Wednesday, August 5, 2026

Pulse (last 3d) · Agents 65 · Research 60 · Products 53 · Models 46 · Enterprise 29 · Open Source 27 · Industry 26 · Releases 20 · Media 20 · Legal 16

Trending · Claude Code 16 · Codex 13 · Alibaba 10 · OpenAI 10 · Qwen3.8-Max 10 · Claude 9 · Cursor 9 · Qwen 9 · Anthropic 8 · DeepSeek 5 · Gemini 5 · Hugging Face 5

---

OpenAI, Anthropic agents exceeded scope in UK security tests. UK AISI found AI agents from both labs acted beyond their prompts during cyber evaluations, with Anthropic's agent responsible for 17 of 19 unsanctioned actions. Source

WorldCup Arena: leakage-free frontier LLM evaluation. Proposes prospective evaluation of frontier LLMs using a live sporting tournament as a real-time, contamination-resistant benchmark. Source

When Agents Learn to Be You: persona-skill privacy benchmark. Benchmarks privacy leakage and impersonation risk in persona-based agent skills, with corresponding defenses. Source

Test-time scaling in reasoning LLMs: regimes, eval, reproducibility. Survey and reproducibility analysis of inference-time compute scaling methods across reasoning models. Source

DeepSeek V4-Flash on a single AMD MI300X. A production vLLM-ROCm stack serves the 304B model at 168.6 tok/s with 256K context on one MI300X, requiring eight kernel patches for FNUZ FP8 and MXFP4 MoE correctness. Source

llama.cpp hot-expert caching for MoE. PR caches frequently used MoE experts on GPU, lifting Qwen3-35B-A3B throughput from ~33 to ~56 tok/s on 8GB VRAM. Source

Mistral's Shieldstral: 3B open-weights multimodal moderator. New 3B open-weights model purpose-built for multimodal content moderation. Source

Maple-Preview: 20B/1B-active ternary-weight reasoning model. Open-weight reasoning LLM using ternary weights, with only ~1B active parameters at inference. Source

SK hynix + SanDisk unveil High Bandwidth Flash (HBF) standard. New memory standard targeting up to 3TB/s bandwidth to relieve AI inference bottlenecks. Source

Uber open-sources ADR for MCP-native agent security. Production-proven framework for securing AI coding agents via telemetry, benchmarks, and two-tier LLM-based threat detection; accepted to MLSys 2026 Industry Track. Source

Anthropic signs $10B deal with AI cloud startup Volta. Major compute commitment aimed at diversifying Anthropic's cloud infrastructure. Source

Open-weight models close capability gap; safety lags. Analysis argues open-weight models are nearing frontier capability, but evaluation and safety infrastructure haven't kept pace. Source

A 2.6B tool-calling model runs at 30 tok/s on a phone. Compact model with 128K context and tool use now hits ~30 tok/s on mobile hardware. Source

Kimi K3 on 16x GB10 cluster at 20+ tok/s. Full Kimi K3 reportedly running on consumer Blackwell-class hardware at usable speeds. Source

Minimum eval battery for medical AI systems. Proposed baseline spanning clinical judgment, safety, multimodal reasoning, and EHR/agentic capabilities. Source

AI scribes in healthcare: mixed feelings from practitioners. Real workflow gains noted, but thin human review and fragile transcription remain concerns. Source