eavaria

@EstebanCervi Si no eres dueño de las maquinas, todas las ganancias se las llevan los que las tienen. Y si lo haces manual pierdes dinero. Conviene instalarse en la zona, comprar maquinas y ser tu quien le quita las ganancias a los otros. Solo ahi puedes tener una plantas de hobby...

La fantasía de comprar unas hectáreas de olivos en Andalucía, tenerlo todo delegado, que de el 10% anual, e ir un par de veces al mes para ver cómo está todo y cómo lo mejoro con tecnología mientras me tomo un vino y como un asado con amigos ¿funciona o es un mito?

@Chile_Ranking Los trenes no son suficientes. Siempre necesitas camiones. Usar un camion para cargar un tren para cargar un camion para llevar a bodega no es rentable precisamente por la geografia... Camion de A a B es mejor... Los trenes brillan en redes complejas. La de Chile es simple...

Abro debate ¿Se puede Habilitar una Línea Ferroviaria de Norte a Sur en Chile? ¿Qué lo impide? 🤔🚅🇨🇱

@mesch Stickers nfc. Son ultra baratos y los lee cualquier celular. Pueden ser tarjetas que representan un producto cada uno y se cambian (donde incluso nfc es irrelevante) o tarjetas de prepago con un sistema web

A ver si algún gordo Claude me da una idea. Vamos a hacer un kiosko en el colegio primario de los nenes. Históricamente se paga en efectivo, que nadie tiene. Los chicos no tienen celular. Como se puede hacer compra anticipada de algo que luego lo usen en el colegio? Los que venden podrían tener un celu quizás... Alguna idea?

@seperdomi @_nasch_ No hago nada particular. Debe ser que tienes 2x12 y no 1x24. Al menos yo tengo el contexto en q8 y me 260k. Pero me preocupo de sincronizar el contexto a mano. Le tengo al lm studio que use el contexto en rolling window y de vez en cuando le mando en texto un sync de contexto

@eavaria @_nasch_ Cómo conectas Paperclip con LangGraph si se puede saber? He intentado Paperclip con Openclaw y qwen 3.5 35 B pero siempre se desborda el contexto para mis 2 rtx 3060 y queda a medio hacer cualquier tarea...

A todos los gurus de OpenClaw, que automatizaciones tienen que les está arreglando la vida?

@Gerardo_Rhdz @_nasch_ En cpu qwen me daba 5 tps, en gpu sin ultra optimizacion me da 50 (hay reportes de 100)... Para correr agentes en casa es suficiente en cpu. Parte por ahi y luego añades una fast layer en tu 3060 con algun modelo que quepa bien...

@eavaria @_nasch_ Entiendo, tengo 64 de ram a parte, crees que afecte mucho si no cabe el modelo en la de 12? Ya que también uso comfyui con modelos como z image, claro que cuando openclaw esté en uso, comfyui no esta corriendo

@CoastalFuturist Add me. I still. Feel I need to learn more about openclaw... Everything I do eventually evolves in an old fashioned agent programmed by myself 🫠

If there’s enough interest I’d like to make a group chat for people using openclaw / hermes agent heavily I really want to understand some good use cases, best practices, and just have a place for people to talk shop Comment if you’re interested

@Gerardo_Rhdz @_nasch_ Tengo una 3090 y una 3070. El problema es la ram. Yo corro stable difussion en la de 8 y qwen en la de 24... Yo lo uso en multilayer... Por ejemplo Grok me da el script de parseo y mis agentes apoyados por qwen parsean... Si algo falla chatgpt me debugea. Gasto la nada en tokens

@eavaria @_nasch_ Hola! Que hardware tienes? Qwen 3.5 es mejor que el gpt 4.1? Tengo una rtx 3060 con 12gb vram

@sciencegirl A lot of great titles. Check this one too

For those who used a computer between 1995 and 2001, what's the computer game from that time that sticks with you the most, and why

@NattyNattLoL Apparently they applied a multiplyer depending on where we are on the season. At the begining, it was all climbing... -14 and +22 LP. I'm plat and living the exact same situation... terrible MM and terrible LP. Ath the start of the season the MM was decent and the LP good

I dont know what you guys done for this season. But you need to fix it. League is genuinely unplayable. This is by far the worst state Ive ever seen league of legends soloq be in. Matchmaking is unfair. The LP gains are not climbable. I don't mind matchmaking being bad if LP gains are good and vice versa but they cant both be terrible at the same time. What is the point to play right now? I genuinely want to know, everyone I know in high elo have dropped the game currently in hope u can fix it asap so we can start playing the game we love. But this, this just isnt what we fell in love with. Please. Please, help us. @RiotPhroxzon @drewlevin

@martinlasek At least at my rates and my setup (dual xeon and 3090), the electricity bill is about $30/month at idle and $90/month at 100% usage. Add that do the bill and it's the same... the issue is if you want to be your own devops or not, and that you can eventually resell/repurpose

You can pay $200/m for Claude for eternity Or buy a Mac Studio for $166/m and run your own model for free Best part? You own it after 12 months at 0% interest: > M4 Max > 36GB Memory > 512 GB SSD Hmmm…

@Potent213 Pretry sure they considered it and the idea is to make tanks more appealing in top...

