William Ruider

@exolabs @openclaw @opencode I already did it for you 9 hours ago 😉 x.com/ruider92545/st…

EXO v1.0.70 is out. This release ships with multimodality and major enhancements for memory usage in long context use cases (e.g. @openclaw and @opencode), as well as updated model support and QOL features.

@exolabs @openclaw @opencode I already did it for you 9 hours ago 😉 x.com/ruider92545/st…

@openclaw @opencode (Sorry for the late announcement this time) Download EXO v1.0.70 here: exolabs.net Or build from source - EXO is open-source (Apache 2.0): github.com/exo-explore/exo

@marcusjihansson IBM docling / granite

@ruider92545 Like even what data models train and "don't train on" is still weird. I would argue that you still need own hardware, a local specialized LLM. For agentic RAG, you might even need to go back and look at GraphRag and optimize that so it works with security principles. 2/2

I don't blame people who voluntarily spend over $200 a month chatting with their beloved OpenClaws or Hermeses, but aren't forced to. But I'm surprised that so many people don't realize that without investing in professional NVIDIA or Apple hardware, certain specific requirements for working with sensitive or private customer data cannot be met. Such work requires running local AI. The point isn't how intelligent these are compared to frontier cloud models, but what can be done in an Air-Gap environment regarding sensitive data and customer privacy. More and more standards, such as ISO/IEC 27001 and NIST CSF/800-53, require accreditation and appropriate hardware and software. At the same time, more and more people are realizing how dangerous online work is. That is why there are more and more people who are able to meet the demand in this niche sector.

For your consideration, I am sending (a reprint) someone's post from FB below. I lost a $1,000 client because of a mistake I am embarrassed to admit. But I am sharing it anyway, because it might save YOU from the same mistake. 👇 Here is what happened. A client hired me to build a project management tool. I was excited. I went ALL IN. 🔥 ✅ Clean architecture ✅ TypeScript everything ✅ Beautiful UI ✅ Proper authentication ✅ CI/CD pipeline 6 weeks later I delivered what I thought was my best work ever. The client used it for 3 days. Then sent me this: "This is very well built. But it doesn't solve our actual problem. We need a refund." 😭 I was devastated. Then I re-read our first conversation. They said: "We struggle to track who is doing what each day." I built a FULL project management system. They needed a SIMPLE daily check-in tool. 😅 I solved the WRONG problem perfectly. The lesson that changed everything: Before writing one line of code, ask: 👉 "What does success look like in 30 days?" That ONE question would have saved everything.

🔥EXO Labs released v1.0.70 with vision support.🔥 EXO v1.0.70 is out. Multimodality, vision support, PDF via text+image, better long-context memory for OpenClaw/Hermes, updated model support. v1.0.69 brought continuous batching. v1.0.70 adds eyes. If you run a local cluster, this matters.

😱 ARE LOCAL LLMs REALLY THAT STUPID? Over the weekend, for less than $20 of burned electricity, the local AI cluster I'm using and is called by some "experts" from LinkedIn, X, FB... - Macs are too shit to even run local models - bullshit devices MLX - unreliable in performing tasks because the local LLM models are primitive and outdated compared to the frontier cloud models - slow due to old (2022) non RDMA hardware - with hardware bottlenecks and latency (300 miliseconds per hop) - using Macs it is pointless and foolish due to poor performance compare to NVIDIA - ... Performed the following: Sorted and segregated, reviewed and grouped by content 27,421 PDF files. - Total 27,421 PDF files - 2 files corrupted, unrecoverable - Total 27,419 PDF files - 24,415 files (open access) reprocessed - 3,004 hash candidates - 1,129 base versions - 933 duplicate groups Then compare them according to their hash and reduce them by getting rid of duplicates and overwritten files. Then searched for content and grouped them into folders. And then it "chewed", distilled and summarized the context from the individual folders, and "spitted out" an .md file containing the summary. What has your cloud frontier AI done for you over the weekend?🤔

@digitalix Trans or not trans, I don't care, I volunteer if you give it away.😆😆😆

I’ve done all the tests i could do. This is all i have left

@ashen_one Wow, congratulations, this is a fantastic moment for you. You can install some serious LLM models on it. Wish you success in your projects.😉

unboxing the 256gb mac studio + why i spent $8500 on it

@pthread_mutex_t @valigo Polak i Węgier to dwa bratanki.

