30
30
Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL
Hugging Face blog describes running async GRPO reinforcement learning with LoRA across HF Jobs using a storage bucket and proxy instead of NCCL.
A Hugging Face blog post titled 'Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL' explains an asynchronous Group Relative Policy Optimization training setup using LoRA adapters distributed across Hugging Face Jobs workers. The architecture coordinates training through an object storage bucket and a proxy server, removing the need for NCCL collective communication. No full article text was available at classification time.
25
How Hugging Face Inference Endpoints, Jobs, and Buckets Power Search on Papers with Code
Hugging Face explains how Inference Endpoints, Jobs, and Buckets power semantic search on Papers with Code.
Hugging Face describes the infrastructure behind search on Papers with Code, built on its Inference Endpoints, Jobs, and Buckets services. The post is a product-focused engineering walkthrough with no security impact.
20
Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets
Hugging Face, Strands Agents, and LeRobot integrate with Storage Buckets for a unified record-train-deploy robotics data workflow.
Hugging Face announced an integrated robotics workflow combining LeRobot, Amazon's Strands Agents, and Hugging Face Storage Buckets. The setup lets developers record robot data, stream it in a data loop, train models, and deploy agents from a single place. No article body was available, so details beyond the title are limited.
30
30
30
dealignai/GLM-5.3-CYBERSECURITY-FP8 — new model trending #13 on Hugging Face
dealignai releases GLM-5.3-CYBERSECURITY-FP8, a 753B MoE weight-modified variant cutting refusals on offensive-security prompts, trending #13.
dealignai released GLM-5.3-CYBERSECURITY-FP8 on Hugging Face, a cybersecurity-domain 'crack' of the 753B-parameter GLM-5.3 MoE model, currently trending #13. The release directly edits bf16 residual writers, keeps FP8 routed experts, and serves with stock vLLM on 8x H200 GPUs with 131k context. HarmBench-320 evaluations show 80-84% direct harm compliance and 89% cyber-offense compliance, while MMLU rose 1.07 points to 86.65%. Copyright-verbatim reproduction remains a known soft-refusal limitation, with an UNCENSORED sibling variant offered.
30
30
30
Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile AI raised a $2B Series B from GreenOaks and Index Ventures to scale human-behavior simulation for Fortune 100 clients like CVS.
Simile AI, co-founded by Generative Agents researcher Joon Sung Park, announced a $2 billion Series B backed by GreenOaks and Index Ventures, with Fei-Fei Li and Andrej Karpathy among backers. The company runs tens of millions of simulations for Fortune 100 clients including CVS, reporting 85-99% accuracy versus human focus groups and digital twins of 1,000 real people at 85% behavioral accuracy. The long-term ambition is foundation models of human behavior, post-trained on interviews, transaction data, and randomized controlled trials, potentially simulating all 8 billion people.
58
30
30
30
The Rise of the Forward Deployed Engineer — and How To Do the Job Right
Palantir veteran Vinoo Ganesh traces the forward deployed engineer role and shares practices for building effective FDE teams.
Kepler CEO and former Palantir forward deployed engineer Vinoo Ganesh argues that labs, startups, and PE firms hire FDEs without a shared definition of the role. He recounts Palantir's Project Frontline rotation, which trained about 250 software engineers as FDEs, many now leading forward deployed teams at OpenAI, Anthropic, xAI, and Anduril. A 2013 failure of the Phoenix transaction store at a bank, where real-world data gaps caused roughly 2.3 million keyspaces and an out-of-memory crash, illustrates why FDEs must own the gap between design and production reality. At Kepler he places the FDE function inside product rather than sales.
30
BreezeBlue/Breeze-TTS-2 — new model trending #19 on Hugging Face
BreezeBlue open-weights Breeze TTS 2, a bilingual text-to-speech model it ranks #1 among open-weight models on the Artificial Analysis TTS leaderboard.