IDEA FOR TOPLANE QUEST! READ! This is for when you run TP (which in competitive is the most common choice) I think a lot of toplaners share this sentiment, that if you are NOT playing a tank the shield on TP is really whatever (it scales off 30% max hp) Often times it even just gets wasted Getting a shield feels really boring and unfulfilling (you do get bonus xp, which is only meaningful if ahead) My idea, is that when you finish your quest with TP, you can choose out of two TP's to upgrade into! The Teleport of Resilience (Defensive, Blue) > This one can offer the shield The Teleport of Fortitude (Offensive, Red) > While this one could offer potentially temporary adaptive force? I think there is many things possible to do (I think the buffs should also last longer, like 1min, but obviously can be weaker as a result) ALSO ITS COOL AF, and will give more options to toplaners, which is a really versatile role, and having a shield just doesn't feel good on a lot of toplane champs @RiotPhroxzon

@QGallouedec To me it's a gamechanger but not in the usual way... I use it to keep my context alive. Still use chatgpt for reasoning and grok for coding, but everything eventually get curated by qwen. Grok is ultra fast for coding, gpt 5.2 good for reasoning and qwen for keeping big contexts

are qwen 3.5 models as good as people say?

@KyleHessling1 Thanks! will see if I can tweak on weekend. I'm a bit scared of Q4, and obviously that's a must for long contexts... will do some tests to see how it behaves

llama-server \     -m Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf \     -ngl 999 \     --fit off \     -fa on \     -ctk q4_0 \     -ctv q4_0 \     -c 1048576 \     -b 2048 \     -ub 512 \     --override-kv "qwen35moe.context_length=int:1048576" \     --rope-scaling yarn \     --rope-scale 4.0 \     --yarn-orig-ctx 262144 \     -t 18 \     -tb 20 \     --parallel 1 \     --no-mmap \   What each setting does according to Claude:   - -ngl 999 — Offload all layers to GPU   - --fit off — Disable auto-fitting (prevents llama.cpp from silently reducing context to fit in VRAM)   - -fa on — Flash attention (essential for large context)   - -ctk q4_0 -ctv q4_0 — Quantize KV cache to 4-bit (1M context KV = 5.76GB instead of ~23GB at fp16)   - -c 1048576 — 1M token context window   - -b 2048 -ub 512 — Batch size 2048, micro-batch 512 (keeps compute buffer at ~3GB to fit in VRAM)   - --override-kv "qwen35moe.context_length=int:1048576" — Override model metadata that caps context at 262k (without this, the server silently limits slots    to 262k)   - --rope-scaling yarn — YaRN (Yet another RoPE extensioN) for context extension   - --rope-scale 4.0 — 4x scaling factor (262k × 4 = ~1M)   - --yarn-orig-ctx 262144 — Original training context length for YaRN   - -t 18 -tb 20 — 18 generation threads, 20 batch threads   - --parallel 1 — Single slot (all resources dedicated to one request)   - --no-mmap — Pinned memory for faster PCIe transfers Hardware: RTX 5090 (32GB), 128GB DDR5, Intel Core Ultra 7 265K (20 cores), PCIe 5.0   Result: 53 tokens/sec with 1M token context, fully GPU resident, zero CPU offloading

@TheAhmadOsman I just hit 53 tp/s with Qwen-3.5-35B-A3B (Q4_K_S) at FULL 1M token context running fully on a single 5090. Is this the new king of consumer inference? Thought you'd be proud! Settings in comments. OWN YOUR OWN COMPUTE! Well done @Alibaba_Qwen can't say it enough!

@KyleHessling1 @dholzric @sudoingX can you share all settings? I'm at 50 tps with 64k context

@dholzric @sudoingX I can pull almost 100 tps at the standard context, but I actually was even able to get it up to 53 tps with 1 Million token context fully on GPU using the Q4_K_M quant: x.com/KyleHessling1/…

Qwen3.5-35B-A3B. 35 billion parameters, only 3 billion active per token. 4-bit quant, 19.7 GB on disk. ran it from 4K context all the way to full native 262K on a single 3090. zero offloading. all layers on GPU. speed didn't move. 112 tok/s at 4K. 114 at 262K. flat line. default settings will OOM at 262K on this card. one flag unlocks it. full breakdown and the unlock command are in the chart below. turns out this model isn't a pure transformer. 30 out of 40 layers are Mamba2 statespace. fixed memory no matter how long the context. only 10 layers carry KV cache. that's why scaling is free. the model doesn't notice the difference between 4K and 262K because 75% of its layers genuinely don't care. full native context on a 4-bit quant on a single consumer card. already impressed before touching code quality. running Claude Code on this next. if it codes anything like it scales, this changes the local inference game.

