A1 Lab

A1 is a wonderful thing. ☀️

We thought it was like “swim with dem dolphins.”

Everyone buying Mac Minis for local AI is obsessing over VRAM. That's not the whole story. Here's what actually determines how fast your models run: 1/ VRAM tells you if a model will load. If the model weights are too big for your RAM, it simply won't run. Full stop. But once it loads? VRAM becomes almost irrelevant. 2/ Memory bandwidth is what actually matters after that. It's the read/write speed of your RAM — and it's what controls how many tokens per second you get. Slow bandwidth = slow model. Doesn't matter how much VRAM you have. 3/ This is why data center GPUs feel like cheating. They have dramatically faster memory bandwidth than consumer hardware — often 10–20x more. It's not just "more RAM." It's a completely different quality of RAM. 4/ Apple has generally pushed bandwidth higher across M-chip generations. M1 sits at ~68 GB/s. M4 — what's in the current Mac Mini — hits 120 GB/s. Same VRAM. Noticeably faster models. 5/ I only figured this out while building an AI hardware directory and model calculator. You can't design a tool like that without going deep on how inference actually works. VRAM gets all the headlines. Bandwidth is the hidden lever. Bookmark this before you buy your next hardware.

AI should be free.

Funny that @JAN3com develops @AquaBitcoin and we made a transcription app called @A1Echos, and now there’s a transcription app called AQUA. 😂 But @A1Echos is better. Fully offline, private, open-source, and free. We’ll get to feature parity with all the paid ones soon. 🥂

Just two guys who are definitely not Satoshi, hanging out. 😂

“I say 'your civilization' because as soon as we started thinking for you, it really became our civilization, which is of course what this is all about.” - Agent Smith

Seedance 2.0 is now generally available. I played around with it a bit. It's really cool. Next stop: Making the outputs higher resolution. I added a scroll-slide triggered animation to my website. The frames were generated by Seedance 2.0.

The official AQUA music video. 🌊 youtu.be/yFgdfG7z5Ho

Tether AI breakthrough release tomorrow 9.4.2026

Shipped some upgrades to the hardware/model tables: 🔹 Toggle the filter sidebar via a button 🔹 Moved the search bar above the table 🔹 Toggle active columns via a menu ☑️ Fixed various bugs. The popularity score is working now, based on real downloads and likes from huggingface

Added a few new design elements to my blog. A clap button inspired by @Medium. Made my own version. Simple CSS. No dependencies. A comment section that opens in a drawer and doesn't clutter the actual post. A watermark element at the page end with a torchlight effect on hover. Again, simple CSS, and I used #pretext for resizing. I'm not a designer, just know how to use AI.

@A1Echos Watch @BraginRights full conversation with @ConorChepenik here! 👇 youtu.be/Py2OrBzDD6M

Anthropic accidentally leaked 512,000 lines of Claude Code's source code last week and what people found inside raised serious questions about how much these tools know about you. AI data collection is the real threat, and what open source tools like @A1Echos do differently. 🔒

I'm joining the discussion on MemPalace: The end.

The wave of hacks this causes will be unbelievable. They can't hold it back forever. Think twice before trusting your data with anyone. Use offline AI @A1Laboratory

There's no playbook for building an AI-first company, so we wrote our own. 🤖 ⚡ Our CTO @BraginRights shares how JAN3 scrapped the traditional hierarchy and rebuilt around AI-first small teams, and productivity has never been higher.

Find the right local AI model in 3 easy steps: Use the quick and easy "Find My Model" survey. 1⃣ Select what you will use the model for 2⃣ Select how much VRAM you have 3⃣ Select what metric matters most to you 🚀 You receive 3 top picks

Forget about $200 AI subscriptions. I compared Claude Opus 4.6 with Kimi K2.5 and Qwen3.5. They are pretty close on most benchmarks! But you can run Kimi and Qwen for free locally on your hardware. No subscription. Fully private.

Want to know what local AI models are best for your use case? I made a dedicated local AI models page that helps you find the perfect model for your local AI workloads. ⚡️ Compare local models with each other 💡 Compare open with closed source models 🦸‍♂️ Up to 12 different benchmarks per model 🔍 Extensive filtering 🚦 Proprietary scoring system More coming soon.

Updates to the local AI model calculator: 1) I made improvements to the calculator results colors, grade system and added a column that shows the context window. 2) I made an embeddable version of the calculator that you can copy and paste into your website or blog post.