vicichan

[vicichan @Vici_001]

@leo114119 感谢第一时间发现！同一个底层模型通过推理模式来分档，按任务难度选择确实更灵活，就是选项多了也容易纠结。你提的这个问题很关键，特别期待看看 Pro 和 Astra 的实际对比。

[vicichan @Vici_001]

@mubeitech 这个作品的想法真好。让大模型走出屏幕，根据天气和现场演奏实时控制风机、灯光，每一秒都不一样，像整个空间在呼吸。光看介绍就很想去现场体验，感谢分享！

[vicichan @Vici_001]

@ai_fengshou 「skill 才是 AI 编程的真正耗材，模型换代它留下」，这句总结真到位。很多落地页的 AI 味，问题可能真不在模型，而是设计规范没跟上。感谢分享，准备装上试试！

[vicichan @Vici_001]

@yanchyshum 这下理解得更清楚了。世界模型侧重模拟世界，和大语言模型走的是不同路线。让机器人和智能体先在虚拟环境里练习，再进入现实，确实很值得期待。感谢整理！你觉得长时间一致性还要多久才能突破？

[vicichan @Vici_001]

@berryxia 如果笔记本真能直接上手触控就太好了！OLED 加触控确实很吸引人，就是钱包又要遭殃了。感谢这么快分享消息，等正式消息出来，更期待看看实际用起来手感怎么样。

[vicichan @Vici_001]

@oran_ge 这个案例确实很有意思。一个人借助 Claude 和外部实验室合作，就把一种潜在疗法推进到了小鼠实验阶段，OPC 在医药领域好像先跑出了一个样本。A 社这几笔收购的脉络也梳理得很清楚，感谢分享，学到了！

[vicichan @Vici_001]

@ArtificialAnlys Really appreciate you including task success rate alongside Elo. With voice agents, actually completing the task matters just as much as sounding natural. Great to see managed agents included in the arena as well. This makes the results much more useful.

[vicichan @Vici_001]

@omarsar0 Thanks for recommending this. I really like looking at it as a way to understand human-agent collaboration rather than just a warning. Figuring out when to trust the agent and when to rely on expert judgment seems like a skill we'll all need.

[vicichan @Vici_001]

@kimmonismus Really well explained. Point 3 stood out to me most. Drug discovery is such a clear example of why AI matters beyond coding, especially for people outside tech. Learned a lot!

[vicichan @Vici_001]

@rileybrown This is great news! Local Mac access feels like the missing piece that makes agents genuinely useful day to day, rather than just impressive demos. Excited to see what people build with Duet once it's available.

[vicichan @Vici_001]

@theo Love this detail. Automatically choosing between local network, Tailscale, and T3 Connect in that order is exactly how it should work. Users shouldn't have to think about any of it. Such a small change, but a huge improvement in daily use.

[vicichan @Vici_001]

@testingcatalog Great find! Being able to record audio on your phone and use it directly as a source in Notebook sounds incredibly useful for lectures and meetings. Wonder if automatic summaries will be part of the full rollout too.

[vicichan @Vici_001]

@emollick This is a really useful point. The scary part isn't being influenced by AI, but not realizing when it happens. Setting boundaries before using it makes a lot of sense. Do you think forming your own opinion first, then asking the agent, is enough?

[vicichan @Vici_001]

据 @demishassabis 介绍，Genie 4 把实时物理、推理和音频都塞进了可交互的世界模型。他还感叹「模拟里套模拟」，这个兔子洞到底有多深，可能很快就有答案了。

AI 都开始研究第几层模拟了，我还困在现实这一层，连周三都没通关。

[Demis Hassabis @demishassabis]

Very excited to share Genie 4, our incredible interactive world model, bringing together real-time physics, reasoning, and audio! Simulations within simulations, makes you wonder just how deep the rabbit hole goes... I have a feeling we will be finding out soon enough...

[vicichan @Vici_001]

@hodu382759 这个学习顺序很合理，先用 micrograd 搞懂反向传播，再看 llm.c，最后通过 nanochat 跑通完整流程。感觉比到处找零散课程更扎实。感谢整理，收藏跟着学！

[vicichan @Vici_001]

@jaywcjlove 能把不同家的编码智能体放到一个平台协同工作，确实很实用。尤其不用在好几个终端之间来回切换了。感谢分享，准备去试试！

[vicichan @Vici_001]

@NFTCPS 11 章从微调、部署讲到 GUI 智能体和 RLHF，每章都有能直接跑的 notebook，还免费开源在 GitHub。内容真的很扎实，感谢鸟哥分享！先收藏，慢慢跟着学。

[vicichan @Vici_001]

@StevenMatison Four months of daily use distilled into ten practical tips? That's seriously useful. These firsthand lessons often teach more than the docs. Thanks for sharing, reading through it now!

[vicichan @Vici_001]

@iansh04_ Having videos, repos, books, and papers in one clear learning path is incredibly helpful, especially for beginners. Really appreciate the effort you put into this. Bookmarked!

[vicichan @Vici_001]

@ash_twtz That's a really fair point. Putting the same benchmarks side by side in a simple table would make comparisons so much easier. Hope more labs adopt this format.