llm

Heretic 2.0 is released as a major update, transforming it from a censorship-removal tool into a general-purpose model modification framework featuring the new Arbitrary-Rank Ablation (ARA) steering method and a plugin architecture.

The tweet shares Andrej Karpathy’s two-hour hands-on AI course, which systematically explains the core workings of AI agents, harnesses, loops, and graphs. It is well suited for developers looking to gain a deeper understanding of AI agents.

AMD’s China Developer Platform has launched free API access to 15 large language models, including DeepSeek V4.1 Flash, GLM-5.3 Flash, Qwen3.8, and MiMo. No payment card is required, and quotas reset daily at midnight. Users can sign in with a phone number or GitHub account to obtain an API key and start making calls directly.

Troy Hunt's Weekly Update 528 discusses the Proton CEO's controversial AI engineering manifesto about developers' roles shifting to guiding LLMs, alongside reflections on reading AI-generated code and the economics of solar.

SWE-Trail is a history-guided trajectory collector for coding agents that uses earlier rollouts to steer where collection effort is spent without exposing that history to the agent, avoiding train-test mismatch. Trained on its trajectories, a 124B Ling-3.0-flash-base student gains 2.33 and 1.60 percentage points in resolution on SWE-bench Multilingual and SWE-bench Pro over independent sampling.

FusionCow/Sulphur-3 is a gated-access model on Hugging Face, finetuned from MiniMaxAI/MiniMax-H3, with 195 downloads last month and no inference provider support yet.

CrowdGPT v2.6 推出了重大更新，通过众包方式在普通 PC 上协作训练 10 亿参数 LLM，加入拜占庭交叉验证、加密签名与协作数据集，并在 6 天内新增训练 4.5B tokens。

Matt Dancho explains why understanding tokenization is a must-have skill for Data Scientists moving into AI Engineering, highlighting how tokens affect context windows, cost, embeddings, and multilingual performance, and links a tokenization handbook plus a free workshop.

介绍康奈尔大学与 ASAPP 的 DecisionRAG 论文：用决策模型（如 Jev 1.13）在文档树上导航取代向量相似度检索，实现 retriever-free RAG，在 Natural Questions 上 top-1 命中率达 65.4%，优于 embedding 检索加 8B reranker 的 48.0% 和 GPT-5 reranking 的 60.8%。

新发现的 Volta-Autodoxer 木马将商用 LLM 集成到信息窃取恶意软件中，能自动对受害者进行画像与优先级排序。Cisco Talos 研究人员通过其 CAIRN 框架发现了这一 AI 驱动的恶意软件变种。

Practitioners discuss 'context soup' — the ballooning of tokens and conflicting instructions when work handoffs pass full transcripts between AI agents or models — and share distilled-state handoff patterns like local ledgers, CLAUDE.md-style files, and structured receipt notes.

SelfMorph 让冻结的大模型在推理时根据输入自适应选择内部计算拓扑（原始前向、层重复、局部并行分支），通过轻量选择器统一管理这些计算形态，在多个基准上取得显著提升，提出测试时扩展的新维度。

Tsinghua University researchers propose a Bounded Dual-Loop Automated Schematic Synthesis Framework that converts natural language specifications into verified KiCad schematics for power electronics, boosting component verification coverage from 22.2% to 98.6% and SPICE pass rates from 0% to 100% versus a vanilla LLM baseline.

This paper systematically analyzes when and where socio-demographic information benefits pluralistic offensiveness modeling, showing gains are clearer at group/population levels than individual ones, and that explicit reasoning about demographics outperforms direct conditioning.

The author explores what to do when AI coding agents produce code faster than a team can comprehend it, and how five layers of governance — compiler-enforced constraints, architectural skeletons, read-only contracts, refactoring loops, and disposable implementations — can counteract system entropy and human cognitive bankruptcy to preserve software maintainability.

PostgreSQL 19 is being delayed until November because LLMs have made it far easier to discover bugs, leading to a 60-74% jump in back-patch commits and several late feature reverts as maintainers are flooded with fixes to review.

Article comparing Alibaba's Qwen 2 7B model against OpenAI's GPT-6 in terms of capabilities and performance.

Google's Project Suncatcher successfully sent a command to a prototype satellite in orbit, booted up a TPU, ran a prompt, and received a complete LLM-generated haiku response back on Earth — a milestone for running AI models in space.

The author recommended the technical reports for MAI-Thinking-1 and Qwen3.8-Flash-Next, considering them the best learning materials for understanding the details of training frontier large language models, covering topics such as data processing, pretraining, reinforcement learning, architecture choices, and training stability.

A creator is building a work-in-progress YouTube playlist/course covering LLM training and inference topics such as distributed training, fine-tuning, RL, OPD, serving, speculative decoding, and optimizers, and is asking for suggestions to make it more complete.