From 'A Severe Misalignment' to a Mathematics Industrial Stack: A Critical Reading and a Blueprint

2026-09-13

A point-by-point reading of the 25-Fields-medalist declaration “A Severe Misalignment of AI in Mathematics”: which arguments hold (digestion bandwidth is the real bottleneck), which are borrowed rhetoric (the misalignment frame, authority as argument, the absence of asks), and what the criticism becomes when translated into design, a four-layer industrial stack for mathematics with three paradigm routes and three falsifiable pilots.

The checkpoint is not the policy: a merged reading of PDM and RLT

2026-09-13

A merged reading of two 2026 project reports from the same author line, Prefill–Decode Kernel Mismatch (an execution-level diagnosis of why same-weight sampler and trainer paths define different policies, plus contracts, audits, and four remedies) and Recurrent Looped Transformer (an architecture that designs the sampler-trainer mismatch away with one history-dependent transition): what each actually solves, what is actually measured, and the end-to-end causal loop neither has closed yet.

Qwen3-TTS on AMD Radeon 780M (gfx1103): A Docker + ROCm Deployment Autopsy

2026-08-03

A full post-mortem of deploying Qwen3-TTS-Openai-Fastapi on an AMD Radeon 780M (gfx1103) APU via Docker + ROCm. Central finding: per-step .item() GPU-to-CPU syncs inside transformers generate() deadlock permanently on this APU because of lost HSA completion signals; routing around them with the project’s fast codebook path made the service stable.

Qwen3-TTS 在 AMD Radeon 780M 上的 Docker + ROCm 部署排障记录

2026-08-03

在 AMD Radeon 780M (gfx1103) 集成 GPU 上用 Docker + ROCm 部署 Qwen3-TTS 的完整 排障记录。核心结论：transformers generate() 每步 .item() 的 GPU→CPU 同步在 APU 上会因 HSA 信号丢失而永久卡死，绕过它的 fast codebook 方案使服务稳定运行。

AI 怎么想问题：从数学推理漫游中读到的认知画像

2026-08-02

OpenAI 的数学推理漫游系列揭示了 AI 解题的认知模式：失败绕路是结构探测器、显式反例是假说过滤器、换坐标系是主要武器、精确不变量必须守恒。这套画像与人类数学家的启发法高度同构，但事后叙事与真实理解之间的张力仍未解决。

From engineered engagement to evidence-based development: the neuroscience of short-form video and a framework for redirecting its mechanisms

2026-07-14

Short-form video platforms exploit dopaminergic prediction error, variable-ratio reinforcement, and algorithmic personalization to achieve compulsive engagement; the same mechanisms, when redirected through Self-Determination Theory-aligned design and evidence-based learning science, can promote skill acquisition and sustainable habit formation.

从提示员到系统架构师：Loop Engineering 的范式跃迁

2026-06-28

Loop Engineering 将工程师的角色从与 AI 模型逐轮对话的操作者，转变为设计自主迭代系统的架构师——这是一次人机协作契约的根本重写。

寻找复数网络的并行扫描 —— 从 Mamba 的训练哲学到复数序列模型的高效训练

2026-06-14

Mamba 通过并行关联扫描实现了序列训练的 O(log L) 并行深度。问题是：复数神经网络能否复用这一策略？本文从数学结合律出发，论证复数线性递推同样支持并行扫描，并讨论硬件落地和现有工作的可借鉴之处。

Two ways to diffuse text: DFlash block diffusion vs DiffusionGemma

2026-06-10

A side-by-side comparison of two recent approaches that apply diffusion to text generation — DFlash using a tiny block diffusion model as a speculative drafter, and DiffusionGemma building a standalone text diffusion model on the Gemma 4 backbone.

Pruning Qwen3.6-35B-A3B for RTX 5090: what I learned pushing MoE compression to its limit on a single GPU

2026-05-18

Seven sessions of REAP-pruning a 256-expert MoE model on a single GPU taught me that calibration data composition matters more than the pruning algorithm, that agentic calibration cannot lift capacity-bound agentic benchmarks, and that the right recipe at the right compression depth can move +17 BugFind points at no parameter count change.