@milesdeutscher If you think that, it means you haven't understood AI yet. A lot of times in the history, valuable things have turned into cheap colmodities (think in salt). The world built around human labor is the one you are describing with AI taking over. The world is no lobger that one.

I got chills when reading this article. I've never been more bullish on AI. And I've never been more terrified of what that means. It's written from the POV of June 2028. But it's long, so I summarised it for you: • AI gets good. Companies lay off workers. Margins expand. Stocks rip. S&P hits 8,000. Everyone celebrates. • But fired workers stop spending. Companies weaken. They buy more AI to cut costs. More layoffs. Less spending. A negative feedback loop with no natural brake. • The top 10% of earners drive 50%+ of all consumer spending. They're the ones getting replaced. A $180K product manager ends up driving Uber for $45K. Multiply that across every major city. • Ghost GDP emerges - the economy is "growing" on paper, but the money never reaches real people. Productivity is booming. Wages are collapsing. • Then it hits housing. $13 trillion in mortgages, all underwritten on one assumption: you keep your job for 30 years. In 2008, the loans were bad on day one. In 2028, the loans were good. The world just changed after they were written. • S&P crashes 38% from highs. Unemployment hits 10.2%. Markets barely react anymore. • The punchline: you're not reading this in 2028. You're reading it in February 2026. Every domino they describe has already started falling. The canary is still alive. Barely.

@Elmiillor @RiotSupportES @RiotSupport Eso yo lo aprendí de los videos de Mauro. Deberias haberle puesto más atencion 😂. Los clicks son en los pies de los models porque ahi es donde están para el juego. Donde clickeaste es la cola del dragon y los pies de Naut... resuelve a Naut.

Alguien (no platas) me puede explicar porque este smite le ha dado al nautilus? Es un bug no? @RiotSupportES @RiotSupport

@Send0o Ahahahhaha. Esto es cine. Seedance: hazme una pelicula de pandilla juvenil de los 90 donde los protagonistas sean Dani, Sendo, Werlyb, Elm, Xokas, Champi y Mellado. No puedo esperar el estreno 😂

Pochipoomer desde la cuna 🚬

@oshodanielas @diego_castro_a Te equivocas. Como sociedad tenemos el deber de proteger a cada uno de nuestros niños. Dale un cigarro y me meto. No le provees educion y me meto. Le gritas y me meto. Es debatible si darle un celular es maltrato, pero habemos muchos que creemos que si y se deberia legislar.

@diego_castro_a No, no es defender el uso; es meterte en la crianza tan personal de cada familia. Nadie tiene derecho a hacerlo.

Jamás imaginé que como respuesta a este tweet aparecería tanta gente defendiendo el uso de celulares para niños de dos años.

@tanveerntu @r0ck3t23 All of them are a consequence of not being able to handle the whole domain. They are suboptimal solutions that allow us to tackle the problems. Not an inherent part of the solution. The same happens with bitcoin. An optimal solution where the suboptimal is already standardized.

What about the following argument ? Compression is not loss — it is abstraction. Abstraction is how intelligence works. Mathematics compresses infinity into symbols. Physics compresses reality into equations. Law compresses conflict into principles. Without compression, cognition collapses under raw data. Language is not failure. It is structured abstraction.

Elon Musk just dated the death of human language and explained exactly why it has to die. Musk: “Our brain spends a lot of effort compressing a complex concept into words.” Language isn’t communication. It’s failed compression. You have a complete thought. You crush it into words. The listener gets fragments and attempts reconstruction. Everything important dies in translation. We don’t communicate. We approximate and hope it’s close enough. Musk: “You would be able to communicate very quickly and with far more precision.” Neuralink doesn’t improve communication. It replaces it. No compression. No loss. Direct cognitive transfer at the speed thoughts occur. Not describing the painting. Transmitting the experience itself. Musk: “You wouldn’t need to talk.” Five to ten years until brain interfaces make speech optional. Talking persists for sentiment. For information? Speech becomes primitive compared to direct neural transmission. Lifetime of memory in one second. Complete schematics transferred instantly. Not summaries. The entire thought structure whole and uncompressed. Not better communication. Actual telepathy at physical information limits. Musk: “Ideally, we are a symbiosis with artificial intelligence.” Humans who don’t merge with AI at high bandwidth don’t just fall behind. They become incomprehensible to the intelligence that matters. We’re already cyborgs with pathetic interfaces. Phones extend cognition through typing at words per minute when bandwidth should be terabytes per second. Neuralink doesn’t optimize that. It detonates the constraint. Five to ten years. Not fiction. Deployment window. From language as default to neural link as standard. From compressing thoughts into inadequate words to transmitting uncompressed cognition. From humans using AI to humans indistinguishable from AI at communication speeds. The species that survived by evolving language is making it extinct with technology matching how fast we actually think. The ones who don’t transition won’t just be slow. They’ll operate at such reduced bandwidth they become effectively deaf to everything happening at neural speed around them. Language served 50,000 years. It has less than a decade before it becomes smoke signals. Functional but hopelessly inadequate for anything that matters.