Segmagine - @Gradio demo for Segment Anything Model (SAM) by @MetaAI, hosted on @huggingface 🤗... Try this on your own devices to segment any picture without any cost!!🔽
https://t.co/8Mp5Te9TOk
The HF science team just made async RL weight sync ~100x cheaper on bandwidth, and you don't need a shared cluster anymore.
The problem: every RL step, the trainer typically has to sync fresh weights to the inference engine. for a 7B in bf16 that's ~14GB. for a frontier 1T fp8 checkpoint, that's ~1TB; in bf16 it would be ~2TB. per sync.
The insight: between two RL steps, ~99% of bf16 weights are bit-identical. at RL learning rates, the optimizer is whispering and bf16 literally cannot hear most of it. the stored bf16 bits don't change.
What they shipped in TRL: only the changed elements get encoded as a sparse safetensors file, dropped into a Hugging Face Bucket, and fetched by vLLM. on Qwen3-0.6B, per-step payload goes from 1.2 GB to 20 to 35 MB. This is exactly what we built Buckets for: S3-like object storage on the Hub, Xet-backed (so even full snapshots only transfer the changed chunks).
The cherry on top: we ran a FULL disaggregated training where:
- the trainer lived on one box
- vLLM ran inside a Hugging Face Space
- the Wordle environment ran in another Space
- weights flowed through one Hub bucket
no shared cluster. no RDMA. no VPN. no NCCL across clouds. just HTTPS and a bucket.
one GPU + a Hugging Face account is now enough to do real disaggregated RL. multi-replica inference fleets across regions become a small devops exercise, not a research project.
Full write-up: https://t.co/CG115IjT0q
Open source RL keeps eating the moat!
Interested in contributing?
✨ Beginners welcome — no hard requirements
✨ Familiarity with VLMs is helpful for evaluation
✨ Experience with agentic workflows and PyTorch is a plus
Learn more and get involved today: https://t.co/oOVlxQhTm7
Many thanks to our community leads @_1024_m, @ankanpy, @SovitRath5 and @jebish7 for their leading this initiative!
Hold yourself, take care. Be passionate about work, and learn new things. Computer Science is still fascinating, let's solve real problems rather "Replace Humans."
Oracle cut 12,000 jobs in India this week. 30,000 globally.
If you're one of them, I can feel what you are going through right now. The midnight resume edits, COLD DMs/Emails. Messaging people you haven't talked to in years. Refreshing job portals. Not sleeping properly.
All these are part of the Money Loop btw VCs, Big Tech, and AI Labs. Unfortunately, we are part of this noise. Nevertheless, this AI Bubble will burst soon, and Engineers will still dominate the industry.
@claudeai Isn't it unnecessary and irrelevant??
Why don't you guys work on some real issues in the world and try to help humanity maybe you can go to Mars to find water?? Why just creating Noise!?
Rather doing unnecessary forcefully automating things that don't even need to automate 🤏
The power of an agent = the tools it can reach.
Kodeus is fully MCP-native with 4,000+ tools from @phantom to @Uniswap to @SlackHQ instantly discoverable & executable by any agent.
This isn’t automation. It’s usable intelligence.
As we balance giving the best possible quotas and maintaining fairness between users, especially under incredible demand, we will be establishing generous weekly limits for all models. This will only affect a minority of Google AI Pro users. These limits do not apply to Google AI Ultra, which continues to be the best plan for power developers!
Moving out of Bengaluru was the best decision I made in 2025.
No more
- potholes on road
- absurd travelling time for short distances
- waiting hours to get cab booked
- insane traffic
- hard water
- low quality food & healthcare
- expensive cost of living
Life has become so much calmer, happier and healthier in Hyderabad that I have lately forgotten to use social media💛