llms

作者解释为何在自己的研究博客中从不使用 LLM 进行研究，认为 AI 掩盖了探索乐趣、无法归属贡献，并批评 AI 公司借网络安全（如 php2shell 事件）进行过度营销，指出 LLM 仍需专家引导才能发现漏洞。

A quote from Kyle Kingsbury (Jepsen) arguing that LLMs erode trust in software, documentation, and the organizations producing them, and asking how maintainers can demonstrate software quality in the age of LLMs.

An opinion essay arguing AGI is effectively already here because LLMs native operate in the language layer where all science is recorded, while brain uploading is far harder than scaling since brains depend on analog, timing-sensitive, electromagnetically coupled computation that digital hardware deliberately suppresses.

作者分享通过大声朗读自己的思考过程（类似链式思维）来提升解谜和问题解决效率的经验，并联系到 LLM 的链式思维机制与橡皮鸭调试法的认知原理。

该论文研究 Verus/Rust 形式验证工作流中，正确性反馈如何反而抑制规格完整性审计所需的证据，发现独立无反馈探索比带纠正反馈的循环产生多一倍的认证反例，并在 NeurIPS 2026 AI for Verifiable Coding 工作坊被接收。

该论文展示预训练语言模型无需任务特定训练即可零样本完成记录链接与姓名族群推断两项社会科学常规任务；作者提出 EnsembleLink 算法，在记录链接上匹敌或超越既有方法，并证明 LLM 在美国选民文件上的族群推断准确率高达 83.8%，优于 name-supplemented BISG，并可推广到宗教、种姓等其他群体类别。

作者探讨了一个思想实验：如果小型自主代理群真的在未来在互联网上长期存在，它们将面临怎样的选择压力？文章借用生物自然选择的类比，追问无控制自复制的大模型如何‘变异’，以及哪些行为特征能让它们在互联网上最成功地复制与传播。

Simon Willison quotes Geoffrey Huntley on enforcing AI coding agent behavior by using pre-commit hooks that block direct observability calls and inject prompts to make agents use his diagnostics library, achieving 'ten out of ten' results.

A blog post arguing that the claim 'AI is inevitable' is a thought-terminating cliché, exploring what inevitability means for technologies and why LLMs likely don't meet those criteria given contingent social and cultural structures.

科技博主 Gergely Orosz 指出，AI/LLM 已成为现代软件工程不可或缺的一部分，不使用 AI 的开发者将难以在初创公司和大型科技公司获得工作机会。

A software engineer's blog post criticizing the rise of LLM-generated 'personalized' recruiter emails that praise a candidate's background but are blatantly disconnected from the actual job description, making outreach feel even more impersonal than old templated spam.

A developer's personal blog post describing how using LLMs to write code helped ease a long-standing repetitive strain injury (RSI) caused by typing-heavy programming work, by reducing the need to type complex special characters and modifier keys.

In a talk at ETH Zürich, Yann LeCun argued that academics should not work on LLMs or generative models, saying there is nothing new academia can bring to the table and that real progress lies in grounded, physical AI — a stance he admits makes him unpopular in Silicon Valley.

StanceEval 2026 is the second edition of the stance detection shared task on Arabic social media, focusing on cross-target and cross-domain generalization, with top systems reaching F_avg2 scores of 0.8994 and 0.9400 and substantially outperforming baselines.

This arXiv paper shows that RoBERTa-based machine-text detectors track statistical complexity rather than machine authorship: "humanizing" LLM outputs via Mistral-7B actually raises Verb Diversity and detection scores, while the detector exhibits a 76.3% false-positive rate on formal human writing and is evaded by 41% of homoglyph attacks. Event-based latent-space detection proved fragile to paraphrasing and character-level perturbations.

Geoffrey Huntley argues that the traditional compile/restart/deploy CI/CD loop built around Docker/Kubernetes can be collapsed, and that the next shift will apply an old idea (actors, repairing images instead of rebuilding) to systems designed for LLMs as programmers rather than humans.

antirez argues that calling classifiers "decision models" misrepresents machine learning, since such models (whether small LLMs using prefill logits or BERT-like encoders) make far fewer decisions than an LLM that processes information through chain-of-thought reasoning.

A software engineer offers practical career advice to junior developers, urging them to ignore outdated ZIRP-era guidance, be pragmatic and helpful, and engage with AI tools without abandoning their own judgement in the LLM era.

报道指出当前 AI 行业的模型发布节奏极快，大约每 4.5 天就有一个主要 AI 模型推出，反映出各大厂商在开源与前沿模型竞争上的白热化态势。

该论文测量了2026年主流LLM分词器对法语及法国地方语言的token溢价（法语比英语多31%–58%，地方语言高达1.6–3.3倍），并提出法语优化的字节级BPE原型Baracoda FR v1.2，在法语上比Tekken少用11.5%的token。