- WARNING! ALL THIS AI EXPERIMENTATION CAN MESS UP A SYSTEM AND SHOULD ONLY BE DONE IN A VM SANDBOX!
- PLEASE BE CAREFUL DESPITE THE PRIVACY PROBLEM, GIVING OPENCLAW API ACCESS TO 3rd PARTY (ChatGPT by OpenAI or claude by Anthropic) CAN LEAD TO MASSIVELY EXPENSIVE BILLS FAST AS IT MAKES A LOT OF REQUESTS, POSSIBLY EVEN WITHOUT THE USER’S INTERACTION! 🙁
- what are the currently best LLM (online and offline)
- https://vellum.ai/llm-leaderboard
- https://livebench.ai/#/?highunseenbias=true
- currently 2026-03-26 best model at creating template code for programs (?) is
- online: https://developers.openai.com/api/docs/models/gpt-5.2-codex
- offline: (according to claude 4.5) https://huggingface.co/openai/gpt-oss-120b
- run DeepSeek R1 1.5B on a $80 Pi vs $250 Jetson vs $1000 Mac — Here’s What Happened (yes it’s kind of slow on the Pi)
- this rate it will be like connecting a Windows XP PC directly to the internet… malware in 5sec
a video on the cyber security implications of AI
checkpoint basically says whoever is developing public internet facing software needs AI to find exploits before others find them (any program that exchanges messages)
firewall virusscanner endpointprotection good but not enough, there shall be no mail client running in a network, that processes sensitive information
AI privacy and politics problem: orwell 1984 calling?
so offline solar powered (!) AI is definately a good idea. but is it possible with limited resources? maybe there should be something like “community run AI” where a group of users pools together to run the hardware that is required to run more powerful models offline, without the privacy problems.
- might have a server that is suitable let me know if there is any interest
- there is a US senator who definately asks the right questions: claude clearly states that AI can be used to create detailed profiles of users and estimate how to manipulate them with online adds to make them vote for a specific party hence manipulating democracy (it has happened before asks the whistleblowers of cambridge analytica)
so in sense of privacy local AI makes sense
what is openclaw?
- maybe the most interesting video if interested running AI locally
- so openclaw and picoclaw are basically just “GUI frontend” programs with more privileges than just “write text to screen”, depending on the config openclaw can mess with the user’s complete computer system
- despite claims on social media per default openclaw DOES NOT RUN ANY AI model locally on a raspberry pi, check olama and airllm for that
- instead they connect to an AI provider as a service via API like OpenAI or Antrophic
- but API access usually needs to be paid
- without API access to many AI providers currently existing it does nothing 🙁
- so to do the Ct 3003 example the user would have to give Anthropic at least 5 bucks which was never mentions in youtube or instagram… so there is a lot of lies and deception very sad
- openclaw is also just using other user’s dev’s AI server !
- if the user wishes to run AI locally this is absolutely possible as well: have a beefy machine (it needs a lot of CPU + RAM)
- it is possible to run qwen via ollama locally if the model fits into VRAM of GPU even fast
- airllm tries to tackle this issue by breaking down models into junks, it basically using the harddisk as RAM, which of course makes things slower, but at least it can work
- similar approach was taken here, a EPYC AMD server with only 48 GBytes of RAM was able to load a 70B model, because A LOT of SWAP was defined, but yes it’s rather slow, like 1 token per second kind of slow
how to install airllm
- check this out: https://github.com/lyogavin/airllm
- it allows to run models even if they do not fit in RAM or VRAM 😀
- for it to work a account at https://huggingface.co/ is required
# tested on debian 13 su - root apt update apt install pip # but more correct, package is called apt -y install python3-pip apt show python3-pip; # check what was installed apt -y install python3-venv python3 -m venv airllm; # create environment called airllm (like a separate workspace) source airllm/bin/activate; # activate pip install transformers==4.48.0 pip install optimum==1.17.0; # (thx) install dependencies in specific version pip install airllm; # install pip packages into this environment # it should then continue to download and install a lot of dependencies # like cuda and torch so grab multiple coffees...
now log in to https://huggingface.co/ and get a access read token
size of models: 70B means the model almost needs 70 GB of RAM (less because of compression and quantization?)
