Chinese contentGLM-5.3 网络安全能力接近 Claude Mythos：ExploitBench 得分 12% vs 14%
开源的 GLM-5.3 网络攻防能力快追上 Claude Mythos 了，20 美元 token 就能挖出 Chrome 漏洞，Andrew Ng 还给出了防御方的判断。
Latest
Every published item, in order · 6305 items
Oct 4
Chinese contentGemini 4 临近正式发布，GLM、DeepSeek、Kimi 一段时间没有动静
Gemini 4 快来了，而且 Anthropic 博客都还在引用 GLM-5.3 的网络能力，国产模型这条线索值得关注。
Chinese contentAnthropic 新模型定名 Fable 5.5，或选在 IPO 前发布
有人爆料 Anthropic 下一个旗舰叫 Fable 5.5，还推测会赶在 IPO 前放出来，想追 Claude 动向的可以先看看。
Chinese content爆料称 Fable 5.5 下周二发布，一条提示词生成带配乐动画
有人晒出 Fable 5.5 的实测：一条提示词直接出 3 分钟动画加原创配乐，比现在的 5.1 能打不少，等周二验证。
Models04:56
Chinese contentUtopai X 登上 Artificial Analysis 文生视频榜单第二，基于 MiniMax H3 构建一家影视工作室自己训的视频模型，基于 MiniMax H3，一上榜就冲到带音频文生视频第二，导演直接参与训练反馈，思路挺特别。
Chinese content传闻 Fable 5.5 性能将超过 Astra 6.1
传闻 Anthropic 的 Fable 5.5 快发布了，可能反超 OpenAI 因安全问题跳票的 Astra 6.1，两家竞争有看头。
Chinese contentOpus 5.5 让 Max 订阅价值回升，用户呼吁 Anthropic 不要削弱它
这位博主用 Opus 5.5 后重新买了 Anthropic 的 Max 订阅，说之前 Fable 模型时期订阅几乎没法用，现在体验完全不同。
Chinese contentWhistle：仅 16.9 MB 的离线语音转文字模型
Cactus 出了个只有 16.9MB 的离线语音转文字模型，11 毫秒启动还能标注单词时间点，做嵌入式语音的可以看看。
Chinese contentAleph Alpha 开源 Kolibri-1：德英双语 MoE 模型，78B 参数仅激活 3B
德国的 Aleph Alpha 开源了 Kolibri-1，78B 参数但只激活 3B，还支持 260K 上下文，做德语项目的可以看看。
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Ideogram 4.5 Debuts at #23 on AA-Image-Editing v2.0
Placement behind HiDream-O1-Edit-1.5 and HunyuanImage 3.0 Instruct but ahead of ByteDance's Seedream 5.0 Lite gives buyers a concrete ranking when choosing an editing model, and the US$0.008 low-tier edit price undercuts most rivals. Teams running iterative image-editing workflows can test it now via the Ideogram API and weigh open weights once released. The tiered pricing makes low-volume experimentation cheap before committing to High quality output.
Aleph Alpha releases Kolibri, a 78B open-weight model
Kolibri positions compliance as a design feature rather than a bolt-on, making it a candidate for European deployments where GDPR and AI Act alignment block US models. The Apache 2.0 license allows local hosting and fine-tuning without licensing fees, a practical option for ASEAN enterprises evaluating sovereign-AI alternatives to Mistral. Teams needing German-English bilingual capability gain a purpose-built tokenizer.
Chinese contentOpenAI 发布 GPT-6.1 Sol (Max)，Agent Arena 排名第五
OpenAI 新发的 GPT-6.1 Sol (Max) 把价格砍到 GPT-6 Astra 的两成，性能只差 1 分左右，做智能体选模型的可以看看这份榜单。
Chinese contentAgent Arena 排行榜上线：查看智能体成绩与 Pareto 前沿
LMArena 的智能体榜单更新了，能看各家 agent 的实时排名和性价比前沿，选模型前可以去对照一下。
Chinese contentClaude Sonnet 5.5 (Max) 登上 Agent Arena 第三名，Anthropic 包揽前三
Anthropic 新版 Sonnet 5.5 排名很猛，Chat 类第一，编码只差 GPT-6 两分还便宜 20%，就是单任务贵了七成。
Cohere Introduces North Small Translate Open Translation Model
Open weights mean teams can self-host translation instead of routing sensitive text through a paid API, which matters for compliance-heavy deployments. Builders localising products for Southeast Asian languages should benchmark it against Google Translate and NLLB on their own language pairs. The open process also makes evaluation methodology inspectable, not just the results.
Chinese contentCohere 用 WMT26 基准测试 North Small Translate 翻译模型
Cohere 讲了怎么给 North Small Translate 做防作弊测试：用模型做完之后才发布的 WMT26 数据，成绩没法刷，方法挺实在。
NVIDIA Releases Nemotron 3 Diarization on Hugging Face
Speaker diarization has usually required heavyweight pipelines, and a 100M-parameter model handling up to eight overlapping speakers makes transcription and meeting-analytics workloads much cheaper to run locally. Developers building voice agents, call-center tooling or subtitle generators can grab it directly from Hugging Face and integrate it into existing ASR stacks without licensing costs.
Chinese contentMeta 发布六篇论文：数学家与 Muse Spark 1.1/1.2 协作求解开放数学问题
Meta 让数学家用 Muse Spark 的普通聊天界面做开放数学题，还发了六篇论文，每篇都标注人和 AI 各写了哪部分。
Berkeley AI team wins RoCo Challenge
This victory demonstrates significant progress in robotic assembly capabilities through simulation-based approaches, potentially accelerating industrial automation development. Researchers and engineers in automation and robotics should follow this work as it may lead to new methodologies for complex manufacturing tasks, with potential applications across Southeast Asia's growing manufacturing sectors.
GPT Code Quality Issues
This critique reveals a fundamental limitation in current AI coding assistants: they optimize for immediate task completion rather than long-term software quality. Developers should be aware that AI-generated code may require significant refactoring before real-world use, especially in projects with evolving requirements. The analysis suggests that while AI can speed up initial development, it may increase technical debt if not carefully supervised.
Oct 3
Claude Sonnet 5.5 ranks third in Agent Arena
The performance-cost tradeoff reveals Anthropic's strategic positioning between premium and efficient models. Developers evaluating cost-performance ratios can now compare Sonnet 5.5's $2.74 per task against Opus 5.5's $1.58, with both leading different categories.
Chinese contentGeneralityLabs Exploit Bench测试结果
GeneralityLabs公布Exploit Bench测试结果，V4.1 Flash性能下降，Claude Code超越Codex。
DSec Software Machine Concept
This reframing of DSec as a Software Machine reveals new possibilities for distributed computing that go beyond traditional RL approaches. Developers and infrastructure managers can explore how this architecture might solve previously intractable problems by coordinating massive agent populations, potentially offering a competitive alternative to existing distributed systems.
Chinese content四大模型重塑文本、代码与智能体前沿
OpenAI和Anthropic的新模型在代码竞技场竞争激烈，Gemini 4在多领域表现亮眼，开源模型MiMo和Flash也加入智能体竞技场。
Chinese contentClaude Fable 5.5 即将发布
Claude Fable 5.5 要来了，继承了 Sonnet 和 Opus 5.5 的好评，OpenAI 的 Bel 得加把劲了。
Chinese content韩国Upstage发布Solar Mini 4模型
Upstage的Solar Mini 4虽参数量大但激活参数少，性能介于Qwen 3.6和3.8之间，专为自动化任务设计。
Only the latest 2000 items are listed; for older ones see the daily archive