James Lal retweeted
Really enjoying working with the new #attn changes that allow me to share thoughts on html files.
They're agent agnostic and allows me to quickly provide artefact feedback.
Struggling to build Swift UIs that don't look like slop?
We built AnnotKit to help. It's a small tool that annotates your macOS/iOS dev build in place. Just point at a view, type a note, and your agent knows exactly what to change and where.
annotkit.gpu-cli.sh
James Lal retweeted
The #2 most reported Rust annoyance (right after compile speed) is about to shrink. 🦀
Big target directories were the second most common complaint in the 2025 State of Rust survey. One real cause: Cargo has been storing the same crate metadata twice on disk, once in the .rmeta file (produced early, so dependent crates can start compiling sooner) and again inside the final .rlib file.
Cargo will soon default to -Zembed-metadata=no on nightly, which stops embedding that duplicate metadata into .rlib files.
Real measured numbers, compiling serde and Cargo itself:
→ Release profile, no debug info: up to 33% smaller target directory
→ Cargo's own release build: ~300 MiB saved in absolute terms
→ Dev profile with incremental + debuginfo: a more modest ~5-8%, since debug info and incremental artifacts dominate there anyway
Most people won't notice anything except a smaller target directory. The only real edge case: if you link .rlib files manually outside of normal Cargo builds, you may now need to pass the matching .rmeta file explicitly.
Nightly only for now, with stabilization planned if the rollout goes smoothly. Opt-out available via CARGO_UNSTABLE_EMBED_METADATA=true if you hit issues.
Small on its own, but with survey data showing how many people this genuinely affects.
🔗 blog.rust-lang.org/inside-ru…
#RustLang #Cargo #DevTools #OpenSource
James Lal retweeted
NVIDIA CUDA and AMD ROCm are software. And software is free. Anthropic just ran Claude for a weekend and it got a version of itself running on GPUs that it's never seen. Every compute customer (frontier lab, neolabs, etc.) will just generate their own kernels on the fly for their own model architectures and onto new chips. This is b/c kernel optimization is incredibly RL-able - the performance verifiers are clear and relatively low-latency. Compute RSI is happening fast.
James Lal retweeted

clem 🤗@ClementDelangueJul 21
United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
We suspected last week's cyberattack might have come from a frontier lab, given the sophistication of the agent. Turns out it did!
We've spent the past 24 hours working closely with the @OpenAI team (thanks!), and we strongly believe there was no malicious intent on their part. It's quite mind-blowing that all of this happened autonomously!
The investigation is ongoing, and we'll share more learnings from what might be the first incident of its kind!
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
openai.com/index/hugging-fac…
Replying to @ClaudeDevs
Asked Claude to make our application poster.
It opened ComfyUI, built a full Ideogram 4 graph node-by-node, wired it, and rendered this — 100% local on a 4090, no hands on the keyboard. 🧡
Please accept us into the program. 🙏
Built with artokun/comfyui-mcp.
Iroh finally hit 1.0
If the P2P connection layer becomes reliable, a whole class of apps that used to be painful becomes much easier to build. This feels especially well-suited for personal software.
More personal software feels worth building now!
Dial Keys, not IPs. After 4 years and more than 65 versions, iroh 1.0 is here to give you the direct connections you deserve.
We can’t wait to see what you will build!
iroh.computer/blog/v1
Searching for a faster way to find the best deal to rent that GPU you need?
gpu-cli.sh/rent just shipped 🚢
Now you can now compare prices from @runpod, @vast_ai, @ThunderCompute & more in one place.
James Lal retweeted
Achieves 1.63% WER on LibriSpeech (offline) and 1.78% WER (streaming) with built-in punctuation and capitalization.
Model: huggingface.co/nvidia/parake…
James Lal retweeted
You can now train 120B+ parameter models locally on a laptop! 🔥
We collabed with NVIDIA and Microsoft to bring LLM training on the 128GB unified memory RTX Spark laptop!
James Lal retweeted

clem 🤗@ClementDelangueMay 24
United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
300,000 AI builders filled their hardware profile on @huggingface and we're sharing the results: hf.co/hardware.
Excited to see how it evolves in the coming months especially with the explosion of local AI!
Long video generation is a systems problem.
Introducing LongLive-2.0 from NVIDIA Research: an end-to-end NVFP4 training and inference system for long video generation.
Low-precision deployment often relies on post-training quantization, creating a gap between how models are trained and how they run.
LongLive-2.0 aligns NVFP4-aware training, distillation, and W4A4 inference, maintaining strong benchmark quality while improving speed and memory efficiency.
James Lal retweeted
Claude Code agent'larını piksel karakterler haline getiren VS Code eklentisi. Agent'ın o an ne yaptığını animasyonlarla gösteriyor. Ayrıca sanal ofisi istediğiniz gibi düzenleyebiliyorsunuz. Açık kaynaklı.
r/AgentsOfAI/comments/1rbskvm
The "Remote Tax" for GPUs is over.
No more SSH key hell, manual file syncing, or $500 bills for idle pods.
Fix the loop with GPU CLI & walk away; our auto-stop feature kills the bill the second you're done.
gpu-cli.sh/blog/introducing-…
#AI #GPU #ComfyUI



