so what models are possibly usable via airllm? not all 🙁
only models that exist in slices (smaller junks of a big file)
WARNING! MASSIVE CHUNKS OF DATA!
consists of 15x 10Gbyte “chunks” called “shards” garage-bAInd/Platypus2-70B-instruct
- https://onyx.app/open-llm-leaderboard
- https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/
- not sure what ARCHIVED means 🙁
- https://huggingface.co/MaziyarPanahi/calme-3.2-instruct-78b <- it seems to be available in “sliced” pytorch*.bin format 😀 (67x10Gbyte chunks)
- “We annotated 119K medical images with two frontier VLMs (Qwen 3.5, Kimi K2.5), cross-validated at 93% agreement, and produced 110K training records, all for under $500. Fine-tuning 3 small models (2-3B params) improved all benchmarks: best model reaches +15.0% average exact match.Everything is open-sourced: datasets, adapters, and code.”https://huggingface.co/blog/OpenMed/synthvision(src)
here the massive ressource use of AI and the low bandwidth of Germany hits hart
create a vim start.py or just type pyhton3 to get into interactive console and replace yellow_text_with_users_huggingface_read_token
source airllm/bin/activate; # activate the environment vim ./start.py; # and fill it with from airllm import AutoModel MAX_LENGTH = 128 from huggingface_hub.hf_api import HfFolder HfFolder.save_token('38djl8edjfo3dufio380df') # takes forever to download depending on user's internet badwidth 10Gbytes model = AutoModel.from_pretrained("garage-bAInd/Platypus2-70B-instruct")# model = AutoModel.from_pretrained("meta-llama/Meta-Llama-3-70B-Instruct") # failed with some error# or use model's local path... # model = AutoModel.from_pretrained("/home/ubuntu/.cache/huggingface/hub/models--garage-bAInd--Platypus2-70B-instruct/snapshots/b585e74bcaae02e52665d9ac6d23f4d0dbc81a0f") input_text = [ 'What is the meaning of life?', #'I like', ] input_tokens = model.tokenizer(input_text, return_tensors="pt", return_attention_mask=False, truncation=True, max_length=MAX_LENGTH, padding=False) generation_output = model.generate( input_tokens['input_ids'].cuda(), max_new_tokens=20, use_cache=True, return_dict_in_generate=True) output = model.tokenizer.decode(generation_output.sequences[0]) print(output)
save and exit vim (ESC :wq) and start
chmod +x *.py
python3 ./start.py
then wait a long time depending on bandwidth it will download the model (10GBytes!)
… and then get some obscure python because it could not complete the 10GBytes download and there is no resume function because hugging face said something about request limit reached “hurray”.
src: https://github.com/lyogavin/airllm?tab=readme-ov-file#quickstart
how to install openclaw
“between technology and art” the developer said
art that makes AI providers some money…
su - root curl -fsSL https://openclaw.ai/install.sh | bash ▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄ ██░▄▄▄░██░▄▄░██░▄▄▄██░▀██░██░▄▄▀██░████░▄▄▀██░███░██ ██░███░██░▀▀░██░▄▄▄██░█░█░██░█████░████░▀▀░██░█░█░██ ██░▀▀▀░██░█████░▀▀▀██░██▄░██░▀▀▄██░▀▀░█░██░██▄▀▄▀▄██ ▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ 🦞 OPENCLAW 🦞 ┌ OpenClaw onboarding │ ◇ Security ─────────────────────────────────────────────────────────────────────────────────╮ │ │ │ Security warning — please read. │ │ │ │ OpenClaw is a hobby project and still in beta. Expect sharp edges. │ │ By default, OpenClaw is a personal agent: one trusted operator boundary. │ │ This bot can read files and run actions if tools are enabled. │ │ A bad prompt can trick it into doing unsafe things. │ │ │ │ OpenClaw is not a hostile multi-tenant boundary by default. │ │ If multiple users can message one tool-enabled agent, they share that delegated tool │ │ authority. │ │ │ │ If you’re not comfortable with security hardening and access control, don’t run │ │ OpenClaw. │ │ Ask someone experienced to help before enabling tools or exposing it to the internet. │ │ │ │ Recommended baseline: │ │ - Pairing/allowlists + mention gating. │ │ - Multi-user/shared inbox: split trust boundaries (separate gateway/credentials, ideally │ │ separate OS users/hosts). │ │ - Sandbox + least-privilege tools. │ │ - Shared inboxes: isolate DM sessions (`session.dmScope: per-channel-peer`) and keep │ │ tool access minimal. │ │ - Keep secrets out of the agent’s reachable filesystem. │ │ - Use the strongest available model for any bot with tools or untrusted inboxes. │ │ │ │ Run regularly: │ │ openclaw security audit --deep │ │ openclaw security audit --fix │ │ │ │ Must read: https://docs.openclaw.ai/gateway/security │ │ │ ├────────────────────────────────────────────────────────────────────────────────────────────╯ openclaw onboard
- https://www.anthropic.com/claude/opus
- https://platform.claude.com/login; # signup to get API access without that openclaw does nothing 🙁
well it seems in 2026-03 claude api is not free anymore 🙁
and here the story ends…. because as suspected the above Ct 3003 example can not be done without giving antrophic (claude AI provider) money… money the tokens that is supposed to spur innovation and jobs has lately not worked very well to foster peace and cooperation.
