Dimasya

I really can’t believe, but SafeSpace 1.6.0 is out. Although it’s not an impressive number, it’s a massive milestone for us. Now it’s possible to store videos! Let it sink for a while how far mobile technologies advanced.

@tireduckquack бросил бухать

Кто бросал курить? Расскажите как смогли

@conductor_build Whoah!

Finally! Go get a request we are collecting feedback and I’ll personally fix all the issues 😗❤️

@WaynesWorldza @sillyswap_xyz @silhouette_ex Bridge from hypernym, sell and buy on spot bridge back? Is it really private though?

@satanworker @sillyswap_xyz @silhouette_ex so it calls Silhouette shielded API and performs shield swaps by routing through the Hypercore orderbooks. Spot assets only

Private swaps on Hyperliquid just got an upgrade > Light mode added > Smart Routing for multi-hop swaps > New markets added: $ZEC, $SOL, $ETH, $BTC > $USDH quote route added > New @silhouette_ex fee model added What else is missing?

I just claimed my .agent domain and joined the .agent community! get yours now and help shape the future of autonomous agents #RWLZHZG2" target="_blank" rel="nofollow noopener">agentcommunity.org/join#RWLZHZG2 @agentcommunity_

@kail_designs @thekitze Absolute stunner!

@thekitze Done by us✨🫶

what a dope launch vid remotion + cc would never

I probably didn’t expect that at all

Opened twitter from vacation today and the timeline looks absolutely wicked

@bobek_balinek @alpennec @SebJVidal does it have collaboration features?

I was the same, ditched Realm for CoreData as soon as MongoDB has announced their acquisition in 2019. It was a crippling decision in hindsight. I've had to work around so many weird constraints. What I've realised is that the frustration with CD/SwiftData is common enough that a project like SQLiteData has every right to reach stable, long-term commitment as an OSS.

Bye bye CoreData, good riddance

@0xSero I find Droid just sometimes hang when starting implementation, and also it spawns headless chromium which prevents my Mac from sleep :D

Here's my current stack ---------- Models: 1. GPT-5.4 - Research + Programming + Parchi + OC 2. Kimi-K2.5 - Frontend Dev + Computer Use + Parchi 3. Opus-4.6 - Writing + Review GPT + Fallback 4. Gemini-3.1-Pro - Writing + Research + Reviews 5. Secret Model - Very fast + Devops + Cheap ----------- Harnesses: 1. Codex - It's the highest value play rn 2. Droid - Best multi-model harness & missions 3. Parchi - Browser use (I genuinely think this is the best one on the market) ------------ Apps: 1. Codex / Opencode - Kitty and OC for OSS models 2. KittyLitter - Mobile Codex access ------------- Terminals & IDEs: 1. Cmux - When I need to read while I prooompt 2. Zed - When I want to read the code 3. Ghostty - GOAT simple terminal -------------- Browser: 1. Brave - I am open to Helium if u wanna tell me what makes it better -------------- Infra: 1. Tailscale - Securely linking all your devices 2. VibeProxy - Use your subscriptions anywhere 3. Amphetamine - Keep computer awake 24/7 -------------- Quality of life: 1. Superwhisper - AI voice transcription 2. Raycast - Clipboard / Calculator / shortcuts 3. OBS - I record a lot of stuff and don't pay :p

@thomaspaulmann @glazeapp PLEASE LET ME IIIIIIINNNNN

@glazeapp Joint waitlist 👉 glaze.app

Glaze apps 🤝 MCP servers Left: An app to monitor your system resources Right: @claudeai asking "which process is eating memory?" via MCP Simple example but powerful concept.

This is like podcast with LLM and Michelangelo youtube.com/watch?v=H9rF1C…

The sweet gap between how things suppose to work and how things really work

this is kinda fire? if anything suspicious government contracts should be quite easy to spot for ai as well as insider trading

what another day of prompting feels like

@0xSero wait did you realy moved to opencode GUI?

I think a lot about how I can run Kimi in a recognizable form at home, at over 20 tokens/s I don't think models smaller than 1T will be able to provide high quality consistent responses across a wide range of topics, the math just doesn't work. I would mark MiniMax as the golden standard for how much you can stuff into an LLM, it's very good at it's narrow set of work, it makes mistakes but it works and is good enough. But to get there you have to sacrifice multi-modal, world knowledge, hallucination rate increases, and multi-lingual. There's a lot of world knowledge in there that is important. ----------- I tried a bunch of things, for example extreme REAPs, it would just collapse after 60% reduction. I tried Sherry ternary quantisation, LoRA to recover accuracy, 1 bit quants. The closest we can get is the 1 bit unsloth quants but vllm doesn't run it fast or at all so what's the point. ------------- New idea: What if you can provide a gigantic dataset to the model and watch it's outputs, then cluster all the weight activations for certain topics like: - Language - Coding - Agentic reasoning - Image description - Multi-Lingual - Casual - Assistant - Roleplay There will be a lot of overlap and fuzzy places, but nothing 1 LoRA can't fix. Now you map your topics to expert activations, next you load your model architecture into vllm, without any weights. You create 10 REAPs each one zeros out everything but the topics associated experts, so you can get it down to 1.25-2 bits pre on average pre-8bit quantisation. Your router will the load and unload expert weights, process the context, then get unloaded, you can basically maintain 16 slices of experts of the model, that get loaded and unloaded dynamically with very little latency if stored on CPU, each slice maintains the original 8 bit weights of the specific topic. So basically you introduce latency only between switching of experts, and you can make that more or less of a problem based on how small you want to go. This means you can run the original 8bit weights on 3.75% of the original FP16 without losing any fidelity in model responses. Will this work over a few turns? it surely will over 1 turn, I've demoed this. the latency is negligible as long as I am not switching the experts too often.

@andrey_sitnik Какой отличный пост!

Очень милый пост, напоминающий нам, что несмотря на трудности эмиграции, детская версия вас очень гордилась бы вашим путём instagram.com/p/DVaqHgvDUDA/

that's what I pay internet for

Introducing Valantis Prime. Unifying the HyperEVM and HyperCore via a singular account layer: