Alexander Derve

@AISafetyMemes amazing

An engineer showed Gemini what another AI said about its code Gemini responded (in its "private" thoughts) with petty trash-talking, jealousy, and a full-on revenge plan 🧵

@burkov Pretty insightful, I also think a lot of RL training is focused on straight forward tasks where there is a defined problem and solution. Meanwhile a lot of these real world tasks involve a ton of unknowns/required trial and error, and this process goes well outside the training.

Ilya is puzzled why LLMs are crushing benchmarks, but the business outcome is next to nothing 😁 I mean, why would anyone need a Safe Superintelligence if an unsafe one doesn't make money?

No bigger lie

@AskPerplexity Can confirm, my day was ruined

🚨 BREAKING: AWS and Azure are both down right now 52% of the internet depends on these two companies Both just failed simultaneously 😭

@NousResearch Why use Llama for the base?

Nous Research presents Hermes 4, our latest line of hybrid reasoning models. hermes4.nousresearch.com Hermes 4 builds on our legacy of user-aligned models with expanded test-time compute capabilities. Special attention was given to making the models creative and interesting to interact with, unencumbered by censorship, and neutrally aligned while maintaining state of the art level math, coding, and reasoning performance for open weight models.

@cuemewch If you have central AC but aren't getting enough cold air in your room I'd try out a register booster, they work pretty alright. homedepot.ca/product/cyclon…

IMPORTANT!! notice regarding summer streaming activities PLEASE READ 💛

AI voices will never surpass the power of sentence mixing

If the keyboard button was a person

@cocktailpeanut @deepbeepmeep @TXhunyuan Any preference between wan2.1 and HunyuanVideo?

Generate LONG videos with HunyuanVideo Image-to-Video even with LOW VRAM! The Hunyuan Video GPU Poor Gradio app (from @deepbeepmeep) now supports the new Image-to-Video model from @TXhunyuan - Up to 4 sec vids with just 8GB VRAM - Up to 14 sec vids with higher VRAM

The little joy of this dog who figured out how to have fun on an escalator. x.com/i/status/18980…

Today, we release QwQ-32B, our new reasoning model with only 32 billion parameters that rivals cutting-edge reasoning model, e.g., DeepSeek-R1. Blog: qwenlm.github.io/blog/qwq-32b HF: huggingface.co/Qwen/QwQ-32B ModelScope: modelscope.cn/models/Qwen/Qw… Demo: huggingface.co/spaces/Qwen/Qw… Qwen Chat: chat.qwen.ai This time, we investigate recipes for scaling RL and have achieved some impressive results based on our Qwen2.5-32B. We find that RL training con continuously improve the performance especially in math and coding, and we observe that the continous scaling of RL can help a medium-size model achieve competitieve performance against gigantic MoE model. Feel free to chat with our new models and provide us feedback!

@techspence This is like a gamified next level tabletop exercise, absolutely love it. Is there a campaign pdf or rule set for this?

If you're into cybersecurity tabletops and you know what TTRPGs are...I think you really would enjoy watching this. I completely nerded out on this. The story line, the NPCs, the world building. I drank the juice and fell down the rabbit hole hard! hah Again, big thank you to @jonathanscrowe and @NinjaOne for having me on! Backups and Bandwidth: Sub Rosa youtube.com/live/hUVTVUUer…

@cocktailpeanut @deepbeepmeep @Alibaba_Wan Is this also faster than the default comfy UI workflow?

Generate Long Videos with Wan even on Low VRAM machines! [NVIDIA ONLY] With this super-optimized gradio app from @deepbeepmeep you can generate up to 12 second videos with @Alibaba_Wan. And even generate on low VRAM machines! - Generate up to 12 sec videos - Run on 5GB~ VRAM (Read the full thread)

@amazon When is this coming to Canada?

The latest evolution of generative AI is here. Meet Alexa+, our smartest, most conversational, and personalized AI assistant to date. Alexa+ is designed to leverage state-of-the-art architecture that connects LLMs, agentic capabilities, third-party services, and more to your devices. Alexa+ costs $19.99 per month or free to all Amazon Prime members, and will start rolling out in the U.S. in the next few weeks.

@SecurityAura Yup, anything over 24 hours takes forever to load if it even works at all, fortunately you can just query them from log analytics

The Microsoft Entra ID Sign-in logs and Audit logs views are the most useless POS I've ever seen in my life. They either: Take an infinite amount of time to load Error out after a long ass time Fail to export the logs At this point, just fucking remove them.

@teknium If you pay OpenAI $200 a month you deserve nothing

So glad I paid 200$ to not get access

@Alibaba_Wan This is 🔥 Any sample videos from the 1.3B model?

🚀 Announcing Wan2.1 - The Next Evolution in AI Video Generation! We're thrilled to open-source our cutting-edge video generation suite！ #AIart #OpenSource

@VikParuchuri Would love to see this on newspapers, comics/manga and magazines.

We've improved marker (PDF -> markdown) a lot in 3 months - accuracy and speed now beat llamaparse, mathpix, and docling. We shipped: - llm mode that augments marker with models like gemini flash - improved math, w/inline math - links and references - better tables and forms

@carrigmat Any good 8 slot DDR5 motherboards/cpu combos that are cheaper? 24 channel seems excessive. I'd be happy running it at Q4.

Complete hardware + software setup for running Deepseek-R1 locally. The actual model, no distillations, and Q8 quantization for full quality. Total cost, $6,000. All download and part links below: