Alignment Blog

_site/index.html

@@ -75,7 +75,7 @@ <h2 class="post-title">
     	</h2> 
         <p> I am an Applied Research Scientist at <a href="https://amazonsearchqu.github.io/" style="color:#4133ff;">Amazon Search/A9</a>, building LLMs from the scratch for Amazon (i.e. <a href="https://www.aboutamazon.com/news/retail/amazon-rufus" style="color:#4133ff;">Rufus</a>). I dropped out of <a href="https://www.cs.washington.edu/research/nlp/people" style="color:#4133ff;">UW CS PhD program (NLP track)</a> in 2021 due to some personal reasons (UW is definitely one of the best places to do NLP research). I received M.S. in Computer Science (Machine Learning Specialization) at <a href="https://www.gatech.edu/" style="color:#4133ff;">Georgia Tech</a> in 2021, where I worked with Prof. <a href="https://cs.stanford.edu/~diyiy/" style="color:#4133ff;">Diyi Yang</a> at SALT lab, currently at <a href="https://nlp.stanford.edu/people/" style="color:#4133ff;">Stanford University</a>. Before that,  I received B.S. in Biological Science and Computer Science at <a href="http://english.pku.edu.cn/" style="color:#4133ff;">Peking University</a> in 2019, where I worked with Prof. <a href="https://pku-tangent.github.io/" style="color:#4133ff;">Sujian Li</a>. In summer 2018, I worked as a research intern with Prof. <a href="https://www.aclweb.org/portal/content/bonnie-webber-receives-2020-acl-life-time-achievement-award" style="color:#4133ff;">Bonnie Webber</a> at <a href="https://www.ed.ac.uk/" style="color:#4133ff;">the University of Edinburgh</a>. I also researched as an intern at <a href="https://en.wikipedia.org/wiki/A9.com" style="color:#4133ff;">Amazon Search/A9</a> (Fall 2021), <a href="https://research.google/" style="color:#4133ff;">Google</a> (Summer 2021) and <a href="https://www.microsoft.com/en-us/research/lab/microsoft-research-asia/" style="color:#4133ff;">Microsoft</a> (Spring 2019), and conducted software development while interning at <a href="https://www.amazon.jobs/en/location/san-francisco-bay-area-ca" style="color:#4133ff;">Amazon</a> (Summer 2020). </p>
 
-        <p> Nowadays, I'm mainly working on LLMs <a href="https://jingfengyang.github.io/gpt" style="color:#4133ff;">[Blog: Reproduction and Usage of GPT3/ChatGPT]</a>, including 1) pretraining (data, infrastracture, scaling laws) 2) post-training (instruction tuning, human and AI preference learning) 3) evaluation 4) language agents (tool using, planning and reasoning, long-context handling) 5) alignment <a href="https://jingfengyang.github.io/safety" style="color:#4133ff;">[Blog: Capability and Alignment]</a> and AI safety <a href="https://jingfengyang.github.io/safety" style="color:#4133ff;">[Blog: AI Safety: Why, What, and How]</a>. The major question I'm thinking about is where is the boundary between LLM capability and alignment.</p>
+        <p> Nowadays, I'm mainly working on LLMs <a href="https://jingfengyang.github.io/gpt" style="color:#4133ff;">[Blog: Reproduction and Usage of GPT3/ChatGPT]</a>, including 1) pretraining (data, infrastracture, scaling laws) 2) post-training (instruction tuning, human and AI preference learning) 3) evaluation 4) language agents (tool using, planning and reasoning, long-context handling) 5) alignment <a href="https://jingfengyang.github.io/alignment" style="color:#4133ff;">[Blog: Capability and Alignment]</a> and AI safety <a href="https://jingfengyang.github.io/safety" style="color:#4133ff;">[Blog: AI Safety: Why, What, and How]</a>. The major question I'm thinking about is where is the boundary between LLM capability and alignment.</p>
 
