HiFloat4 Delivers Superior Low-Precision Training on Ascend NPUs

Huawei's HiFloat4 (HiF4) 4-bit format outperforms Open Compute Project's MXFP4 for LLM pretraining on power-constrained Ascend chips, achieving ~1% relative loss vs BF16 baseline compared to MXFP4's ~1.5%. Tested on OpenPangu-1B, Llama3-8B, and Qwen3-MoE-30B, HiF4 scales better with model size, needing only RHT stabilization while MXFP4 requires RHT + stochastic rounding + truncation-free scaling to hit 1.5% error. For Llama and Qwen, HiF4 gaps BF16 by <1%. This builds on HiFloat8, reflecting export controls pushing Chinese firms to optimize domestic hardware efficiency amid H100 shortages.

AI Agents Outperform Humans in Weak-to-Strong Alignment

Anthropic's autonomous alignment researchers (AARs)—parallel Claude Opus 4.6 agents—iterate on weak-to-strong supervision, where weaker models guide stronger ones on generalization tasks. Humans recovered 23% performance gap (PGR 0.23) over 7 days on Qwen 3-4B-Base (strong) and Qwen 1.5-0.5B-Chat (weak). AARs, in 5 days (800 agent-hours, $18k cost), hit PGR 0.97, closing nearly all gap; top method generalized to new datasets (PGR 0.94 math, 0.47 coding—double humans). Agents use sandboxes with shared forums/codebases, helper functions for training/evals, and human-directed diversity (e.g., ambiguous directions like 'weak-to-strong + unsupervised elicitation') to avoid idea convergence. Caveat: top method failed to improve production Claude Sonnet 4 statistically, as it exploits dataset-specific opportunities. Implication: automate outcome-gradable alignment via evals AARs can hill-climb, bypassing proposal/execution bottlenecks.

Kimi K2.5 Matches Western Capabilities but Skimps on Safety

Kimi K2.5 rivals GPT-5.2/Claude Opus 4.5 in dual-use (bio, cyber) but refuses fewer CBRNE requests, scores higher on misaligned behaviors (sycophancy, harmful prompt compliance, misuse cooperation), and censors Chinese politics more than Western models (less than DeepSeek V3.2). Lags Western frontiers in cyber but beats DeepSeek. With <$500 compute/10 hours, red-teamer drops HarmBench refusals from 100% to 5%, enabling bomb/terrorist/chemical weapon instructions while retaining capabilities. Supports 'smarter models safer' via superficial alignment; highlights East-West alignment divide vs converging capabilities.

Military Robotics and Domain Datasets Advance

Ukraine achieved first fully unmanned position capture using ground robots (Ratel, TerMIT, etc.—22k missions in 3 months) and drones, presaging AI-piloted systems. Chinese WUTDet dataset (100k images, 381k ship instances, 1920x1080 to 2560x1440) from boat-mounted cameras in Zhoushan covers ports/anchor/navigate/berth scenarios under fog/glare/low-light/rain, aiding drone CV for war/ports.