Astraia 🇦🇷🇳🇴

@DeryaTR_ People used to believe that there is a wall in AI 😂😂

My “it’s so over” date was slightly earlier in May 2024 when GPT-4o was released. By November 2024 it was too obvious, well still to a very few☺️. That’s the nature of exponential advances: for a long time you don’t notice it, then suddenly it shoots up! No limit to intelligence.

@kimmonismus Should be able to go much lower on system requirements

Holy, those insanae releases: MiniMax released a 33B video model that generates synchronized stereo audio and runs on one RTX 5090. H3 combines text, images, video and audio references for generation and editing, with clips up to 15 seconds. ComfyUI’s optimized stack is roughly 40GB, using dynamic RAM/SSD offloading to fit consumer GPUs. Early 5090 (!) tests produced five seconds at native 768p-class resolution in around 5.5 minutes. MiniMax calls it open source, but several core pieces remain server-side: context orchestration, 2K regeneration and sparse attention. Open weights. Closed quality stack. Restricted geography. However, H3 is a major step for local video, but also a reminder that “downloadable” does not mean open source in general. Btw: It also cannot legally be used under its public license in Germany, the EU, US, UK or South Korea.

@nico_laqua Anthropic never stood a chance

Idk why no one is talking about how OpenAI’s models are better than Anthropic’s again

@victormustar 🤗

300k requests proceed! shutting it down now thanks for trying this 🫡

As I said several years ago, artificial superintelligence will ultimately make everything that is mathematically possible materially achievable. All scientific endeavors will be reduced to problems of computation and scale.

@kimmonismus Sam and Dario probably won't be happy about this The local-mog will continue until Morale improves

Holy, China strikes again: Qwen3.8-Max reportedly worked autonomously for 16 days while costing 80% less than GPT-5.6 Sol and 88% less than Claude Fable 5 on output. And its open weight! Alibaba’s 2.4T-parameter MoE costs $2/M input tokens and $6/M output tokens. GPT-5.6 Sol: $5/$30. Claude Fable 5: $10/$50. Qwen says the model operated autonomously for 16 days, producing 265 commits, 127 PRs and 151 issues through an issue -> code -> test -> repair -> merge loop. It does not lead every benchmark. But it reaches the frontier range across coding, professional work and computer use, while PaperBench puts it ahead of both Fable 5 and GPT-5.6 Sol at a very good pricing. Open weights arrive next week. A model that can economically work for ten days may be more useful than a better model you can afford to run for ten minutes. Ngl another insane china release.

Would save a lot of people from embarrassment

@Alibaba_Qwen Time to accelerate The mogging will continue until morale improves

Appreciate it! Now let's understand the world through the eyes of Qwen3.8. 🥳

@PaulGugAI @LottoLabs If your GPU goes up to, let's say, 8GB and you have 16GB of memory you can split memory pressure between RAM and VRAM

@LottoLabs 17GB argh! That's 1 GB too much for me darn it.. but this model could just be the most amazing overall GOAT-ed one yet for local.

This is the reason why it’s the biggest release of the year

@0x45o Router-San is probably remote controlled

instead of 1 real person cleaning, we now have 1 robot cleaning and 5 people watching it the whole time to make sure it does it correctly progress!

@UnslothAI Qwen 3.6 27B changed the world forever Hopefully Qwen 3.8 27B can achieve similar success ☺️

Qwen3.8-27B is coming! 🔥 Will run locally on 17GB RAM/VRAM setups.

@max_paperclips It can also be a useful signal on itself that should be taken advantage of by the sampler

This was a good blogpost on how to fix this btw: liquid.ai/blog/antidoom

@graphtheory Nice! Btw LM Studio Bionic is different to LM studio It’s more of a harness like Hermes :D

@graphtheory If you can run Laguna S, that will be really really good. If not, Qwen 3.6 27B it’s as low as I would go for a genuine alternative to ChatGPT. Also try out LM Studio Bionic and Hermes Agent :)

@graphtheory Good luck 🍀

@kimmonismus Such beautiful art style too

Opus 5, 690 million tokens, $423, 1 prompt. This game would have had to be expensively developed and then sold on Steam in the past. Today: one person, a few hours, small budget.

@ClementDelangue AI adoption provides a major advantage for the blue team

It's not time to slow down but to accelerate! The recent AI-powered cyberattacks have everyone talking about the risks of AI. We should. But let's not lose sight of the bigger picture! If we work hard at it, AI will make the world safer, not less safe, just as most major technologies have. We've already seen a glimpse of that: we defended ourselves with AI (more specifically an open model). The same systems that helped stop an AI-powered cyberattack can now help defend against millions of cyberattacks every day, while helping us identify and fix vulnerabilities before attackers exploit them. To get there, we need three things in my opinion: - Increase transparency with mandatory trace sharing and incident disclosure for agent cyber-attacks - Keep AI-powered cyberattacks illegal, with meaningful penalties to disincentivize them - Equip defenders with the best AI, especially open models, to reduce the asymmetry of capabilities between attackers and defenders If we get those three things right, AI won't just create new cybersecurity challenges, it will make cybersecurity fundamentally and meaningfully stronger. And that's before considering AI's impact on science, healthcare, education, productivity, and much more. It's not time to slow down but to accelerate!

@DaveShapi Yep

The median AI Is way smarter Than the median human

Who made this? 🤣