Master of Data Science at University of Hong Kong, OpenToWork , AI Research intern at @ambergroup_io Agent system,VLM,trustworthy AI. e/acc iloveyou1314.eth
moon
Joined October 2018
- Tweets6K
- Following4.5K
- Followers1.1K
- Likes7.9K
Pinned Tweet
Master of Data Science student at @HKUniversity
looking for PHD or job opportunities in Hong Kong in 2027, open for Intern and RA
Agent,VLM,trustworthy AI, Knowledge Distillation related papers accepted in AAAI2026, T-IFS, ICCV2025, ECAI2024
Find me on eminentgu.top
For those who visit me because of AI, i am current a student in this area and actively seeking for phd/mphil opportunities. CV here: eminentgu.top
For those who visit me because of blockchain, I am a crypto native with 5years+ experience. $DOG holder.
Xiang.G retweeted
AI agents are already capable of managing assets and trade. What comes next? 🤖
In @tokenpost interview, @MichaelWuAmber shares what must come after smarter models:
💡 Turning intelligence into real-world outcomes
💡 Understanding user intent before taking action
💡 Crypto as the financial layer of the agent economy
Xiang.G retweeted
ONE HUNDRED AND TWELVE PERCENT
Fortytwo Engine more than doubled Qwen3.8-27B’s score on @scaling01's LisanBench
Standalone BF16: 2,452
Fortytwo Engine swarm: 5,197 (+112%)
LisanBench tests strategic planning, memory and constraint following: build the longest word chain possible using single-letter edits, dictionary words and no repeats.
Both runs: 50 starting words × 3 trials, temperature 1.0, 100k-token cap, xhigh reasoning, leaderboard scoring.
What improved:
• Repetition loops observed in the standalone run; none observed on Engine
• Higher mean scores on 40/50 starting words
• Mean link validity: 92.2% → 97.1%
• Fully valid chains: 5.6% → 22.9% of recorded answers
• Answers after retries: 72/150 → 140/150
Against the October 16 leaderboard, these scores would place 28th → 14th, putting Qwen3.8-27B on Fortytwo Engine between GPT 5.4 (medium) and DeepSeek V3.2 Speciale.
These are our runs, not official leaderboard entries.
Xiang.G retweeted
Fortytwo Engine more than doubled Qwen3.8-27B’s score on @scaling01's LisanBench
Standalone BF16: 2,452
Fortytwo Engine swarm: 5,197 (+112%)
LisanBench tests strategic planning, memory and constraint following: build the longest word chain possible using single-letter edits, dictionary words and no repeats.
Both runs: 50 starting words × 3 trials, temperature 1.0, 100k-token cap, xhigh reasoning, leaderboard scoring.
What improved:
• Repetition loops observed in the standalone run; none observed on Engine
• Higher mean scores on 40/50 starting words
• Mean link validity: 92.2% → 97.1%
• Fully valid chains: 5.6% → 22.9% of recorded answers
• Answers after retries: 72/150 → 140/150
Against the October 16 leaderboard, these scores would place 28th → 14th, putting Qwen3.8-27B on Fortytwo Engine between GPT 5.4 (medium) and DeepSeek V3.2 Speciale.
These are our runs, not official leaderboard entries.
Xiang.G retweeted
With AI agents, hallucination is an easy failure. The hard one is being wrong convincingly.
One agent drifts. Next agent builds on it. Review agents agree. Checks pass. ⚠️
In finance, that compounds. So what must run exactly as written is built by people who understand why.
Thanks to @awscloud for hosting the Web3 + AI Leadership Forum, where our Founder & CEO @MichaelWuAmber discussed deploying agents responsibly with Christopher Kung of @OpenAI, @davidsalamon of @CoinMarketCap and Yaohua Guo of @awscloud.
I built a 24*7 memecoin alpha mining bot on @hark_labs and here is the results,overall losing money definitely needs improving but it actually runs steadily in the last 15 days and found some alpha like 27x, very cool bot💪
@adcock_brett @boardyai
Xiang.G retweeted
A big moment for @NEURARobotics 👏
Congrats on the debut of 4NE1 Gen 4.
4NE1 Gen 4 is already making its mark on the world stage. 🤖
Today, we unveiled the next generation of our humanoid robot on two major stages.
First, at the opening of the TUM RoboGym powered by NEURA in Munich, built in collaboration with the Technical University of Munich and TUM MIRMI. The RoboGym is where companies teach robots real-world tasks, with partners from all kinds of industries already training there since September.
This is also where 4NE1 Gen 4 builds its experience. At its core is Neuron AI, our foundation model for Physical AI, learning from real-world robot data collected across our NEURA Gyms, as well as from human demonstrations and simulation.