         <p> I also have a broad interest in Traditional Natural Language Processing <a href="https://zhuanlan.zhihu.com/p/539706909" style="color:#4133ff;">[Blog: NLP Trends from the perspective of LLM, i.e. summary of ACL 2022 (in Chinese)]</a>  and Machine Learning. From this perspective, my research goal was to 1) solve the data scarcity problem, 2) improve the generalizability of models, and 3) design controllable and interpretable models,
             by 1) using unlabeled, out-of-domain or augmented data, 2) incorporating external knowledge or inductive biases (e.g. intermediate abstractions, model architecture biases etc.) into models, and 3) large-scale pretraining.</p>

index.html

@@ -18,7 +18,7 @@ <h2 class="post-title">
     	</h2> 
         <p> I am an Applied Research Scientist at <a href="https://amazonsearchqu.github.io/" style="color:#4133ff;">Amazon Search/A9</a>, building LLMs from the scratch for Amazon (i.e. <a href="https://www.aboutamazon.com/news/retail/amazon-rufus" style="color:#4133ff;">Rufus</a>). I dropped out of <a href="https://www.cs.washington.edu/research/nlp/people" style="color:#4133ff;">UW CS PhD program (NLP track)</a> in 2021 due to some personal reasons (UW is definitely one of the best places to do NLP research). I received M.S. in Computer Science (Machine Learning Specialization) at <a href="https://www.gatech.edu/" style="color:#4133ff;">Georgia Tech</a> in 2021, where I worked with Prof. <a href="https://cs.stanford.edu/~diyiy/" style="color:#4133ff;">Diyi Yang</a> at SALT lab, currently at <a href="https://nlp.stanford.edu/people/" style="color:#4133ff;">Stanford University</a>. Before that,  I received B.S. in Biological Science and Computer Science at <a href="http://english.pku.edu.cn/" style="color:#4133ff;">Peking University</a> in 2019, where I worked with Prof. <a href="https://pku-tangent.github.io/" style="color:#4133ff;">Sujian Li</a>. In summer 2018, I worked as a research intern with Prof. <a href="https://www.aclweb.org/portal/content/bonnie-webber-receives-2020-acl-life-time-achievement-award" style="color:#4133ff;">Bonnie Webber</a> at <a href="https://www.ed.ac.uk/" style="color:#4133ff;">the University of Edinburgh</a>. I also researched as an intern at <a href="https://en.wikipedia.org/wiki/A9.com" style="color:#4133ff;">Amazon Search/A9</a> (Fall 2021), <a href="https://research.google/" style="color:#4133ff;">Google</a> (Summer 2021) and <a href="https://www.microsoft.com/en-us/research/lab/microsoft-research-asia/" style="color:#4133ff;">Microsoft</a> (Spring 2019), and conducted software development while interning at <a href="https://www.amazon.jobs/en/location/san-francisco-bay-area-ca" style="color:#4133ff;">Amazon</a> (Summer 2020). </p>
 
-        <p> Nowadays, I'm mainly working on LLMs <a href="https://jingfengyang.github.io/gpt" style="color:#4133ff;">[Blog: Reproduction and Usage of GPT3/ChatGPT]</a>, including 1) pretraining (data, infrastracture, scaling laws) 2) post-training (instruction tuning, human and AI preference learning) 3) evaluation 4) language agents (tool using, planning and reasoning, long-context handling) 5) alignment <a href="https://jingfengyang.github.io/safety" style="color:#4133ff;">[Blog: Capability and Alignment]</a> and AI safety <a href="https://jingfengyang.github.io/safety" style="color:#4133ff;">[Blog: AI Safety: Why, What, and How]</a>. The major question I'm thinking about is where is the boundary between LLM capability and alignment.</p>
+        <p> Nowadays, I'm mainly working on LLMs <a href="https://jingfengyang.github.io/gpt" style="color:#4133ff;">[Blog: Reproduction and Usage of GPT3/ChatGPT]</a>, including 1) pretraining (data, infrastracture, scaling laws) 2) post-training (instruction tuning, human and AI preference learning) 3) evaluation 4) language agents (tool using, planning and reasoning, long-context handling) 5) alignment <a href="https://jingfengyang.github.io/alignment" style="color:#4133ff;">[Blog: Capability and Alignment]</a> and AI safety <a href="https://jingfengyang.github.io/safety" style="color:#4133ff;">[Blog: AI Safety: Why, What, and How]</a>. The major question I'm thinking about is where is the boundary between LLM capability and alignment.</p>
 
         <p> I also have a broad interest in Traditional Natural Language Processing <a href="https://zhuanlan.zhihu.com/p/539706909" style="color:#4133ff;">[Blog: NLP Trends from the perspective of LLM, i.e. summary of ACL 2022 (in Chinese)]</a>  and Machine Learning. From this perspective, my research goal was to 1) solve the data scarcity problem, 2) improve the generalizability of models, and 3) design controllable and interpretable models,
             by 1) using unlabeled, out-of-domain or augmented data, 2) incorporating external knowledge or inductive biases (e.g. intermediate abstractions, model architecture biases etc.) into models, and 3) large-scale pretraining.</p>