Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-09-16 · source-backed
Michael Noukhovitch named the failure mode: standard GRPO concentrates compute on easy problems while hard ones stagnate, and a scalar average hides it. NGU starts with small completion counts and re-queues unsolved problems with probability p, producing a geometric distribution of attempts weighted toward hard cases. On GSM8k, k=4 with p=0.9 beat every standard GRPO configuration on the hardest subsets; on DeepScaler with Qwen 3 4B base (~120 H100 hours) it improved AIME 2025 and BRUMO 2025 without regressing easy problems. The transferable rule is to read per-difficulty metrics, not one average.
Each link below shares sources, entities, or timing with this story.
New batching algorithms enable ~7x, up to 12x+, longer-context GRPO training with no accuracy or speed penalty versus optimized FA3 and chunked-loss setups (Unsloth Docs). Qwen3-8B GRPO reaches 110K context on one 80GB H100 via vLLM plus QLoRA. For solo builders doing reasonin...
The letter to Senators Tim Scott and Elizabeth Warren, dated June 10 and surfacing publicly this week, frames it as model distillation run against Claude at scale (Anthropic). A related claim pegs it at 28.8 million fraudulent exchanges, though that figure is single-sourced an...
Alibaba released Qwen3.6-27B on April 22. Dense architecture. Open weights. 77.2% on SWE-bench Verified, within 3.7 points of Claude Opus 4.6. On SkillsBench, it scores 48.2% versus its own 397B MoE predecessor's 30.0%. That's a 77% improvement with 14.8x fewer parameters. Let...
Moonshot AI dropped Kimi K2.6 today and the numbers are hard to ignore. One trillion parameters total, 32 billion active per token across 384 experts, 256K context window, and native multimodal input. It scores 58.6 on SWE-Bench Pro versus GPT-5.4's 57.7 and Claude Opus 4.6's...
Alibaba post-trained its flagship in place on September 1, keeping the 2.4T-parameter base and 1M context. All eight published coding benchmarks improved: TerminalBench 3.0 from 11.3 to 29.0, DeepSWE 1.1 from 56.6 to 69.3, QwenSWEbench V2 from 55.1 to 70.0, JobBench from 53.4...
Staged on ModelScope for 23:00 Beijing time August 26, it's roughly 125B parameters plus a separate N-gram embedding table of about 51B, activating 6B per token, with GDN gated-delta hybrid layers and Qwen Sparse Attention. Alibaba frames it as a technology preview of the Qwen...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.