Ozzy Arreguin (AG-47)

@ErikVoorhees Commercial idea (free/my ip) for Venice/VVV and DIEM: Scene 1: Guy vibe coding like a rockstar. Speaking into a boom mic, orchestrating AI agents. Completely in the flow. Scene 2: The screen stops mid-sentence. Retro pop-up flashes giant red text: "LIMIT REACHED!!!" Scene 3: Panic sets in. Frantically opens Apple Store app to buy hardware. Configures a Mac Mini: "OUT OF STOCK." Voiceover: "Out of compute? Mint some when you need it." Scene 4: Smash cut. The same guy, total zen, in a beach chair under a palm tree. Phone buzzes: “Idle daily compute rented out. +$5.00 USDC.” Voiceover: “...and lend it when you don't. Stop paying rent for your intelligence, own it." $VVV / $DIEM

@MATLAB @lmstudio

128 GB of memory is nice, but you can get started with local agentic AI workflows with much less. By connecting gemma 4 in @lmstudio to MATLAB MCP Server, you can run a local AI model that uses MATLAB as a tool on a 16 GB RAM MacBook Pro. 🔗 spr.ly/6010BD2aXC

The "Brain Size" Argument (Memory vs. Math)Twitter focus on TOPS (Trillion Operations Per Second) which is just speed, not smarts.The Frontier Model: A subscription model (like GPT-4o or Claude 3.5) has hundreds of billions—sometimes trillions—of parameters. It requires warehouses full of giant supercomputers just to hold the model in memory. It can write complex code, reason through hard logic, and understand deep context.The 67 TOPS Device: This hardware usually only has 8GB of memory. Even if the chip can do math quickly, it physically cannot hold a frontier model. You are forced to use a 3-billion parameter model.The Analogy: A 67 TOPS device is a hyper-fast calculator. A frontier model is a university library. The calculator can do simple math instantly, but it cannot write a master's thesis on history because it does not have the library's data.The Use-Case Clash (What They Do)They are built for completely different worlds. You can show your friend this simple contrast:FeatureFrontier Subscription (Cloud)67 TOPS Hardware (Local)The Smart LevelExpert-level writing, coding, and logic.Basic chat, simple summaries, strict tasks.Internet AccessCan browse the web for live facts.Totally offline (unless you build a web-scraper).Best TrickSolving complex, unpredictable problems.Counting cars, reading cameras, running gadgets.Cost$20 a month forever.$100–$400 one-time hardware cost.How to Say It to Them SimplyIf you want a quick script to text them, use this:"The 67 TOPS chip is like a custom sportscar engine. It is incredibly fast, but it is built to run a small, lightweight car—like counting objects on a camera or running a basic offline chatbot. A frontier model like ChatGPT or Claude is a massive cargo airplane. It is way bigger and handles massive, complex jobs that the sportscar engine physically cannot carry. You can't fit a cargo plane's worth of brainpower into an 8GB hardware box, no matter how fast the chip is."They are comparing a specialized local tool to a giant cloud supercomputer. They are both great, but they do completely different jobs.

@starmexxx Meanwhile apples putting models on phones.

AMD CEO LISA SU HELD A MINI PC ON STAGE THAT RUNS A 235B MODEL AND REPLACES YOUR $440/MONTH AI STACK amd's ryzen ai max+ 395 is the first x86 chip that runs a 200 billion parameter model on one piece of silicon. cpu and gpu share 128gb of unified memory, no separate graphics card needed the gmktec evo-x2 runs qwen3 235b fully, deepseek v3 comfortably and llama 3.3 70b with headroom. on linux you get 110gb of usable vram out of 128gb amd claimed the chip beat an nvidia rtx 5080 by more than 3x on deepseek r1 inference. a lunchbox sized pc outrunning a $1,000 discrete gpu on a real ai workload a heavy ai user pays $200 for claude code max, $200 for chatgpt pro, $20 for cursor and $20 for gemini. that's $5,280 a year and the box pays itself off in 9 to 10 months install ollama, pull the model, point claude code at localhost. same interface, nothing leaves the machine, nothing costs per request bookmark this and read the article below

@josephfkrause @localghost You should just trade in your privacy tokens for convenience tokens. PRIV tokens are overrated anyways.

@localghost I’m big in privacy but always annoyed by how hard it makes it in everyday life. What made you go this way versus just give up and add it all?

Got a mac mini for clawdbot. Had a lot of fun setting this up today. Instead of access to my accounts, I gave it: ✅ its own apple account for messages ✅ its own gmail to sign up for stuff ✅ its own github to push code

@itsnoahd its based on who needs it the most. real time analysis of your phone usage is used.

I am generally curious as to how the Siri waitlist is being handled because I’ve been on it since launch with no luck.

