이번 주 하이라이트

sqlelf 用 SQL 声明式探索 ELF 文件，作者提出用 SQLite 替代 ELF 作为可执行格式的激进构想。

评论区普遍欢迎降价，认为价格战利好开源模型和用户，但也有人认为降价不影响订阅用户，且模型命名混乱。

SDAD 论文将 Spec-Driven Agentic Development 形式化，主张规范质量是 agent 自主交付的执行燃料。

Terminal Agents 综述以终端执行为组织视角，建立七维终端能力画像，系统梳理命令行环境中的 LLM agent。

Kern 是一个 1.5MB 单二进制、无守护进程的容器与资源运行时，基于 OCI 镜像、cgroup v2 和 namespace，启动约 3.5ms。

FreeToken splits MoE cache misses between PCIe fills and CPU execution using measured bandwidths, unlocking frontier models locally The post Meet FreeToken: An Edge-Native MoE Serving Engine that Runs 753B GLM-5.2 on a Single Workstation GPU appeared first on MarkTechPost .

Our devs got their hands on it a few days ago. One wired it into Codex to compare with GPT Luna, our usual workhorse right now for its cost effectiveness. Another tried it out on one of our OCR pipelines. It's comparable to Luna for coding and ***OCR quality appears to be better than Gemini 3.5 Flash Lite***. That's huge. We pay a ton of money for OCR. This is the first local model that feels like more than a toy. It's truly as capable as the frontier models from a year ago. For the first time e

Harvey 发布首个后训练法律 Agent 模型 Harvey Tenet，基于 Kimi K3 用 Fireworks 异步强化学习训练，LAB 任务完成率接近翻倍。

评论区普遍认可Qwen 3.8 27B在逆向工程等任务中的实用性和持久性，但也有人认为其他模型表现更优。

Qwen 3.8 27B is great, however it takes me ages to do tasks on xhigh. I need Qwen 3.8 35B A3B. It'll be a little dumber but faster. I am also aware of the fact that 27B gets its "intelligence" from the long thinking time. I therefore assume that 35B would also be a long-thinking model, however running Qwen 3.8 27B over night on my M1 Max for just one task is impractical and no fun. I love the progress and the work of alibaba with 27B but... yeah I sadly don't own a faster RTX. What are you guys

The key skill required to make productive use of coding agents is being able to confidently instruct them on how to make changes and then confidently verify that those changes have been applied in the correct way. Sometimes this involves reviewing every line of code they have written, but there are other ways to achieve that goal. Eyeballing every line of code has never been the most effective way to validate a chance to a piece of software. Tags: code-review , coding-agents , generative-ai , ag

开发者从零训练 250M 参数 LLM，30B token，量化后仅 60MB，CPU 上约 400 tok/s。

Built by DeepMind alumni, British AI lab Inherent released Faraday, an AI agent whose ability to replicate scientific papers could be a stepping stone for innovation.

And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work. I'd like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it. I suspect those things have been trained by people who may not be quite as stubborn as I am. But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed. So credit wher

Guidelight AI Standards 研究发现，主流 AI 实验室几乎未公开 rogue model 的遏制响应计划，OpenAI 得分最高，Anthropic 与 Meta 最低。

评论区普遍认可DeepSeek新增视觉能力是重要升级，但也有人认为其图像分辨率偏低、识别精度不及竞品。

EnvHarness 通过可编程插件层动态重塑静态环境，针对 agent 弱点进行强化学习协同进化。

🤖EnvHarness and EnvRigger dynamically reshape static environments via programmable plugins to target agent weaknesses and improve reinforcement learning co-evolution.

FACET 框架在共享修复环境中保持源意图与可执行状态，为终端 agent 训练生成高质量可执行任务。

🤖FACET constructs executable terminal tasks by preserving source intent and grounding instructions, solutions, and verifiers in a shared repaired environment to enable scalable agent training.

SWE-bench Science 基准覆盖 20 个科学领域 119 个任务，揭示编码 agent 修复科学软件时的失败机制。

🤖SWE-bench Science benchmarks coding agents on scientific software repair, revealing failure mechanisms and mixed effects of scientific guidance.

Nvidia research shows that AI agents can perform well, and not go off the deep end, through fine-tuning, even if the AI model isn't that great at the task.

开发者以 250 美元预训练 10.2 亿参数 Kimi K3 复刻模型，在 50 亿 token 上已超越 GPT-2 (124M)。

用户报告 Qwen3.8-27b 在单次提示下自主执行 80 次工具调用，成功从大学网站抓取课表，展现本地模型罕见的 agent 能力。

评论区普遍认为Rust生态缺乏安全管控，需加强沙箱和依赖审计；但也有人认为应减少第三方依赖并采用更完善的开发环境。

Cryptographic Context Injection is only the latest way to break an LLM safety guardrail.

Matt Pocock tells us about his /wayfinder skill, for greenfield projects or for when the way forward is unclear.

论文《Demystifying Agent Skills》通过对照实验揭示：Skills 主要通过程序化锚定稳定执行，而非注入缺失知识，检索瓶颈与脆弱假设限制其效果。

🤖Skills enhance LLM agents primarily by stabilizing execution through procedural anchoring rather than injecting missing knowledge, though retrieval bottlenecks and brittle assumptions limit their effectiveness.

Agentic ESOpt 用进化策略替代 RL 微调长程 LLM agent，支持全参数优化且 GPU 需求极低。

🤖Agentic ESOpt uses evolution strategies for scalable full-parameter fine-tuning of long-horizon LLM agents via trajectory-level reward-weighted updates and parameter-context co-evolution.

FreeToken 是边缘原生 MoE 推理系统，将个人机器视为弹性推理平台，动态调度 CPU-GPU 与专家驻留。

🤖FreeToken is an edge-native Mixture-of-Experts serving system that dynamically maps computation and model state onto heterogeneous local hardware to run large open-weight models on personal machines.

Thoughts About Scaling Law Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions. The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — rou

用 4× RTX 3060 12GB 跑 DeepSeek V4 Flash Q4_K_XL，在 360k–376k 上下文窗口下达到约 100 tok/s 提示处理速度。

ByteDance Seed and Tsinghua AIR have released CUDA Agent, an agentic reinforcement learning system that trains a large language model to write GPU kernels that beat a compiler. The gap it targets is narrow but stubborn: frontier models already produce correct CUDA, they just produce slow CUDA. On KernelBench, the base model Seed1.6 passes 74.0% […] The post ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation appeared first on MarkTech

Long-horizon agents can fail even when their underlying models can solve the constituent steps. They may lose track of mutable state, fail to reactivate lessons from earlier executions, skip known procedures, or stop prematurely. We bet on harness scaling to improve the execution system around an agent without changing its model weights. We introduce StateM, an agent-native runtime that organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and

🤖StateM is a runtime system that improves long-horizon agent execution through durable states, recoverable runbooks, and enforceable procedural controls without altering model weights.

Secret parameter allowed hackers to steal passwords when a target clicked on a link.

Working on Mac, Linux, and Windows now. I include a simple GUI to find new models and get things built and set up. It is working quite well across a few models for me. The GitHub README and DESIGN.md files go into detail of the how/why and it's working remarkably well so far. Comments URL: Points: 38 # Comments: 6