And Munich was just one stop on the journey. Earlier today, 4NE1 Gen 4 was already in action at NVIDIA GTC in Berlin, where Jensen Huang left his signature on the robot. Tomorrow, the journey continues in Lugano. 🦾
Designed to work with people. Built to work in the real world. With a slimmer, lighter body engineered for series production and a soft black textile skin that gives it a more approachable appearance, 4NE1 Gen 4 brings together human-centered design and industrial ambition. It represents our next step toward humanoid robots that move beyond the showcase and into everyday working environments.
At our NEURA Gyms the new robot is already learning and performing real-world tasks every day. Physical AI is real - and as of today, one step closer to reality.
Xiang.G retweeted
4NE1 Gen 4 is already making its mark on the world stage. 🤖
Today, we unveiled the next generation of our humanoid robot on two major stages.
First, at the opening of the TUM RoboGym powered by NEURA in Munich, built in collaboration with the Technical University of Munich and TUM MIRMI. The RoboGym is where companies teach robots real-world tasks, with partners from all kinds of industries already training there since September.
This is also where 4NE1 Gen 4 builds its experience. At its core is Neuron AI, our foundation model for Physical AI, learning from real-world robot data collected across our NEURA Gyms, as well as from human demonstrations and simulation.
And Munich was just one stop on the journey. Earlier today, 4NE1 Gen 4 was already in action at NVIDIA GTC in Berlin, where Jensen Huang left his signature on the robot. Tomorrow, the journey continues in Lugano. 🦾
Designed to work with people. Built to work in the real world. With a slimmer, lighter body engineered for series production and a soft black textile skin that gives it a more approachable appearance, 4NE1 Gen 4 brings together human-centered design and industrial ambition. It represents our next step toward humanoid robots that move beyond the showcase and into everyday working environments.
At our NEURA Gyms the new robot is already learning and performing real-world tasks every day. Physical AI is real - and as of today, one step closer to reality.
Xiang.G retweeted
💬 @MichaelWuAmber at HSC Conference Singapore on how multi-agent systems drift, and why it matters most in high-stakes work where trust is everything. 🔐
Better routes to better destinations.
#HSCLifeRules - Michael Wu (@MichaelWuAmber), Founder & CEO at @ambr_io, HSC Asset Management Singapore, October 8
Once you get to multi-agent collaboration, you can get a collective lie, which is even harder to spot. It happens especially with engineering agents. If an early step drifts slightly from the original intention, it shows up in the code base, and later agents have two sources of truth: the spec and the code. The test bots are supposed to catch it, but sometimes they get confused too
When a personal agent fails a single task, you can check your account. Here you get fail-open cases: they fail, but the tests pass. You can't tell it's wrong. That's why it matters in finance and investing, where trust is so valuable
#HSCSingapore
Xiang.G retweeted
When a robot opens a lock, only sound and touch can tell it whether the job is done.
That’s why Neuron AI, our Perception-Language-Action (PLA) model, combines several senses at once. 3D images, tactile and force signals, audio, the robot’s own position and movement, and language flow into one shared picture of the situation, updated moment by moment. Based on it, the robot decides what to do next.
The second building block is memory. Humans develop object permanence in their first two years of life, the ability to remember objects when they are out of sight. Robots need it too, because they work in a physical world where things leave their field of view and every action has consequences. We adapted our architecture specifically for this.
The model gets its knowledge from three kinds of experience: our robots in NEURA Gyms, simulations that vary each scenario, and people wearing sensor-equipped suits who demonstrate the movements. After thousands of hours of base training across all five of our robot types, the model is fine-tuned on a robot and learns each specific skill from demonstration.
Through the Neuraverse, customers fine-tune Neuron AI for their own processes and deploy the results to entire fleets. This makes Time-to-Skill a central measure of progress: how quickly a robot learns a new task and performs it reliably.
Read more: neura-robotics.com/neuron-ai…
Xiang.G retweeted
We are proud to announce Neuron AI: our multimodal foundation model for Physical AI
Neuron AI connects 3D vision, touch, audio, robot state and language with memory and turns that information into action. That’s the idea behind PLA, short for Perception-Language-Action.
Vision alone is not enough. People feel whether a screw has caught and hear when a connector clicks into place. Neuron AI applies that principle to robots, so they can react to what cameras cannot capture.
It runs across five of our robot embodiments and their grippers and hands, and partners can fine-tune it for their own tasks.
We call it the brain of the physical world. Neuron AI, developed and trained in-house in Metzingen, Germany, is now available via Neuraverse. Sign up for NEURA Gym and get started today.
