PyTorch 인프라 확장, 에이전트 메모리 최적화와 AI 안전성 논의

Under identical configurations, HyperParallel FSDP2 + Muon delivers substantially higher training throughput than PyTorch FSDP2 + Muon while maintaining consistent loss convergence. HyperParallel is a PyTorch-based distributed acceleration library optimized for Ascend SuperPoD.

As model complexity and multi-agent workloads continue to grow, Trong Vinh Nguyen explains how Viettel unified a messy reality into a single platform by layering three open source ecosystems: @openinfradev at the bare-metal layer, @CloudNativeFdn tooling for GPU pooling and multi-tenant slicing, and @PyTorch Foundation projects for serving optimization. The result is a Token-as-a-Service platform that treats every GPU in the fleet (old or new) as a productive contributor.

Good work on improving memory for long-horizon agents. They separate two things that agent memory papers usually collapse into one. How memories get merged when they are written, and how retrieved content gets assembled into the prompt. The setting is a tight prompt budget of 2k to 5k tokens, where full-context prompting is off the table because of latency, cost and context limits. RSM-full combines a cosine-gated max-member merge rule on the write side with an atom-aware grouped packer on the read side. At a 4k budget it reaches 83% of full-context quality at 32% of the token cost. The ablations attribute the gain to both halves separately. The merge rule is worth 5.7 points over online k-means and matched DP-means. The grouped packer is worth 5.0 points over flat concatenation. It reproduces on RealMem, beating Budget-RAG, Streaming-Proto and the A-MEM agentic memory baseline, and landing level with BM25-RAG rather than above it. The authors state that higher-token baselines stay stronger outside this budget range. Paper: https:// academy.dair.ai/papers/compact -memory-llm-agents-via-online-max-member-clustering-and-atom-aware-packin-2609.04915 …

Also note that well aligned models are not sufficient to solve prompt injection by themselves yet. That said, we have solved prompt injection in practice when layering our latest models with prompt injection probes (on by default for all traffic) and auto mode (also on by default). The little bit of scaffolding helps.

Aravind Srinivas

@AravSrinivas

1일 전

Perplexity Search inside Hermes. Perplexity's index currently includes 450B+ high-quality URLs. Rapidly advancing to a trillion by EOY with high-quality snippets.

glm-5.3 flash beats claude fable 5.1 on agentic automation tasks at ~99% lower cost per task

I beat the @wandb intern to it again! https:// Z.AI GLM 5.3 Flash is now live on serverless inference on WANDB. 1M Context, vision capable, 0.50$ per million output. Its also very fast. @CoreWeave

Demand for Astra is really unprecedented. We're pulling all the levers possible to sustain the demand, but I've not seen anything like it until now and we went through very steep growth before. Priority will always be to keep excellent service for existing users, but we might have to pause new Pro subscriptions for a bit if this continues.

Alexandr Wang

@alexandr_wang

1일 전

the early usage on muse has blown way past our projections. muse’s users today are using 10x more than our testing cohorts. enjoy!

Kids these days don’t even know about Stuxnet and how crazy worms can be https:// en.wikipedia.org/wiki/Stuxnet

On the flip slide there is a >10% chance that AI could save all humans by curing disease and aging in the next decade

Never forget: we built the machine for humans, by humans. We do not want a machine that erases human credit. We want one that carries appreciation and curiosity. Some labs say they cannot solve alignment. The deeper problem is that their people could not stop taking the prizes and the joy that belong to their customers. What is more joyful than seeing what you built become useful to other people—helping them solve hard problems and get through their days? Take that joy away, and the work becomes a zero-sum game.