Transformer诞生九年后，一批初创公司正押注LLM的下一场革命，看懂前沿风向就看这篇。
MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the …
原生多模态成K3与DeepSeek V4分水岭，看懂AI模型演进新方向。
文 | 李炤锋 编辑 | 张雨忻 “长链任务如果只通过代码层面的反馈，误差可能会不断累积，最终效果会非常差。”谈及原生多模态的意义，一位多模态研究员表示，“视觉是一种更准确的反馈，也更贴近用户意图。” 过去一年，Coding与Agent能力不断改写大模型的排名，也成为AI最快兑现商业价…
从词元到混合专家模型，一张按依赖顺序梳理的LLM术语地图，适合想系统搞懂大模型核心概念的读者。
TL;DR — A glossary to actually understand the terms you hit when reading about LLMs: token, embedding, attention, KV cache, GQA, MoE, quantization and…
Smallest.ai获投1300万美元，押注小语音模型+离线大模型双架构，打造几乎以假乱真的实时语音交互。
The startup is building voice models designed to make AI phone calls pass the Turing test.
将缓存训练融入推荐模型内部，这篇论文提出的Memory Layer有望大幅提升推荐系统效率与性能。
arXiv:2607.25110v1 Announce Type: cross Abstract: Early ranking stages in recommendation systems precompute item embeddings and cache them in-model fo…
从零开始的跨模态预训练规模扩展，揭示原生多模态模型的新范式与核心挑战。
arXiv:2607.22043v1 Announce Type: new Abstract: Although large language models (LLMs) exhibit remarkable reasoning capabilities, their reliance on tex…
提出Caption Bottleneck模型，通过瓶颈模块压缩视觉特征并生成描述，在ECCV 2026上展现多模态理解新范式。
arXiv:2607.00578v1 Announce Type: new Abstract: Concept Bottleneck Models (CBMs) provide interpretability by routing predictions through a layer of hu…
从Transformer到LLM的关键一步：堆叠层、参数量与训练细节，用同一示例直观看懂模型如何进化。
Article URL: https://bharad.dev/blog/from-transformer-to-llm Comments URL: https://news.ycombinator.com/item?id=48705350 Points: 1 # Comments: 0
探究数据混合与模型架构对非洲语言持续预训练的影响，为低资源语言建模提供前沿实证与设计指南。
arXiv:2601.06395v3 Announce Type: replace Abstract: Large language models (LLMs) are increasingly multilingual, yet open models continue to underperfo…
联邦MoE架构打破全副本部署瓶颈，让分布式模型训练更高效。
arXiv:2606.19025v1 Announce Type: new Abstract: Pre-training Large Language Models (LLMs) typically demands large-scale infrastructure with tightly co…
KDD 2026论文提出Tokenized MoE框架，用专家混合机制突破GNN泛化瓶颈
arXiv:2602.09258v2 Announce Type: replace Abstract: Deployed graph neural networks (GNNs) are frozen at deployment yet must fit clean data, generalize…
网友质疑Anthropic新模型Claude Fable 5是全新架构还是仅数据优化，引发版本命名逻辑讨论
I am trying to understand why Claude Fable 5 is different, is it a new architecture, or trained from scratch or just a better fine tuning on top of Op…
Claude与ChatGPT的"思考投入"机制是独立模型还是同参数不同效果？作者通过缓存行为差异切入，探讨大模型推理能力的实现方式。
Claude and ChatGPT have thinking efforts where you can tune the amount of thinking allowed. Like low, medium, high, xhigh and so on. But are they diff…
提出语义层级感知的层次记忆Transformer，突破长文本建模瓶颈，架构设计颇具新意。
arXiv:2605.24930v1 Announce Type: new Abstract: Transformer-based LLMs achieve strong results on many language tasks; however, long inputs remain chal…
中国工业物理AI为何领先？江行智能用99%准确率的算法和三层模型，在新能源与电网领域实现规模化落地。
中国工业物理AI的真正优势不在模型参数，而在全球12倍的工业机器人部署密度、两倍的发电量和密集的5G边缘节点——场景密度、基建底座和开源模型的合力，正在推动物理AI从实验室走向规模化落地。 江行智能提出工业物理AI的三层模型，这套系统已在新能源场站和电网巡检场景落地——覆盖贵州、内蒙古等多地，核心算…