Ismail Salim

Amazing video from @juliarturc 🙌 The best explanation you'll find for legacy, K & I quants in GGUF/llama.cpp: youtube.com/watch?v=vW30o4… Also nice to hear a nod to why we exist: “The challenge is choosing the right quant settings. This depends on your model, target hardware and trade-offs between quality and speed."

@ThatSonti @browserOS_ai Awesome!

Introducing @BrowserOS_ai – an open-source agentic browser, an alternative to Perplexity Comet. We believe browsers will become the new OS, where we offload work to AI agents that'll have access to all your sensitive data. Open-source, privacy-first alternatives need to exist.

@joekndy Insane! Congrats 🙌

Omg

@fleetwood___ 🐐

@idode_k Very cool - looks super useful. Definitely going to share with my recruiter friends!

i built an agent that helps recruiters source candidates. reply if you'd like to try it out.

@joekndy Very cool idea - right up my street. Interesting you're deploying local models for iOS too. We pretty much spend all day every day doing that. Just DM'd you

Mixy looks simple but it’s the hardest thing my cofounder and I have built in 17 years of working together All proprietary models running directly on the phone, on top of a highly performant audio engine. Mixy is built to last and be the only tool for making music you’ll need

@Aussie_Pete @ReallyEpicTuts @Blackmagic_News @Aussie_Pete - You are the 🐐 for DaVinci customer support. (Just sent you a DM btw!)

@ReallyEpicTuts @Blackmagic_News This is a known issue with iOS 18 on the iPhone 16. We are in discussion with Apple about this. For now, if you need stabilisation and you see this jump in standard mode, try cinematic mode, being aware that LUTs are not shown in the preview but the recording should be correct.

Blackmagic Camera for iOS 2.1 Update! Get support for using new iPhone 16 camera control features with Blackmagic Camera, improved Blackmagic Cloud organisation syncing and new recording bit rate options for H.264 and H.265 codecs. Download now from apps.apple.com/us/app/blackma…

@rohit_bmd @OliCoDev Awesome customer support @rohit_bmd 💪 (Just DM'd you btw)

@OliCoDev This is a limitation of Windows Media Player with Windows 10. When you render there is an option of compatibility with Windows Media Player on Deliver page.

I'm not exactly sure what is going on with Davinci Resolve right now, but whenever I update it to the latest version and try to export a simple video, for some reason it feels like the audio and video are both desynced from one another by a quarter of a second.

@awnihannun We'll be staying tuned at @RunLocalAI :)

I'm committing to only use local LLMs for the next few weeks to get a real vibe-check on the gap in perf between closed/server-side and open/local (powered by MLX of course). My favorite tools for that right now are: - The raw terminal (mlx_lm.generate / mlx_lm.chat) - LM Studio

@GavinSBaker @GavinSBaker - What are your thoughts on demand for on-device/local inference as general inference demand grows? (Both in absolute terms and proportionally to server-side)

Shifting from a pre-training centric world to an inference centric world is likely positive for compute overall. Intelligence may scale even better with test-time compute (inference) than it does with pre-training per the charts below. The balance of compute always had to move from pre-training to inference to generate an “ROI on AI.” Just going to happen a lot faster than expected. And while shifting to a test-time compute, “inference first” world is probably good for compute demand, this shift does change the type of compute. And this has an impact on who wins and who loses from a supplier perspective. More 50-100 megawatt datacenters geospatially and cost-optimized for inference. More inference “Hondas.” Fewer 1 gigawatt plus datacenters (which can be anywhere) with the networking, storage, and cooling (which enables density which simplifies networking while increasing potential cluster size) technologies necessary for coherence. Less pre-training “Ferraris.” And the number of companies doing pre-training in a “Ferrari” likely steadily shrinks over time. Satya explained this in the clearest way possible in his most recent podcast. All the back and forth about the Cowen note vs. Microsoft IR commentary in Australia is missing the forest for the trees - the CEO literally just told you he was going to shift investments away from pre-training focused compute to inference optimized compute, which he noted was different! Also Grok-3 voice mode is epic.

@idode_k 🔥🔥🔥

built a lil prototype this week! if interested in trying out let me know and i'll send you the test flight 🫡

I can't wait to try more local models 🤠

@mervenoyann Thoughts on OLMoE??

I was actually serious 😂

@fleetwood___ 🐐

🔥 VLMs on mobile devices with world-facing cameras key for proactive, intelligent computing. Local/on-device inference key for real-time, private experiences. Great to see an emphasis on smaller VLMs. Excited to see where @huggingface, @moondreamai, etc. take things 🚀

@huybery Hey @huybery - Just DM'd you on Twitter about obtaining a commercial license for Qwen2.5-3B. Please let me know if you can help :)

👏🏻Great to see such an outstanding work based on Qwen. Thanks to the Sailor Team for continuously advancing LLM democratization!

@JustinLin610 Hey @JustinLin610 - Just messaged you on Twitter DM about obtaining a commercial license for Qwen2.5-3B. Please let me know if you can help :)

Pretty cool to see our latest vl model with rag!

Snap's announcement about their on-device text-to-image model seems to have slipped under the radar… Apparently, it generates 1024x1024 images with quality that's comparable to cloud-oriented models like Stable Diffusion XL. But it can do that locally on an iPhone 16 Pro Max in <1.5 seconds! 🤯 Snap are planning to ship it soon to their ~450m daily active users, and I wouldn't be surprised if it's free. I wonder how all these subscription-driven, cloud-based image generation apps will respond… Announcement: newsroom.snap.com/ai-text-to-ima… Paper: arxiv.org/abs/2412.09619

@Snap's announcement about their on-device text-to-image model seems to have slipped under the radar… Apparently, it generates 1024x1024 images with quality that's comparable to cloud-oriented models like Stable Diffusion XL. But it can do that locally on an iPhone 16 Pro Max in <1.5 seconds! 🤯 Snap are planning to ship it soon to their ~450m daily active users, and I wouldn't be surprised if it's free. I wonder how all these subscription-driven, cloud-based image generation apps will respond… Announcement: newsroom.snap.com/ai-text-to-ima… Paper: arxiv.org/abs/2412.09619

Short but sweet talk about the WebNN API: youtube.com/watch?v=FoYBWz… Def worth checking out the YouTube playlist from @jason_mayes WebAI Summit last year. It's packed with great talks! Looking forward to the next summit!