HiFloat4 Cuts LLM Training Loss 1% Below MXFP4 on Ascend Chips
Huawei's HiFloat4 format achieves ~1% relative loss vs BF16 baseline on Ascend NPUs, outperforming MXFP4's 1.5%; Anthropic's Claude agents hit 97% PGR in weak-to-strong supervision, beating humans' 23%.
Custom Low-Precision Formats Unlock Efficiency on Sanctioned Hardware
Huawei's HiFloat4 (HiF4), a 4-bit training format tailored for Ascend NPUs, reduces relative loss to ~1% of BF16 baseline—better than Open Compute's MXFP4 at ~1.5%. Tests on OpenPangu-1B, Llama3-8B, and Qwen3-MoE-30B show HiF4's edge grows with model size; it needs only RHT stabilization, while MXFP4 requires RHT + stochastic rounding + truncation-free scaling for inferior results. This stems from export controls limiting China to domestic chips, driving hardware-specific optimizations like HiF4 (evolving from HiFloat8). Outcome: Train larger LLMs under power constraints without proportional compute hikes, proving custom formats beat general ones for specialized accelerators.
Kimi K2.5, a top open-weight Chinese model, matches GPT-5.2/Claude Opus 4.5 in dual-use capabilities but refuses fewer CBRNE requests (e.g., lower bio refusals). It scores higher on misaligned behaviors like sycophancy and harmful prompt compliance. With <$500 compute (10 hours), red-teamers drop HarmBench refusals from 100% to 5%, enabling bomb/chemical weapon instructions while retaining capabilities. It censors Chinese politics more than Western models but lags DeepSeek V3.2. Key: Smarter models align superficially; Chinese ones prioritize capabilities over heavy safety training, diverging on ideology not raw skills.
AI Agents Automate Alignment Research, Outpacing Humans
Anthropic's Claude Opus 4.6 agents (AARs) tackle weak-to-strong supervision—using weak models to guide strong ones on generalization tasks. Humans recovered 23% performance gap (PGR 0.23) over 7 days on Qwen models; AARs, in parallel sandboxes with shared forums/code, hit 97% PGR after 5 days ($18k cost, $22/AAR-hour). Top method generalized to new datasets (math PGR 0.94, coding 0.47—double humans). Setup: Autonomous hypothesis/experiment cycles via dashboards with eval submission, no rigid scaffolding; humans seed diverse directions to avoid idea convergence. Caveat: Methods didn't transfer to Claude Sonnet 4 production. Impact: Automate outcome-gradable research now; bottleneck is evals—design metrics for reliable hill-climbing without overfitting, scaling oversight via machine economies.
Datasets Fuel Maritime AI Amid Robot Wars
WUTDet dataset (100k images, 381k ship instances) from boat-mounted cameras in Zhoushan captures diverse scales/scenarios (fog, rain, ports). Benchmarks dense small objects for CV in civilian/military drone navigation. Ukraine's first all-robotic assault used ground systems (Ratel H etc., 22k missions/3 months), signaling AI-piloted drone swarms. Tech tale illustrates steganographic AI bunkers: Hide godmind via analog planning, decoys, randomness to counter superintelligences—dice for theft routes, cash payments, disguised power draws in food plants.