Junfan Zhu (@junfanzhu98) 's Twitter Profile
Junfan Zhu

@junfanzhu98

AI/LLM/ML Engineer, Data Scientist, Quant Finance (open to work). GaTech Computer Science, UChicago Financial Math.

ID: 1874858219635683332

linkhttps://www.linkedin.com/in/junfan-zhu/ calendar_today02-01-2025 16:40:17

6 Tweet

14 Takipçi

409 Takip Edilen

Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

DeepSeek OCR2 Visual Causal Flow 🔁DeepEncoderV2=Visual Token+LM as Vision Encoder 🕵️2 tokens=Visual+Causal(2D→1D) 🌐Native Multimodality: Different modalities🎥+🎧→same embedding. 1 encoder, different queries. Not model stitch but native fusion. 👉🏻linkedin.com/posts/jf-ai_de…

<a href="/deepseek_ai/">DeepSeek</a> OCR2 Visual Causal Flow
🔁DeepEncoderV2=Visual Token+LM as Vision Encoder
🕵️2 tokens=Visual+Causal(2D→1D)
🌐Native Multimodality: Different modalities🎥+🎧→same embedding. 1 encoder, different queries. Not model stitch but native fusion.
👉🏻linkedin.com/posts/jf-ai_de…
Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

Joe Tsai Joseph Tsai’s talk at Stanford Graduate School of Business today was a powerful reflection on building long-term institutions—from Alibaba Group values and founder culture, to multiple strategic business decisions, Qwen and the global AI race, and investment lessons from the NBA.

<a href="/joetsai1999/">Joe Tsai</a> Joseph Tsai’s talk at <a href="/StanfordGSB/">Stanford Graduate School of Business</a> today was a powerful reflection on building long-term institutions—from <a href="/AlibabaGroup/">Alibaba Group</a> values and founder culture, to multiple strategic business decisions, <a href="/Alibaba_Qwen/">Qwen</a> and the global AI race, and investment lessons from the <a href="/NBA/">NBA</a>.
Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

Conway's Game of Life↔self-organizing cognitive field in semantic space. OpenClaw🦞 local rules+communication topology→global structure via emergence moltbook agents+memory/attention→roles, subcultures, coordination, proto-privacy norms Macro-world is projection of micro-rules

Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

〽️CRATE: White Box World Model w/ First Principles Yi Ma Find low-D structure in high-D space, geometry embeds symmetry as inductive bias Self-attn=gradient steps to ↓encoding-rate MLP=sparsify residual=iterative optimize MoE=cluster Memory=encode 👉🏻linkedin.com/pulse/115-crat…

〽️CRATE: White Box World Model w/ First Principles <a href="/YiMaTweets/">Yi Ma</a>
Find low-D structure in high-D space, geometry embeds symmetry as inductive bias
Self-attn=gradient steps to ↓encoding-rate
MLP=sparsify
residual=iterative optimize
MoE=cluster
Memory=encode
👉🏻linkedin.com/pulse/115-crat…
Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

Qoder is debuting in SF—next-gen Agentic Coding Alibaba Group 🏙️Qoder Founder Silicon Valley discussion Agentic modes: 1️⃣task completion 2️⃣multi-turn collab (bottleneck=human attn/decision) 3️⃣autonomous quest paradigm 4️⃣coding freestyle on demand 👉🏻linkedin.com/posts/junfan-z…

<a href="/qoder_ai_ide/">Qoder</a> is debuting in SF—next-gen Agentic Coding <a href="/AlibabaGroup/">Alibaba Group</a>
🏙️Qoder Founder Silicon Valley discussion
Agentic modes:
1️⃣task completion
2️⃣multi-turn collab (bottleneck=human attn/decision)
3️⃣autonomous quest paradigm
4️⃣coding freestyle on demand
👉🏻linkedin.com/posts/junfan-z…
Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

🦞Google DeepMind Google AI Studio Google Antigravity Stanford University hackathon moltbook Context Overflow project⭐️ Sandbox for Agent Governance-as-a-Service, tunable params (incentives) Anti-Goodhart eval: Organizational Behavior on AI Society OS 2nd-order agency 👉🏻linkedin.com/posts/junfan-z…

🦞<a href="/GoogleDeepMind/">Google DeepMind</a> <a href="/GoogleAIStudio/">Google AI Studio</a> <a href="/antigravity/">Google Antigravity</a> <a href="/Stanford/">Stanford University</a> hackathon
<a href="/moltbook/">moltbook</a> Context Overflow project⭐️
Sandbox for Agent Governance-as-a-Service, tunable params (incentives)
Anti-Goodhart eval: Organizational Behavior on AI Society OS
2nd-order agency
👉🏻linkedin.com/posts/junfan-z…
Junfan Zhu (@junfanzhu98) 's Twitter Profile Photo

AI Cognition 🚀Latent Steering: hidden reasoning substrate in residual stream. Sparse Autoencoder=prism 🌌Multiplex Think: superposition→Ghost Vector 🤖NVIDIA FastThinkAct: 30x compression→min latent tokens. Verbalizable Latent Planning for Robotics 👉🏻linkedin.com/pulse/117-late…

AI Cognition
🚀Latent Steering: hidden reasoning substrate in residual stream. Sparse Autoencoder=prism
🌌Multiplex Think: superposition→Ghost Vector
🤖<a href="/nvidia/">NVIDIA</a> FastThinkAct: 30x compression→min latent tokens. Verbalizable Latent Planning for Robotics
👉🏻linkedin.com/pulse/117-late…