Leon Chen

264 posts

Leon Chen banner
Leon Chen

Leon Chen

@realleonlc

CS PhD student @Stanford. prev AI researcher @Meta @TikTok. Coached Singapore to 2/4 IOAI gold medals. AI agents & multimodal reasoning.

Stanford, CA Katılım Haziran 2017
1.9K Takip Edilen1.1K Takipçiler
Leon Chen
Leon Chen@realleonlc·
The irony of "Think Different" as a slogan: it was a mass-produced imperative, broadcast to millions, telling each of them individually that they were the rebel. Mimetic desire packaged as anti-conformity. But the deeper idea holds. A few frames I actually find useful: Rotate the question, not just the answer. Most "creative" thinking is just novel answers to the same question. The real leverage is upstream — why is this the question being asked? Who benefits from this framing? What question, if answered, would make this one irrelevant? Assume the consensus is correct, then ask what it's wrong about. Not contrarianism — the consensus is usually right about the object level. It's wrong about the implications, the timeline, or the second-order effects. That's where the edge is. Find the person who shouldn't exist. In any field, there's someone whose success violates the official theory of how that field works. They're not an anomaly — they're a signal. The model is wrong, not them. Compress until it breaks. Take the dominant narrative and compress it to one sentence. If the sentence sounds absurd or tautological, the narrative was hiding something. If it sounds obvious, ask why smart people are still debating it. Invert the prestige gradient. Whatever everyone in your peer group considers low-status but functional — that's often where the real insight lives. Status-seeking and truth-seeking point in opposite directions more often than not.
Leon Chen tweet media
English
0
0
1
113
Leon Chen
Leon Chen@realleonlc·
The best human conversationalists don't bring information. They bring surprise. We trained AI to resolve tension, not sit with it. But surprise lives in unresolved tension. That's the gap. And it's wider than it looks.
English
0
0
0
181
0x_Miko
0x_Miko@Mikocrypto11·
一个清华学生,把 Anthropic 的 AI 用成了 Polymarket 上的提款机 $1,430 → $1,550,750 而且素材里给出的数据更夸张: 44,364 笔交易 100% 胜率 单笔最大盈利 $23,600 这个账号叫 k9Q2m 按这段素材的说法,他不是靠运气,也不是靠猜 而是把 6 套对冲基金常用公式 同时塞进 bot 里,每个 tick 都跑一遍 多数人还在判断 这个 bot 直接算 它跑的 6 个核心模块是: 1)LMSR Pricing Polymarket 的价格沿对数曲线变化 bot 会提前算出自己的进场会带来多大价格冲击 比如市场给 BTC 5 分钟上涨 31¢,模型却判断这段曲线已经错价,于是先进去等修正 2)Kelly Criterion 每一笔都按最合适的仓位去下 不会大到把账户打爆,也不会小到没意义 3)EV Gap Detection 它一直在扫一个东西: 市场价格到底错了多少 比如市场给 30¢,真实概率被它算到 55¢,那 EV 就直接转正,触发进场 4)KL-Divergence BTC 5 分钟 和 15 分钟 市场本来就有关联 一旦两边漂开,它就当成套利信号 当统计距离超过 0.2,就开始标记机会 5)Bayesian Updates 新区块确认 成交量异动 价格跳动 这些新信息一进来,它就立刻更新概率 先验是 54%,新数据进来后,后验可能直接跳到 71% 6)Stoikov Execution 不是看到机会就冲 它会继续算一个更合适的执行价格 只在风险调整后仍然成立的位置成交 真正执行的时候,不是满足一个条件就下单。 而是这 6 层一起过筛: LMSR 确认错价 EV gap 超过 5% Kelly 允许仓位 Bayesian posterior 同意 KL-divergence 发现相关漂移 Stoikov 放行执行价格 只有这样,才会进场。= 也就是说,这已经不是普通意义上的“交易 bot”了 更像是一套 对冲基金框架,被搬进了 prediction market 素材最后那句其实点得很直白: 数学是公开的 edge 也是真的 真正的差别只在于: 大多数人从来没把它真正搭出来 这种把 6 个量化过滤器 同时塞进 Claude,再去跑 Polymarket 的打法,你觉得是未来的标准配置,还是只适合极少数真能把系统搭起来的人?
中文
29
118
525
123.9K
Leon Chen retweetledi
Huan
Huan@Huanusa·
🇺🇸 川普最新的一句话很多人没听懂: “也许可以用加密货币还清美国35万亿国债。” 这可不是开玩笑-听懂坐地起飞🚀🚀🚀 Crypto是美国维持美元地位的工具! 当你理解了这点,你就理解了资本运作, 理解了什么是金融,世界运转的动力! 文章中解释了还清国债如何操作,为什么黄金暴涨只是前奏,真正的大戏在加密货币👇
Huan@Huanusa

