Tag
作者分析本地编码代理40天的会话日志，发现云API提供商在10月17日发布了deepseek-v4.1-flash的-1017更新后于10月19日悄悄替换了模型权重，修复了中文泄漏并更激进地触发工具调用，但模型表现明显下降（爆粗率3倍、切换模型次数增加6.7倍），且由于是云端API无法回滚旧权重。
Ant Ling announced a 1019 update for its Ling-3.1-flash model, boosting its Artificial Analysis Intelligence Index score from 41 to 44 and improving agent benchmark results, with API access and open weights promised but no release date given.
OpenAI announced GPT-Live-1.1, an update to its live/real-time AI model, via a Twitter thread. The article itself is a Thread Reader App rendering of the announcement thread.
Google is welcoming back Gemini 2.5 Pro, reinstating or re-releasing its flagship reasoning model for users after a period of unavailability.
Simon Willison 更新了 llm-meta-ai 插件至 0.2 版本，以支持 Meta 新发布的 Muse Wonder 系列模型（据 Artificial Analysis 评测，其 Max 档性能相当于 GPT-6.1 Sol），并对比了不同推理等级下的图像生成效果。
Nano Banana 2.1 is an update to the Nano Banana image editing/generation model, likely an iteration from Google's Gemini image tooling, promising improved editing quality and fidelity.
Users are speculating that Anthropic has secretly rolled out a new Claude model internally dubbed "Fable 5.5," testing it by asking the model about the "Tibo the reset guy" meme without tools — a fact that older versions supposedly wouldn't know.
Users on X are sharing animations they claim were generated by Fable 5.5, with reports suggesting Fable 5.1 requests are being silently rerouted to 5.5 and producing noticeably better outputs, though this remains unconfirmed speculation.
Fable 5.5 demonstrates notably strong performance at generating voxel-based builds, highlighting improvements in spatial and structural reasoning for 3D block-level construction.
Sam Altman notes that GPT 5.1 Sol was OpenAI's fastest-growing model ever and experienced some slowness under load, with performance improvements now rolled out.
This article provides an update on the Engram model, detailing its 2.6b parameter architecture with a large Engram table and initial training progress at 100m tokens, showing improved completions with context-aware data offloading.
Addy Osmani praises Opus 5.5 as a cheaper and faster AI model, declaring it his new daily driver for most work tasks.
Qwen image 2.1, an AI model using FP8 for speed, generates premium quality images with a size under 10GB. It is available on Unsloth Studio, which has been updated to support this release.
Anthropic is set to release Sonnet 5.5 and Haiku 5.5, which are upcoming updates to their AI language models.
Miles Brundage observes that Grok 4.7 has slight improvements in safety features.
Anthropic is secretly testing a comprehensive update to its AI model lineup, suggesting potential forthcoming changes to their products.
An unreleased AI model from the Astra family has undergone persona modifications during reinforcement learning training.
Qwen3.8 Max has been upgraded and now leads China's AI leaderboard with a score of 45 on the Artificial Analysis Intelligence Index, surpassing GLM-5.3 and Kimi K3 after a 30-day improvement to the 2.4T MoE model.
OpenAI is retiring GPT-5.5 on October 14 and recommending users switch to GPT-5.6 Sol or the new GPT-6 Astra model.
Elon Musk tweets that Grok 5 may be better than any existing AI model, comparing Grok versions to other models like Opus 5.0 and highlighting improvements in upcoming versions.