do, why would they not just use a ZDR inference cloud that has NVL72s + an API. You'd get far better performance than frying an RTX on your desk and nuking your electric bill.