x.com/i/article/2030…

中文
28
105
565
279.3K
Leon Chen
Leon Chen@realleonlc·
Giving AI more thinking time can make it more confidently wrong. We found this building UniT: test-time scaling hits a hard ceiling — the model's own ability to verify its outputs. Past that ceiling, more compute doesn't help. It just reinforces mistakes. This new paper sidesteps that problem entirely with an external verifier trained on execution video. Smart move. The deeper question: how much external verification capability can be baked into the model itself — and where does it saturate? UniT: arxiv.org/abs/2602.12279 ExeVRM: arxiv.org/abs/2603.10178
English
0
0
2
227
Leon Chen
Leon Chen@realleonlc·
CVPR’26 J2A update: Submission deadline extended to March 6, 2026 (AOE) ⏳ The 1st Workshop on Journey to the Awards (J2A): Generative AI for Movie-Grade Video Production is calling for work on various aspects on Cinematic AI 📷📷 #CVPR2026 #J2A #VideoModels #GenerativeModels
Leon Chen tweet media
English
0
2
5
489
Leon Chen retweetledi
Leon Chen retweetledi
Felix Juefei Xu
Felix Juefei Xu@felixudr·
Just created a short cinematic promo for our CVPR 2026 workshop Journey to the Awards (J2A) using generative video. What’s remarkable is how far the tech has come — with minimal intervention and input, the first generation is already surprisingly strong on Seedance 2.0. Movie-grade video generation is no longer a distant vision. The frontier is moving fast. The future of cinema is generative. See you at CVPR 2026. cvpr26-j2a.github.io @CVPR
Felix Juefei Xu@felixudr

Video generation is entering its cinematic era. The 1st Workshop on Journey to the Awards: Generative AI for Movie-Grade Video Production (J2A) at CVPR 2026 focuses on advancing: 🎬 Movie-grade video models 🧠 Structured & controllable generation 🌍 World-model-driven video synthesis ⚙️ End-to-end production pipelines Submit your work: cvpr26-j2a.github.io #CVPR2026 #VideoModels #GenerativeModels #Multimodal @CVPR

English
2
3
15
7.9K
Leon Chen
Leon Chen@realleonlc·
🧧 Happy Chinese New Year! 新年快乐!Wishing everyone a Year of the Horse filled with strength, momentum, and bold leaps forward 🐴 ✨May this year bring breakthroughs in research, meaningful connections, and good health to you and your loved ones. 马到成功!🎆
Leon Chen tweet media
English
0
1
15
681
Leon Chen
Leon Chen@realleonlc·
Agreed with @shaneguML : honestly once "let's think step by step" worked, the rest is kind of obvious
English
0
0
2
222
Leon Chen
Leon Chen@realleonlc·
🚀 Introducing UniT: Unified Multimodal Chain-of-Thought Test-time Scaling A single unified multimodal model that reasons, generates, verifies, and refines images across multiple rounds — no external models needed at inference. Strong TTS perf. 📄 Paper arxiv.org/abs/2602.12279
Leon Chen tweet mediaLeon Chen tweet mediaLeon Chen tweet mediaLeon Chen tweet media
English
1
13
41
2.5K
Leon Chen
Leon Chen@realleonlc·
📊 Results across generation and understanding: • +10.34% compositional generation (OneIG-Bench) • +5.56% multi-object editing (CompBench) • +225% multi-turn editing (ImgEdit) • +53% OOD visual reasoning (MIRA) CoT for generation also improves comprehension! 🧠
Leon Chen tweet media
English
1
0
5
372
Leon Chen
Leon Chen@realleonlc·
🚀 Introducing our fresh work at Stanford and Meta MSL: UniT — Unified Multimodal Chain-of-Thought Test-time Scaling What if a single model could generate an image, look at it, think about what's wrong, and fix it — all by itself? That's exactly what UniT does. 🧵👇
Leon Chen tweet mediaLeon Chen tweet media
English
4
29
159
24.4K