X / Twitter
~"This one trick will make your LLM creative like hell"~
😉
maybe it's known but a system prompt similar to:
'You are a writer writing your diary in first person.'
is enough to push all LLMs to get further from slop
and make each generation more diferent from others
🤷🤨🤔
generative UI!
s/o @BraceSproul @__dqbd for pushing the boundary here
https://t.co/WY8fAO5GxP https://t.co/S1TGodz4C9
🎥🤖 Build a Video Summarizer with Gemma
Build a video summarization app using Gemma LLM through Ollama locally. This Streamlit application uses LangChain to process videos and generate concise summaries automatically.
Watch the full tutorial here https://t.co/r1prF1H3o3 https://t.co/Dl9C7x24Fr
I didn't see this announced, but now when you ask ChatGPT for feedback, it'll leave comments on your canvas - so cool... and feels like something that's certain to be staple of the emerging interaction language for agentic editors https://t.co/n18z4StT9i
Introducing StyleGlide, the @shadcn design system editor.
AI powered theme generation. Distributed on the registry.
It uses natural language to understand your project and creates starter kits that are targeted to your niche.
Easily edit the results in real-time to make it your own.
Generate versatile Tailwind palettes from any color, with OKLCH and P3 support.
Instantly get typography options and try out 100s of fonts in your style kit previews.
I’m excited to make it available to all. You can start using it today ⬇️
📢Pixel3DMM: Versatile Screen-Space Priors for Single-Image 3D Face Reconstruction📢
-> highly accurate face reconstruction by training powerful VITs via surface normals and UV-coordinates estimation.
The geometric cues from our 2D foundation model backbone constrain the 3DMM parameters, which allows us to achieve remarkable reconstruction accuracy - works for both single image and videos!
In addition, we introduce a new 3D face reconstruction benchmark that evaluates both neutral and posed face geometry.
🌍 https://t.co/2UBeLPeXEa
📷 https://t.co/pa1DgRgfcN
Great work by @SGiebenhain @TobiasKirschst1 @martin_ruenz @LourdesAgapito
ColBERT (a.k.a. multi-vector, late-interaction) models are extremely strong search models, often outperforming dense embedding models. And @LightOnIO just released a new state-of-the-art one: GTE-ModernColBERT-v1!
Details in 🧵 https://t.co/yYzf9BbHWO
Introducing the registry mcp. One command to make any component registry mcp-compatible.
Your Design System. Now with AI. Zero config.
We've got a lot to cover. Let's get started. ⬇️ https://t.co/uH73P2FlTR
Aero-1-Audio is out on Hugging Face
Trained in <24h on just 16×H100
Handles 15+ min audio seamlessly
Outperforms bigger models like Whisper, Qwen-2-Audio & commercial services from ElevenLabs/Scribe https://t.co/8lyqB4Ypdt
It's screenshot sorting time. https://t.co/xyzmpRQn95
raindrop is the AI monitoring platform of the future!
ofc i set up trad evals for dot. but i wanted to know what was going wrong or right in broad terms as well. just like sentry "just works" & captures all 500s.
what is a 500 error for your AI product? raindrop will tell you https://t.co/NEqc80Zf8U
Snowflake CEO Frank Slootman explains why your company priorities are wrong
“I found out early on that if you can whittle things down to just one thing, you become unstoppable. Unfortunately, people resist whittling things down to one thing because it’s really hard to decide what that one thing is.”
The former Snowflake and ServiceNow CEO continues:
“People have a very easy time telling you what their top 3-5 things are because hopefully the right things are in there somewhere… I can’t tell you how many board meetings I’ve been in where the CEO puts a PowerPoint up and it’s one bullet after another listing all of the things that are their priorities. You just know that they’re going to be a mile wide and an inch deep, swimming in glue, moving like molasses. The energy is leaving my body already just watching a long list of priorities… You’ve basically devalued what you should be doing because you’re time-sharing now with all of these other things.”
Mr. Slootman urges founders to think really hard about the one thing that matters most to your business and focusing entirely on that.
If you can’t decide, just pick one:
“I like to do things in sequence. Even if you’re not sure, do it anyways. Because in the process of doing, you’re going to find out whether you’re right, wrong, or somewhere in between, and you can adjust.”
When you prioritize just one thing, things move much faster:
“Things are going to go much quicker because have a narrower plan of attack. It’s energizing. The pace picks up.”
Video source: @twistartups @Jason (2022)
📷 Can AI understand camera motion like a cinematographer?
Meet CameraBench: a large-scale, expert-annotated dataset for understanding camera motion geometry (e.g., trajectories) and semantics (e.g., scene contexts) in any video – films, games, drone shots, vlogs, etc. Links below!
We contribute a taxonomy of motion primitives, co-designed over months with professional cinematographers, and apply rigorous quality control to label and caption all aspects of camera motion.
CameraBench shows that even the best SfMs and VLMs struggle with real-world, dynamic videos. Yet, a generative VLM post-trained on our high-quality data matches SOTA SfM (MegaSAM) in geometric understanding and outperforms SOTA VLMs (Gemini-2.5 / GPT-4o) in semantic understanding, e.g., describing how the camera moves.
📄 Paper: https://t.co/DozKB5sRF1
🌐 Website: https://t.co/RnVQfzSiUc
Work led by CMU, MIT-IBM, UMass, Adobe, Harvard, Emerson with @censiyuan1, @chancharikm, @JayKarhade, @du_yilun, @gan_chuang, and @RamananDeva.