Wujiang Xu
98 posts
Wujiang Xu
@wujiang_ai
Agentic Memory | Second-year PhD student @RutgersCS , Ex Research Intern @Meta | MS @xjtuofficial | Previous Full-time Senior MLE @AntGroup
Katılım Kasım 2020
269 Takip Edilen93 Takipçiler
@wujiang_ai @fnruji316625 Congratulations! Would you be willing to share your experiences or any referral opportunities? I would really love to join a team working on multimodal agents/models next summer!
English
Excited to join Meta MSL this summer as a Research Scientist Intern, working on multimodal agents!
Before this, I’ve been exploring LLM agents from several angles: prompting-based memory systems, training agentic RL in long-horizon tasks, and evaluating memory in long-horizon agent tasks.
If you’re in the Bay Area and working on related research or cool agent projects, I'd be happy to connect and chat!
English
English
English
I asked GPT to draw a clean NeurIPS/ICLR-style ML framework diagram with an OS environment, task pool, verifier, checklist, actor, and optimization loop.
It gave me a random beautiful woman taking a selfie.
Not even close. Not “slightly misunderstood.” Just fully left the research pipeline and entered the Instagram explore page. #gpt
English
最近我把和 Claude Code 所有的聊天切到英文交流了，开始会有些蹩脚，但是发现后面越用越舒服，大多数 AI 模型底层的英文训练数据远多余其他语言，我不想经过一层看不见的翻译，加上我想着提高自己的英文输出和说的能力，想立一个 flag 有机会做一次全英文的技术大会分享，应该会比较有意思。
于是我给 Waza 加了一个能力叫做 English Coaching，你尽管输入你的蹩脚英语，他会在任务执行完毕后，输出一个 😇 然后给你纠正你的问题，是语法错误还是单词或者是不够地道，这样你会使用起来越来越熟练，相比每天打卡学习多邻国，我更喜欢这个寓教于乐的感觉，假如你也需要，可以去看看 Waza，一条命令直接安装即可。
github.com/tw93/waza
中文
越来越觉得做research就是做菜。。。
不是说做米其林三星。而是你在厨房里（实验室），客人一小时后到（deadline），看手头食材，赶紧想办法做一桌菜。
为什么有人做研究会失败？
1. 有人不洗锅就开始做（没有跑通干净的base pipeline）
2. 有人做了一桌好菜但没时间装盘（结果很强，paper写得很烂）
3. 有人发现自己有水就开始做菜了（只跑最基础的baseline）
4. 有人没打开冰箱找食材（不看related work）
5. 有人拿错了食材（跟了错误的/不重要的paper）
6. 有人想用面粉做分子料理——给你一个月也许能做出来，但你会先饿死（scope远超可行范围）
7. 有人菜还没熟就掀锅盖捞出来了（不做analysis和ablation）
8. 有人追求好吃，有人追求好看（好数字 vs. 好framing）
9. 有人非要复刻小红书上的菜，但自己家没空气炸锅（没有对应的算力和数据就硬复现别人的方法）
10. 有人把所有时间花在前菜上，主菜来不及做了（死磕一个实验，忽略了整篇paper的story）
……
做菜不是只有好看, 最先满足的应该是能吃。缺点小东西可以下楼去买，但你没法坐一小时地铁去大商场。可以做的很惊艳，粗茶淡饭也很好。
学会做菜，就是做research最重要的能力。
中文
How do you train AI agents that can use computers like humans?
Introducing OSGym: Scalable OS Infrastructure for Computer Use Agents. It’s large-scale training made possible by extensive infra optimization:
• 1024 OS sandboxes running in parallel
• $0.23 / sandbox / day (90% cost down)
• 1420 trajectories per minute
• 37× faster disk provisioning
• 88% less physical disk usage
• RAM-bound orchestration (10× scalable than CPU-bound)
🧵
English
@trq212 got the email, but UI doesn't give me the ability to redeem. cross posting from Boris' announcement in case you can get someone to take a look:
x.com/chungusfarmer/…
chungus@chungusfarmer
@bcherny I don't see the ability to redeem any credits. Support bot is lacking. I'd give you my email but not trying to share with entire thread: I have a session ID you can link to my user: 530e36d3-6688-4dfd-b3db-1cc502f34d12
English
claim a month of free credits on us, thanks for bearing with us
Boris Cherny@bcherny
Subscribers get a one-time credit equal to your monthly plan cost. If you need more, you can now buy discounted usage bundles. To request a full refund, look for a link in your email tomorrow. support.claude.com/en/articles/13…
English
@0xcherry 赞，有没有hyperlink to code🙏，学习下。对了，btw除了高频交易，会不会觉得中长线的更合适金融类agent分析，我理解是有些专业的数据我们拿到延迟是比较高的。
中文
Big respect to Anthropic.
They saw us opening 4 Claude tabs like degenerates and said: “Nope, too overwhelming, go touch grass.” Truly the only AI company that cares about our well-being 😂
Thank you for the intervention.
Lydia Hallie ✨@lydiahallie
Thank you to everyone who spent time sending us feedback and reports. We've investigated and we're sorry this has been a bad experience. Here's what we found:
English
RAGEN achieves 200 citations within a year!
There's still a long way to go in agentic RL 🤗
Zihan "Zenus" Wang@wzenus
In Agent RL, models suffer from Template Collapse. They generate vast, diverse outputs (High Entropy) that lose all meaningful connection to the input prompt (Low Mutual Information). In other words, agent learn different ways to say nothing. 🚀 Introducing RAGEN-v2 -- Here's how we define and fix such silent failure modes in Agent RL. 🧵
English
@LegalVoting It's a bug. All they have to do is downgrade everyone while they fix but they don't
English
Anthropic is committing suicide right now.
Everyone is exhausting their AI credits in hours. I just got back from vacation and exhausted nearly all my credits in 3 hours.
This bug has been up for over 3 days and no fix has been issued. No AI credits refunded.
reddit.com/r/Anthropic/co…
github.com/anthropics/cla…
English
@beihuo 我觉得区别在于cc会倾向于读更多的文件 干任何事情前，而codex则会略读几个就给出答案，在project比较大的过程中很容易有问题。
中文
是的，最近因为额度用满，尝试了 codex，结果发现和 claude 差异巨大。可以说 codex 5.4 xhight 在 claude opus 4.6 面前像是个废物。
宝玉@dotey
恰恰相反，codex cli 做的很难用，不如 claude code 代码有价值。 但如果你只是要做一个 ai agent，没必要用 claude code 或者 codex 代码，pi-mono 这种轻量级的库更适合你。
中文
PRBench: End-to-end Paper Reproduction in Physics Research
- Presents a benchmark of 30 expert-curated tasks spanning 11 subfields of physics
- All agents exhibit a zero end-to-end callback success rate
English
@bcherny Hey Boris, as a Max plan user I wanted to flag this: I sent a single message — “how’s going” — and it used up 8% of my session limit. The token consumption feels extremely aggressive. Any plans to optimize this?
English