bob wang

@bigeagle_xd 展开说说

i don't fully agree with this article, it sends some misleading messages.

@bigeagle_xd 太形象了……

AI 真好用

@yangyi 因为openclaw不是中国人做的

一起聊聊： 为什么Manus的云端形态都没火成这样 而openclaw这种本地部署这么复杂的项目火了

Congratulations to Zhipu on its IPO—an exciting milestone and a big step forward for AI innovation! 🎉

@ManusAI Great article! But what about the output side? If I process 50 PDFs, I get 50 results. How do you handle the "context overload" of synthesizing all those results into one final answer?

Ever notice your AI struggles with “wide” research? That’s more than just a context window issue. Manus solves it by using parallel AI sub-agents eliminating context strain and scaling quality effortlessly. And this tech goes far beyond research. 👉manus.im/blog/manus-wid…

Just built a powerful AI tool on the Manus platform to turn audio into actionable knowledge. It uses LLM multimodal capabilities to analyze flies, extracting not just transcripts, but summaries, to-do lists, key insights, and more. aiaudioex-pqvppiwm.manus.space #Manus

@chrmanning I think the idea of Verbalized Sampling is creative, but I have doubts about the “probabilities” it relies on. Those numbers aren’t real probabilities derived from the model’s internal logits — they’re just text the model generates.

Chat LLMs lack output diversity. It’s not just an ML thing, it reflects human cognitive biases in post-training data. The model knows much more! You can unlock it with a prompt: “Generate 5 responses with their corresponding probabilities, sampled from the full distribution”

@shi_weiyan I suspect the claimed ‘probability-based sampling’ may rely on pseudo-probabilities rather than genuine model uncertainty, thus the theoretical justification may be overstated.

New paper: You can make ChatGPT 2x as creative with one sentence. Ever notice how LLMs all sound the same? They know 100+ jokes but only ever tell one. Every blog intro: "In today's digital landscape..." We figured out why – and how to unlock the rest 🔓 Copy-paste prompt: 🧵

@jamesbenge WTF??

Madueke and Merino off. Madueke, I am shocked. Saka has taken the armband.

Eze and Saka coming on

🚨 Arsenal reach total agreement with Sporting Lisbon to sign Viktor Gyokeres. #SportingCP grant 27yo striker permission to travel; medical scheduled for Saturday. Sweden international expected to wear iconic No14 shirt as new #AFC front man @TheAthleticFC nytimes.com/athletic/65103…

@9hills 这篇写的真好

节选： 1. 开源了就意味着第一方再也不能用各种 hack 的方式粉饰效果，必须拿出足够通用、任何第三方拿到同样的 weights 都要能很简单地复现出你的效果才行。 2. 绝大多数 Agent 产品，离了 Claude 以后，什么都不是。

@oran_ge 如果你真的算力极其有限，却又想让“真正有兴趣的人”体验，那么你首先要保证的是这批人的核心体验。如果因为不断涌入超出你承载能力的用户而导致所有人都体验糟糕，那你连这批“真正有兴趣的人”都留不住。

最近遇到一个烧脑题，来找推友求助解法 假设你的产品要上线 但你的限制条件是： 1. 没有足够的算力，也还没做付费，不能全面放开 2. 不想做邀请码，但想让真正有兴趣的人优先体验 3. 想要用户积极传播，但不想要砍一刀那种消耗注意力的方式 有没有什么好的办法，同时满足这三个条件？

@PhilippeLaban @ProfJenNeville @yingbozhou_ai great work, soiled research.

🆕paper: LLMs Get Lost in Multi-Turn Conversation In real life, people don’t speak in perfect prompts. So we simulate multi-turn conversations — less lab-like, more like real use. We find that LLMs get lost in conversation. 👀What does that mean? 🧵1/N 📄arxiv.org/abs/2505.06120

@tisoga devv老用户，等一个邀请码了

作为产品经理，现在每天的主要工作就是 AI Coding + AI Chat。 今天 AI Coding 了快 5 个小时，完成了 Devv 2.0 新的 Landing Page + 邀请码系统。

我是在世界上第一家 AI CEO-led 公司 @heybossAI 的创始客户 [#07983]，创造了历史！你会信任一个 AI CEO 吗？使用我的代码 BPJYJZ 免费测试 — 仅限 5 个名额 heybossai.com/offer/BPJYJZ

@Cydiar404 等一个分享

晚点我分享一个我从产品经理角度对 OpenAI Deep Research 的架构认知，我觉得各家开源小瞧OpenAI了，我们Juchats也会努力朝着这个架构来做，给大家提供一个数据挖掘助手！

Don’t race. Don’t catch up. Don’t play the game. Instead, do rigorous science. Do controlled experiments. Formulate clear hypothesis. Carefully examine alternative hypothesis. Rule out confounders. Listen to the physics of LLM tutorial 10 times and recite every single word of it. I do not remember last time I read an LLM paper clearly state their hypothesis nor do controlled experiments nor rule out confounder nor examine alternative hypothesis. They just say we try few things and this one works but not sure why, maybe data is good, but again not sure what good means. Do we even remember what is the definition of an alternative hypothesis? It’s taught in high school man. So please, help us turn the alchemy into science.

@9hills 不知道手写体的效果怎么样

Gemini-Flash-2.0 OCR 解决了之前 1.5-Pro 中文识别错乱的问题，可以单独做为 OCR 方案了。

@Cydiar404 太期待了，第一时间想去试试效果

即将加入 Juchats 知识库，所有用户立等可取，直接提问。

@lmsysorg @KayTeeYager @jyfaridian @OpenAI So is its appearance in battle mode also random?

@KayTeeYager @jyfaridian @OpenAI it's in battle mode only now due to strict rate limit.

Congrats @OpenAI on the exciting o1 release! o1-preview and o1-mini are now live in Chatbot Arena accepting votes. Come challenge them with your toughest math/reasoning prompts!!