Santiago Rodríguez Díaz

@justbyte_ More than ever! I don't like those AI-first, maybe I'll change my mind in 6 months xD

Is VS Code still the default choice for developers?

@github Lixux AppImage performance is terrible! (running Fedora 44 and more than capable machine :))

The GitHub Copilot app is now generally available. 🙌 The new home base for your work. Pick up what's next, direct agents in parallel, and land your PRs, all in one place. ⬇️ github.blog/changelog/2026…

@AIPandaX It’s a shame but I literally do all this whenever I start with a new TV (friends and family included) 🤦🏻‍♂️

A guy was ready to drop $1,500 on a new OLED TV because his 3-year-old Smart TV was freezing up and took 5 seconds just to respond to the remote. He unplugged it. Deleted old apps. Cleared the cache. The lag kept coming back. He went to Best Buy to get a replacement. The home theater installer in the blue shirt stopped him: "Before you spend a grand, let me show you something." He grabbed a remote and shook his head. "There are 8 hidden tracking settings throttling your TV's processor right now. Manufacturers turn them all on by default. Nobody tells you they exist. Let's fix this." Here's what he showed him in the next 8 minutes. 🧵

Gemma 4 Quantization-Aware Training (QAT) weights are now available on Ollama! They reduce memory requirements while maintaining model quality. E2B: ollama run gemma4:e2b-it-qat E4B: ollama run gemma4:e4b-it-qat 12B: ollama run gemma4:12b-it-qat 26B: ollama run gemma4:26b-a4b-it-qat 31B: ollama run gemma4:31b-it-qat Try them with ollama launch integrations to use with your favorite tools 👇👇👇

Dear GitHub Copilot team, I am happy to announce that I successfully burned all of my monthly tokens in under 3 days thanks to your garbage new pricing model. I'd also like to inform you that I won't be renewing my subscription or adding more budget. Best, A former customer.

@LyalinDotCom @ollama Only MLX for the moment… 😭

If you're waiting Gemma 4 12b through @ollama, its here: gemma4:12b gemma4:12b-it-q4_K_M gemma4:12b-it-q8_0 gemma4:12b-it-bf16 gemma4:12b-mlxMLX gemma4:12b-mlx-bf16MLX gemma4:12b-mxfp8MLX gemma4:12b-nvfp4MLX ollama.com/library/gemma4… You'll need Ollama version 0.30.4-rc0

@NVIDIAGeForce #RTXon Control Ultimate

Over 1,000 RTX games and apps are available now with ray tracing and DLSS. To celebrate, we're dropping Steam cash in the replies to upgrade your library with a #RTXOn title... Comment #RTXOn to enter 💸

@2010MisterChip Los patrocinadores estarán que trinan por las audiencias, jajajaja

Que os parece el nuevo horario de la final de la Champions League?

@2010MisterChip Horrible, me he perdido el partido!!

@NVIDIAAI @grok explain this

A new era of PC. 25.0528, 121.5990

@witcheer The price of ‘freedom’ an investment of 3000+$ and monthly electricity bill 50-60$ to keep this machine alive 24/7 to run an open weight model? Does anyone see this? I know there are many advantages but there’s no break even point compared to cloud providers!

BOOM Qwen3.6 27b is up! day 1 with the headless 5090 linux box. local ChatGPT running on my own hardware. and this without API costs or rate limits or data leaving my network. my current stack: - ubuntu server 26.04, fully headless - ollama serving qwen3.6 27B, entire model in VRAM - open WebUI as the frontend - tailscale for access from anywhere - tmux for persistent sessions over SSH my current architecture: browser → open webui → ollama API → RTX 5090 VRAM every layer is swappable. I can replace ollama with vLLM, swap the frontend, add models. nothing is locked in. can't wait to wire Hermes in for local inference and start benchmarking models.

You can now bring your own key and inference endpoint to the Warp Agent, no paid plan required.

.@openclaw I've migrated to @NousResearch Hermes Agent. full_stop.

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

@witcheer Why don’t you run qwen3.5:4b with a decent ctx? It runs full on 8 Gb and is not that ‘dumb’ with pi coding harness

can a local LLM replace cloud coding agents on 8GB VRAM? tested it. short answer: not yet. setup: RTX 4060 Ti 8GB, WSL2, llama-server (turboquant fork), Hermes Agent + Pi as agent frameworks. two coding tasks: a port scanner (easy, single file) and a log watcher CLI (hard, multi-file + debug loop). tested 3 model configs: >Qwopus3.5-9B-Coder (43 tok/s, fully in VRAM) clean code in general, but the model can't produce valid JSON for structured tool calls at 9B params. >Qwen3.6-35B-A3B MoE + thinking (35 tok/s, ncmoe=30) thinking tokens consumed most of the budget. tool calls: reliable, never broke. >Qwen3.6-35B-A3B MoE, reasoning OFF (35 tok/s, ncmoe=30) 2.7x faster than thinking mode. tool calls: reliable, but still super long to produce the output. also tested Pi (4-tool agent: read/write/edit/bash) vs Hermes Agent with Qwopus 9B. I got the same result: model wrote good code but got stuck in a reasoning loop during test verification. 9B breaks regardless of tooling. ---- MoE expert offload adds ~2-5s latency per API call. agent tasks need 20-50+ round trips. that compounds to 18-87 minutes for tasks that take seconds on cloud APIs. code quality was good across all configs. the models can write code. they just can't do the multi-turn planning loop fast enough (or reliably enough at 9B) to work as agents. what would fix this imho: - a 9B model with reliable structured output (fits fully in VRAM at 43+ tok/s) - faster expert offload (PCIe 5.0, faster RAM)

@AishwaryaDevv I’ve just dropped Hermes and OpenClaw was about to do it… If you need a coding harness used Codex/Claude Code period, if you want a personal assistant cool and fancy then fight against the two first

Why use Hermes or OpenClaw? Genuine question. The more I use these AI assistant wrappers, the more they feel like overlays with memory bolted on. If I can build my own AI-agnostic agent with skills, rules, context files, and a feedback loop that keeps it updated… what’s the actual value add? Is it just faster skill setup? Better tooling? Less maintenance? I even had to wipe Hermes memory because it got full. Feels like I’m missing something. People using Hermes/OpenClaw heavily, what am I not seeing?

@hgruenhagen @openclaw node-llama-cpp and whatever model it sets when you define ‘local’

Which embedding models are you using for memory-core in @openclaw ? Mostly curious about retrieval quality and latency.

@petergyang I’m more lean towards OpenClaw the level of effort from devs will come back with stability and reliability soon \ In the long run OC will lead

I caved and downloaded Hermes to try. For those of you who have tried both Hermes and OpenClaw what difference do you notice? No shilling please, just want some honest opinions

@KarenPayneMVP GPT5.4 as main, Opus for long-term plans

For Visual Studio, what is your go-to model?

@NoahKingJr To me there’s no ‘race’ or at least we are far from the finish. Indeed a loser now would be dramatic, we need a balanced share of usage

Who's gonna win the AI race? > OpenAI > Anthropic > Google > xAI