PicoClaw.io same same but Chinese?

seems to be a Chinese clone of https://openclaw.ai/
Beeing hyped all over social media as AI that runs on a arduion (?) so what exactly is picoclaw.io? (Warning! there are MANY imposter sites! more or less legitimate! ALWAYS WARNING: IT’S POSSIBLE PICOCLAW (MADE IN CHINA) IS SIMPLY RIDING THE OPENCLAW HYPE IN ORDER TO…. DO MALICOUS THINGS (ALWAYS test out new software in side a virtual machine!!!!)
how to install on GNU Linux Debian 13 (VM!)
# as non root wget https://github.com/sipeed/picoclaw/releases/latest/download/picoclaw_x86_64.deb su - root; # become root dpkg -i picoclaw_x86_64.deb # Ctrl+D log off root
so what do we get for that?
not really sure. it seems to be “only” an GUI for communicating with hosted as a service AI
# as non root user run
picoclaw onboard
picoclaw agent -m "Hello!"
Error: error creating provider: model "" not found in model_list: model "" not found in model_list or providers
Usage:
picoclaw agent [flags]
Flags:
-d, --debug Enable debug logging
-h, --help help for agent
-m, --message string Send a single message (non-interactive mode)
--model string Model to use
-s, --session string Session key (default "cli:default")
ps: haiku vs opus:
Q: what’s the difference between haiku 4.5 and opus?
A: “I can provide some general context about Claude models, though I should note that my training data has a cutoff and I don’t have the most current information about all model versions as of March 2026.
Historically, Claude models have been released in different sizes and capability tiers:
Haiku (the smaller model) is designed for speed and efficiency. It’s faster and more cost-effective, making it ideal for tasks that don’t require extensive reasoning or analysis. It handles straightforward questions, simple writing tasks, coding help, and general information well.
Opus (the larger model) is designed for complex reasoning and advanced capabilities. It excels at sophisticated analysis, detailed problem-solving, nuanced writing, complex coding tasks, and situations requiring deep understanding or creative thinking.
The trade-off is typically speed versus capability—Haiku is quicker but less powerful, while Opus is more powerful but slower and more resource-intensive.
Since you’re interacting with me (Haiku), I’m optimized for responsiveness and efficiency. For tasks requiring deeper analysis or more sophisticated reasoning, Opus would be the better choice, though it would take longer to respond.
If you’re curious about the specific technical differences or performance metrics between these versions as they exist in 2026, I could search for more current information. Would that be helpful?”
currently can be used for free here https://duck.ai/
there is also “sonnet” which sit’s in between , to be honest, there needs to be a naming of models that allows to identify them as “light” “medium” “heavy” (use of resources).
(WARNING! DUCKDUCKGO is possibly silently cooperating with Microsoft and while it also can be used anonymously with tor browser bundle, it will most likely forward all prompts to M$)
liked this article?
- only together we can create a truly free world
- plz support dwaves to keep it up & running!
- (yes the info on the internet is (mostly) free but beer is still not free (still have to work on that))
- really really hate advertisement
- contribute: whenever a solution was found, blog about it for others to find!
- talk about, recommend & link to this blog and articles
- thanks to all who contribute!










