ChinAI #195: A Handsome Panda Riding a Motorcycle
Wenxin Yige: Baidu's AI generated content platform
Greetings from a world where…
Ohtani has to be the MVP again this year, right?
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: Wenxin Yige - Baidu's AI generated content platform
Context: With the launch of models like OpenAI’s Dall-E 2 and NVIDIA’s GauGAN2, we’re seeing AI-generated images enter the mainstream. The New York Times recently profiled Dall-E 2 in a piece titled “We Need to Talk About How Good A.I. Is Getting.” Last week, at the 2022 Chinese Congress on Image and Graphics (CCIG) conference in Chengdu, Baidu released Wenxin Yige — an AI art platform that supports the generation of different styles of images. Let’s check it out, using this jiqizhixin [Synced] article (original Chinese) as a guide.
Key Takeaways: Does it look like art?
Here’s what you get when you enter the prompt: 太空中不断延伸的环形轨道, which roughly means “an ever-extending circular orbit in space”:
My favorite prompt so far is 骑摩托的帅气熊猫, which translates to “handsome panda riding a motorcycle”:
To see more AI-generated content on the Yige platform, see this page. Halfway down, there will be images sorted by popularity, or you can sort them by date of creation.
This one, generated with the prompt of “cyberpunk city in the rain,” was pretty impressive.
Art has some borders. The biggest difference between Yige and other Dall-E-like models? It’s tuned to a Chinese audience. Prompts are entered in Chinese. To apply for access to Yige, you have to have a Baidu account, which requires a Chinese phone number.
The underlying 10-billion parameter ERNIE-ViLG model was trained on a large-scale dataset of 145 million image-(Chinese) text pairs.
Baidu pitches another distinctive feature of Yige. Unlike Dall-E 2 and Google’s Disco Diffusion, which the article describes as “stylistically biased” [风格偏向性], you style your art creations in different ways (e.g., cartoon/animation-style or sci-fi) with Yige. Jeff’s note: I don’t think this is that special or technically impressive. Still, the article frames the value-add of this Yige platform as a way to commercialize large-scale AI models in a way that supports the needs of artists.
FULL TRANSLATION: Is there a bottleneck in creativity? Baidu Wenxin·Yige: It doesn't exist, just move your finger to generate a masterpiece of art
ChinAI Links (Four to Forward)
Should-attend: 2022 Politics of Emerging Technologies Virtual Mini-Conference
This upcoming Thursday, I’m co-organizing a mini-conference on the politics of emerging technologies. Baobao Zhang, an assistant professor at Syracuse University, has really led the way to put together five great panels for this conference. I’ll be the discussant for the last panel. Check out the website and register to get the Zoom link if any of the papers pique your interest!
Should-read: Web site blocking as a proxy of policy alignment
I came across this 2020 First Monday article by Nick Merrill and Steve Weber, both at the University of California Berkeley Center for Long-Term Cybersecurity, as I was prepping from syllabi this week. Their efforts to measure the Web site blocking policies of different countries produced some counterintuitive results:
“In general, our metric captures significant differences between countries typically considered to have “closed” Internets. For example, China and India stand apart from both this loose, four-country cluster. China, which is popularly imagined to “export” its model of the Internet to its Belt & Road allies (Shahbaz, 2018), in fact has low similarity with such countries. It is relatively unique, 90 percent similar to Hong Kong. Otherwise, it shares notable similarities only with Lichtenstein, for which News and Media sites comprise 19.8 percent of blocked content (compared to 57.4 percent in China).”
Should-apply — two opportunities in AI risk
I wanted to share a few open positions on AI governance:
First, The Center for a New American Security’s excellent team is launching a new initiative on AI safety and stability. They have six new positions open.
Second, Cambridge University’s Centre for the Study of Existential Risk (CSER) has a post-doc opening in AI governance. They’re open to a range of expertise relevant to areas we're working on, across long-term AI risk, foresight and governance. One of the topics they’re especially interested in is international governance and cooperation, which would build on CSER’s links to Chinese research and ethics leaders.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99
The ERNIE-ViLG samples seem worse than DALL-E 2/Imagen/Parti/Stable Diffusion. Wonder what's going on there - undertrained possibly, n=140m isn't a lot these days, especially for 10b-parameter models.