@valigo Win32 is Hungarian. Of course everyone can distinguish Hungarian from Polish.

Let's play a game - win32 types vs Polish language: LPCWSTR PSZCZYNA WCSLEN WCZESNY LPCTSTR BYDGOSZCZ WSTRZAS HGDIOBJ DOWOD HWINSTA DLUGOSC LPCSTR DWORD KAL LPWSTR SZCZECIN PCWSTR BLAD PUHALF CHUJ UHALF

So Mr. Valentin Ignatev since you're already playing polyglot, you could educate yourself. So to be clear, the Polish alphabet uses "special" characters, so the correct words are as follows: PSZCZYNA, BYDGOSZCZ, WSTRZĄS, DOWÓD, DŁUGOŚĆ, KAŁ, SZCZECIN, BŁĄD, CHUJ. The rest is crap invented by primitives.

@digitalix Yes, 5 please...🤣

64GB anyone?

The local LLM world has a "noise problem". Half the conversation right now is benchmarks. Tokens per second. VRAM numbers. Unified memory bandwidth. People shouting dyno readings at each other like the numbers themselves are the work. They're not. A dyno test isn't a journey. A benchmark isn't a finished task. It's lab noise: useful to the engineer, useless to anyone who actually needs something delivered. And here's the part nobody talks about. That hardware cost real money. The electricity keeping it running costs real money. Every hour the rig is awake is burning watts off your meter, burning your real money. Benchmark numbers won't buy you a coffee. Benchmark numbers won't pay back what you spent on the machine. The only thing that pays the hardware back is finished work: real output you can ship, sell, use, or bank. The question that matters is simple. Did the job finish the way you wanted, from prompt to final result, and was the result worth the money and the power it took to produce? That reframes the whole NVIDIA vs Apple argument. NVIDIA is a beautiful engine for short, fast, high-throughput work: real-time chat, coding, rapid iteration. Dedicated VRAM and CUDA cores earn their keep there. Apple Silicon is a beautiful engine for long, heavy, context-wide work: 250-page PDFs, overnight cron jobs, multi-step automations that need to finish without falling over. Unified memory earns its keep there. Neither platform is wrong. The operator is wrong when they pick the tool without first picking the job, and without first understanding what the job is actually worth to them. Before anyone spends thousands on equipment, they should take the time to familiarise themselves with the hardware and with what LLM models can actually do. Understand the model sizes you really need. Understand your context requirements. Know your power budget. Figure out how much the end result is worth to your wallet, your business, or your time. It's you, the human, who decides. Not the spec sheet. Not the forum consensus. So do yourself a favour and educate yourself before you become somebody's profit. You're the one looking at what has to come out the other end: shipped, sold, built on. Ultimately, the numbers on your paycheck matter more than the numbers on your test screen. Ask yourself what's more valuable: a heavier load of cash in your pocket, or a beautiful animated dashboard with 50 agents and 50 sub-agents built for your beloved OpenClaw or Hermes. Everything else is “static hum between radio stations”. The old proverb says: what goes in, goes out. I wish you all less 💩 in and more 💰 out.

@TheAhmadOsman So there are two of us 🙃

I am a simple guy I just want my own data center

@Merocle You can compare your tests with these: youtu.be/QbtScohcdwI?si… and youtube.com/shorts/yhWSS6v…

Without going into too much detail, all the devices performed very similarly, with a difference of no more than 2.5% over the course of a 15-minute test. The software and driver versions were identical. There was absolutely no throttling. I won’t be naming any favourites just yet. Perhaps longer tests will reveal some difference in cooling or noise. But I was only concerned with performance. I also plan to get a unit from Gigabyte and HP to compare