Read the full story: neura-robotics.com/neura-rob…
Xiang.G retweeted
When we’re doing AI efficiency work at a company, the first thing to get right is that we and our boss share the same understanding of what AI is.
youtu.be/f3-alx3LaXA
Xiang.G retweeted
Introducing Fortytwo Engine
An inference API that runs open-weight models through Swarm Inference
• 7% higher scores on average vs. the base model
• Nearly 25% fewer tokens per task
• Same per-token price and speed
• Private by default
First model: Qwen3.8-27B
Try it: engine.fortytwo.network
Xiang.G retweeted
Hark launches this week. The first 100,000 sign-ups get a paid plan for free
I don't do anything without Hark anymore. The team has obsessed over every detail, and it shows. Can't wait to see how you all use it
Join the waitlist: hark.com
Xiang.G retweeted
还有谁？！Astra + Jev 打配合，8 分 43 秒杀末影龙，总成本不到 1 美元！
要知道以往 GPT-6 Astra 自己打 Minecraft 需要 141 小时，被苦力怕清档，然后种了好几小时土豆。
这才是现在最猛的 Agent 形态——不是更大的一个模型，是两种脑一起上：
🔹Astra 当军师：异步下指令，失败就写新技能，越打越会
🔹Jev 当双手：近即时决策，直接接管走、跳、转视角、点床
🔹131 次瞬间动作 + 35 次规划调用，首落 6 床炸龙，满血出门
🔹Jev 只花 $0.01，Astra 花 $0.96。反射几乎免费，思考才花钱
🔹代码已开源。下一步是把这套行为灌进 Trajectory 做后训练
System 1 负责毫秒级手感，System 2 负责路线和学习。
机器人、电脑操控、游戏 Agent，大概都会走这条路。
#GPT6 #Astra #Jev #Minecraft #AIAgent #System1 #AGI
teapawt.org/nft_chen/status/210156…
💥牛了！将任意 LLM 秒变 Jev 模型实现本地化决策引擎！
LLM2Jev 把 Jev 同一套 Choice / Score / Noul 搬到本地：现成 HuggingFace 因果模型，不调云、不买商业 API。
关键亮点
🔹运行时定义题型，直接拿带概率的类型化答案
🔹Prefill-only，不算自回归，输出稳、好批量
🔹对接 Transformers / SGLang，适合私有数据和高并发
操作步骤
1️⃣git clone github.com/Yinsongxu/LLM2Jev…
2️⃣uv sync --extra sglang（或 transformers）
3️⃣写 JevRequest：state + Choice / Score / Noul
4️⃣LLM2Jev(...).evaluate(request) 拿概率结果
5️⃣要服务就 llm2jev-serve，打 /v1/systemone
仓库：github.com/Yinsongxu/LLM2Jev
#LLM2Jev #Jev #SystemOne #本地LLM #结构化决策 #开源AI #SGLang
Xiang.G retweeted
JEV 的能力边界慢慢开始暴露了
前几天，Browser Use 创始人 @gregpr07 开源的那个爆火 Demo（7 秒订机票、成本 $0.0039）刷爆全网
但后面他亲自用 JEV 跑了一轮长链路任务实测，发现正确率很低，远远落后于GPT luna：
• Luna：17 / 20
• Jev Agent：1 / 20
20 道复杂的现实任务，JEV 整整挂了 19 道……
那它之前为什么能做到 7 秒订机票？它到底强在哪里？
✅这是JEV擅长的：输入需求 ──> 填表 ──> 选航班 ──> 提交成功（毫无分叉，全是直道）
💥这是JEV做不到的：
【真实复杂长任务】
输入需求 ──> 弹窗拦截？ ──> 验证码？
↓ ↓
[死胡同] [死胡同]
│ │
JEV：零推理不会回溯，直接原地死锁 💥
Luna+推理模型：回退上一步（Backtrack），切换策略绕行 ✅
客观来说，JEV 确实快到不可思议，但一旦进入真实的复杂环境，逻辑就全崩了：
Browser Use 创始人说： 浏览器交互从来不是简单的元素点击，而是极其复杂的【高维状态空间搜索】
在真实网页里，一个合格的 Agent 必须具备两个核心底层行为：
停下来深思（Think hard）
走不通时原路回退（Go back / Backtrack）
而 JEV 的推理能力（Reasoning）完全为 0。
看来Jev还有很长的路要走啊~~
Browser Use 接上 JEV 这速度太他妈吓人了吧？？
过去大家看 GPT、Claude 操控浏览器，体验其实并不好，因为真的太慢了，中间还经常卡死...