@Kyriakos_Pelek @HowToAI_ not only that but our cell phones our the modern day version of our secrets keeper. A journal, ledger, news and communications device..all that data, just when data is becoming scarce.

@Kyriakos_Pelek @HowToAI_ offline AI on an always on, cellular network connected device. How could this go wrong?

Apple just did something nobody expected. They turned 2 billion iPhones into local AI machines. They open-sourced coreai-models, the entire toolkit that lets you export any HuggingFace model and run it natively on iPhone, iPad and Mac with zero cloud. → Runs 100% on the Neural Engine → No cloud. No API keys. No subscriptions. → Fully offline. Your data never leaves the device. It even ships with skills for Claude Code, Codex, and Gemini, so your coding agent already knows how to use it. 100% Open Source.

@HowToAI_ then you realize how much a 2b model really can do. fail

@TaylorHabrock @adrgrondin @lmstudio In mother china, kimi runs you.

@ozzy_ag47 @adrgrondin @lmstudio How do you plan on running K2.6 on 128gb? 🤨

LM Studio on stage for WWDC at the Steve Jobs Theater demonstrating MLX distributed on 4 Mac Studio! Coming later this year @lmstudio 👾

@JulianGoldieSEO Why didn’t you just use dictate.

I asked my AI agent how its day was going. It replied out loud: "Splendid as always, sir." So I pushed it further. → I said "open Google" and the page loaded instantly → I said "schedule a nap at 3pm" and it confirmed → I checked my calendar... and it was actually there This isn't a chatbot that just talks back. It runs real tasks on my computer using the new voice + computer skills. That's the part that feels like the future. 🚀 Save this video, you'll want to build your own Jarvis. Want the SOP? DM me.

You PUSHED it to schedule a nap in your calendar….and it did it??? Damn bro how does it feel to be on the bleeding edge of tech? Siri been doing this for hella long.

@OpenAI What a great commercial. Everyone with a fully autonomous AI agent running 24/7 working as their personal assistant and then *ding* the hot pockets are ready and suddenly the camera pans out to the dude sitting in his mom's basement watching the AI agent work on his computer.

It's time to fly.

@AlexFinn “They’re taking are jerbs!” If only people thought about how many jobs the infrastructure is gonna create, even if just for the build.

Scariest video I've watched this week America's next generation actively booing the most transformative technology our species has ever seen In China grandmas line up to get OpenClaw installed In America, supposedly our most educated people BOO even the mention of AI The west simply does not stand a chance if this continues We have a massive AI marketing problem in this country and nobody is doing anything to fix it Tomorrow Meta will announce 8,000 layoffs. They will blame it completely on AI They won't blame it on their irresponsible hiring in 2021 or extended elevated rates or horrible market conditions or bad inflation No, in order to not tank their stock, they'll blame it on AI College students will read that and learn to HATE the technology They'll protest outside datacenters holding ridiculous signs that say "SAVE OUR WATER NO MORE DATACENTERS" Politicians will see this and run on blocking data centers just to get a few votes All of it will be a cycle that leads to America losing to China This should be a warning sign to all frontier labs and CEOs: messaging matters. And if your messaging doesn't change the West is cooked

And by prompt, there’s levels. In open source models that will allow it, you can make a top level system prompt, then there’s the “one shots” prompts many people are used to: “You are an expert in property law.” Then, I don’t know if I would say lastl because in some models like Claude Code for example, it’s literally the message you send “prompting” it to do something. If you ever used dispatch you may have noticed that while you chatted or “vibed” with Claude chat, every now and then it would send a concise, formatted list of steps to take in the background. The macOS app with the UI will let you toggle the background display on.

@Alex_Rogov_js @AlexFinn The prompt. As in prompt engineering. Recently also rebranded as prompt caching or “multi step prompting” for this agentic AI craze. It just means writing clear introductions in proper format.

It's official. Claude Code just released /goal The single most underrated AI feature of 2026 Now Claude Code, Codex, and Hermes agent has it It allows your agent to complete long running tasks, sometimes for days EVERYONE should be immediately running this prompt: 'Based on what you know about me, my goals, ambitions, and what we've built together already, what are the 3 /goals we can run right now that would run for long time periods and produce the best results?' Choose one, then ask for it to build you a prompt You should get a few options for super powerful goal prompts that will have your agent of choice complete long running tasks that will deliver your mind blowing results Carve out 15 minutes tonight to do this. Thank me later.

@Alex_Rogov_js @AlexFinn I’m I mean ai did what is supposed to do?

@AlexFinn /goal + CLAUDE.md is the combo. /goal sets the destination, CLAUDE.md constraints prevent drift. ran a 4h session migrating a 20k line TypeScript monorepo agent stayed on track because architectural rules were explicit. it could plan long but couldn't break patterns