BreezeBlue released open weights and Apache 2.0-licensed PyTorch inference code for Breeze TTS 2 on 2026-08-25. The text-to-speech model supports English and Chinese, voice cloning, reference-free voice design, voice direction, and inline vocal events like (laugh) and (sigh). Reported performance includes #1 open-weight ranking on the Artificial Analysis Elo leaderboard, under 40 ms time-to-first-audio, a 0.32 real-time factor on an NVIDIA H100, and about 7.7 GiB GPU memory for eager inference.
45
30
30
50
30
Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks
Survey of four harness mechanisms—context budgeting, compaction, todo-state, and memory—that keep long-horizon LLM agents on task across 200+ tool calls.
The article details how agent harnesses, not larger context windows, solve context overflow and goal loss on long-horizon tasks, citing Chroma's Context Rot report showing 18 LLMs (GPT-4.1, Claude 4, Gemini 2.5, Qwen3) degrade on long inputs. Concrete implementations include LangChain Deep Agents offloading tool responses over 20,000 tokens to the filesystem and truncating old tool calls at 85% window usage, and Claude Code capping auto memory at 25KB while re-reading the 5 most recently modified files after compaction. OpenAI's Responses API now offers server-side compaction via context_management with a standalone /responses/compact endpoint, which Codex uses for long-running coding tasks. Manus reports a roughly 100:1 input-to-output token ratio per ~50-tool-call task, motivating todo.md state recitation to prevent goal drift.
52
Ask HN: How do you manage skills files?
A Hacker News thread debates whether agent skill files are worth managing, citing 2–4x output-token reductions on flagship models in one company's testing.
Commenters argue skills are stored prompts that help less-technical users compensate for weak prompting, while one participant reports company testing found skills reduce flagship-model output tokens by roughly 2–4x, a gap growing with newer models. Others note skills can bundle reusable scripts and inline commands for deterministic context building, and that harnesses now execute backticked commands before the agent sees the skill. Some argue improving model capability makes downloadable skills redundant.
15
Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed
Perplexity details its GPU embedding serving stack (Ivy, Tulip, ROSE), which reuses LLM prefill/decode kernels, CUDA graphs, and LazyTensors to cut launch overhead.
Perplexity engineers published a deep dive on the serving infrastructure behind pplx-embed, used across Perplexity Search and its API platform. The stack comprises Ivy (Rust HTTP gateway), Tulip (gRPC scheduling and batching), and ROSE (Runtime-Optimized Serving Engine), which reuses LLM prefill and decode kernels rather than running a separate embedding engine. Optimizations include whole-model CUDA graphs with lazy capture and a LazyTensor abstraction that overlaps CPU batch preparation with in-flight GPU work. Benchmarks are reported against vLLM v0.22.0 in BF16, with FlashAttention 4 generally fastest but FlashInfer 3 winning on Qwen-based models at very long sequence lengths.
33
[AINews] Hot Chips: OpenAI’s Jalapeño, Cerebras CS-5, Groq 3 LPX, Apple M6
OpenAI unveiled Jalapeno custom inference chip claiming 1.5-1.9x better perf-per-watt than NVIDIA GB200/GB300, deploying in-house by year-end.
At the 37th Hot Chips conference, OpenAI published first benchmark details for its custom Jalapeno inference chip, claiming 1.5-1.9x more work per watt, 1.7-3.6x lower end-to-end latency, and 2.1-4.1x higher interactive-workload performance versus NVIDIA GB200/GB300, with the 700W-rated part staying at or below 550W in tests. Deployment into OpenAI's own infrastructure begins by year-end, with Gen 2 deep in development and Gen 3 underway. OpenAI also said GPT-Astra and Codex helped write low-level kernels, reportedly 1.5-1.8x faster than human-expert code for selected attention and MoE blocks. Cerebras CS-5, Groq 3 LPX and Apple M6 were also featured at the conference.
72
30
30
30
30
35
30
30
30
30
30
30
30
30
30
30
30