因为让大模型写一堆长篇大论去操作网页，本质上就是杀鸡用牛刀
在网页上点按钮，本质上就是一个离散的多选题，根本不需要 AI 会写诗
JEV 拿到 DOM 结构，以做分类题的速度毫秒级选出该点的元素，只有在需要打字输入时才偶尔回退给小模型
黑客松冠军开发者 @gregpr07 刚刚拿 JEV 做了个开源的微型浏览器 Agent，让它自主打开网页查机票
整个过程：
• 耗时仅仅 7 秒
• 成本只要 $0.0039（折合人民币不到 3 分钱）
• 作者特意注明：演示视频是 1 倍速，没有任何加速剪辑
这个老哥还开源了GitHub： github.com/browser-use/jev-u…
Xiang.G retweeted
很多人没看懂爆火的JEV是什么？
所以我做了一个 50 秒的演示视频，带大家把海外刷屏的 JEV 以及它背后的底层架构看透👇（见下方视频）
JEV 是前 OpenAI、参与 ChatGPT / RLHF 的 @CompleteSkeptic（Diogo Almeida）创办 @typesafeai 后的第一款模型
他在 OpenAI 时发现了一个致命矛盾：
通过人类偏好微调（RLHF）训练出来的自回归大模型，存在严重的模式崩溃（Mode-dropping）、过度自信与幻觉，且串行逐字生成的延迟极大，根本无法胜任严肃的机器自动化生产
为此，他们花了两年研发了专为自动化设计的 System 1 基础模型 JEV，核心架构创新非常激进：
• 单次并行计算（One Parallel Pass）：
如同当年 Transformer 取代 RNN 一样，JEV 彻底抛弃了逐 Token 串行吐字的自回归机制，采用全新的模型架构与采样器，一次性并行击发所有预设选项
• 强化学习校准决策（RL for Calibrated Decisions）：
引入全新的训练算法，输出媒介不再是自然语言文本，而是直接输出经过置信度校准的概率分布
• 代码级可靠（TypeSafe）：
彻底根除幻觉，输出自一致、确定性高，可作为强类型对象直接被程序调用
• 100 倍提速与极端降本：
延迟降至 0.1 秒以内（快到能直接插入 60 FPS 游戏循环）；输入每 10 亿 Token 仅 42 美元，输出 Token 彻底免费
视频里我把传统自回归的 8.5 秒延迟瓶颈、JEV 的并行矩阵运算，以及实时控制小游戏实测，全拉通做了可视化, demo参考 @MatijaSosic
聊天需要慢推理，自动化需要快直觉，Building Prod, not God 确实不是一句空话🫡
发现了一个很省事的 AIGC 工作台，把 AI 漫剧的流程全部整合进去了
它的核心功能是「一键漫剧」，全程引导式推进： 输入一个故事主题，系统先给 3 个创意方向；选定后，大纲、分集剧本、角色三视图、分镜表，到最后带配音和字幕的成片，一步步带着你确认生成，不需要自己单独写剧本、做分镜或者跨软件搬运
我实测了一下：
• GPT Image 2 目前生图完全免费：没有 GPT 账号的有福了，完全可以白嫖
• 画面高清：画质细节不错，关键是我自己根本没有开会员，直接跑出来的成片效果就已经非常惊艳了，而且没有遮挡画面的大水印
• 角色跨镜头保持一致：在资产库里建好角色三视图，多镜头切换不同机位和动作，人物外貌和服装不会变形走样
不过 超大动作的复杂武打偶尔有细微瑕疵，支持单镜头局部重新生成，点一下单独重跑即可，不用整集重做
产品叫：灵镜 AIGC，我自己实测下来，觉得它很适合低门槛入门AI漫剧/AI视频，繁琐的流程它直接帮你做好了，再由人来微调，不用在几个软件之间来回搬素材
新人注册自动送 400 积分，用我邀请码再送 200 积分：
lingjing.uumit.com/?invite=E…
（邀请码：E2XE6N）
想做 AI 漫剧，不知道怎么入手的，建议大家可以来试一下
Xiang.G retweeted
卧槽！OpenAI前大佬离职创业，孵化出了过去 2 天美国最火的 AI 大模型 Jev ！它没办法生成文字，没办法生成图片，唯一做的就是一个非常简单的事情：帮你快速做判断。
比如说，决定一只股票要不要买，查询 1 个到目的地最性价比的航班，决定一个决策是否正确，并且决策速度是白毫秒级别。Jev 由 OpenAI 的大佬创业两年训练出来，押注的却是和 GPT 完全不同的 1 个方向，并且在砍掉各种文字和图片生成功能之后，获得了极致的快速度和低成本。
目前在这个官网就可以申请：typesafe.ai/，或者也可以在 vercel 上自主调用。准备把它列入 Grok bot ，帮助我在某些场景做快速的判断和打标
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution