Shawn Wang on Building AIE Shanghai, the AI Engineer Role, and the Future of SaaS

The conversation opens with the claim that AIE Shanghai is the first AI Engineer conference held in China. Wang says it is run in partnership with Standard Partners, one of the local partners, and that his team does not operate every AIE event worldwide. He says he had wanted to bring AIE to China for some time, pointing to a mix of east-meets-west exchange and a body of progress in China that he feels should be shared with the global community.

The host also notes strong interest in Chinese AI on both sides of the Pacific. Wang credits a fan who translates nearly every episode into Chinese and congratulates China on progress in open-source models, robotics and startups, citing Manus as one example. The episode is set to cover OpenClaw's adoption, the startup culture around the 996 work schedule, and the origin of Latent Space.

Wang corrects the framing that he brought AI to Shanghai, saying the local partners, including Hanu Leu and Ming, made the event happen. He says he is not a Chinese citizen and had no contacts in China, so the local team was essential. He describes his background as coming from Singapore, followed by about eight years in trading, investment banking and hedge funds before he moved to tech at Amazon and several startups.

Four years ago he went independent and, he says, saw that a new AI engineer role was emerging. That observation led him to start AIE. He says he wanted to make the role a recognized industry discipline and that the simplest name for it was AI engineer.

Asked about his current roles, Wang says Cognition is his largest engagement. He worked there as an employee for about six months after the company acquired his personal startup vehicle, then stepped back to an advisory role. He explains that AIE had grown into a global conference organization that required him to act as CEO, so he could no longer hold a part-time employee position.

He lists three main vehicles: AIE, Latent Space and a miscellaneous category that used to be called Small AI. He describes his specialty as combining technical skill with writing and speaking, which he says makes him useful to startups that want someone to talk about their products. He adds that this is not a typical career and that he would not recommend it to most people.

Wang contrasts finance and tech. In finance, he says, knowledge is zero-sum: once you share something you know, its value drops, so people tend to keep information private and hide their failures. In most of tech, he argues, knowledge is positive-sum, which is why people open-source their work and share mistakes.

His first job at Two Sigma left him bored and unable to pursue what interested him, so he put his energy into writing and speaking outside work. He says learning in public also lets people choose their own boss and work for themselves, instead of being assigned a team they may later find to be poor.

Latent Space began when Stable Diffusion appeared, while Wang was at a startup he describes as not doing well. He was not an AI expert but wanted to enter the field, so he began a blog analyzing and writing up what he learned. He says the blog gained attention when Emad Mostaque, then CEO of Stability AI, read a post analyzing Stable Diffusion and called it good. Wang says that work was how he taught himself AI over several years, and that he is still not highly qualified but more so than four years earlier.

Wang says research work such as pre-training, post-training, GRPO and manifold hyperconnections, the methods associated with DeepSeek, is still viewed as prestigious. He thinks the opposite overcorrection, which treats all other AI work as worthless, is wrong.

He argues that building harnesses does not require advanced math or large-scale supercomputing experience. In his view, a good software engineer who is open to learning how to integrate AI into traditional software code can do this work. He says that engineers able to do such things were in demand even before the term harness existed, and the AI engineer label gave them a name.

Wang says the job title has not changed much, though the topics inside AI engineering have. Early work centered on RAG and context management, and more recently it has moved to coding agents, harness engineering and what he called loop engineering. He notes other disciplines such as memory, and says AIE's flagship conference organizes its content into 30 tracks.

He describes an AI engineer as the individual contributor version of the role, while AI architect is an emerging term for a company-wide position that decides how everyone in a company uses AI. He says other titles such as AI designer and AI PM also exist. For him, the test is responsibility: the person who writes and checks the code and is woken at 4 a.m. when something breaks is the engineer.

Wang describes the conference as a place where people from many labs, including Chinese labs some of which are on the entity list, can talk. He says many American companies cannot officially take meetings with these organizations and may not even respond to them, though not by choice. Being in the same venue, he says, lets people run into each other.

He presents such encounters as a source of opportunity: people get hired, find co-founders and sometimes start relationships. He describes one couple who started dating after meeting there and says they are now basically married. He jokes that his number one KPI is the number of marriages and babies he has created.

Wang says AIE aims for about 6,000 attendees this year and reported more than 7,000 plus 1.5 million online views of its talks. He stresses that every talk and workshop is published, so people who cannot travel can still watch. He says he focuses on things that work and offer clear role models.

Wang says China is the tricky case because he is a Singaporean citizen of Chinese descent and his main business is in the US, and there are many restrictions between the two countries. He says he cannot set up a local entity in China without risking trouble for his US business.

For that reason, Lei and his team own the event in China. Wang describes this as a franchise arrangement, comparing it to McDonald's and McDonald's China, and says he is happy the event can be brought to Shanghai under local leadership.

Wang says the most exposed companies are mid-tier software businesses, especially those serving SMBs. He argues that large platforms are less at risk, while smaller customers are increasingly able to code their own tools. He says that a vendor with 50 customers paying $40,000 each could lose its whole business if those customers switch to in-house tools.

He also points out that current models are good at clicking through every screen of a software product and replicating it independently. In his view, SaaS owners have to pivot, either becoming vertical agent builders in a specific field or owning specialized data and acting as data vendors. He says the right answer is not yet clear for anyone.

Wang describes the workload behind AIE Code, his next conference, which has 1,000 submissions to choose from for 25 seats. He says the software the team currently uses requires clicking through each record one by one. One contestant in the Kill My SaaS contest changed that by creating a view that lets users reorder everything in a single table.

He says that view was not available in the existing product, and that getting it from the vendor would have taken six months of requests. He presents this as evidence that a custom tool can make a team much more productive, while acknowledging that his example is simplified for audio.

Wang says he and his team spent about two months evaluating the contest entries. The conference software has five roles, namely organizer, speaker, attendee, sponsor and vendor, and all of them need to agree on how the event works. Testing required creating an account for each role, entering data and switching between accounts repeatedly.

He says that is why he pays about $40,000 for a mid-tier SaaS product, because he does not want to do that work himself. He also says that as a business owner he is doing the project mainly because it makes good content for the podcast, and that without the podcast he probably would not do it.

The hosts say Jev has prompted new ways of thinking about LLMs. Galen Gao says a podcaster in China with no coding background used Jev to judge whether clips from her show would go viral. Wang says Latent Space was the first podcast with Jev's creator, whom he had known for four years, and that the creator co-authored InstructGPT, which became ChatGPT.

Wang sees Jev as a version of his original vision of AI engineering: instead of building a large harness around a model and handing off work, the model's scope is reduced to making choices or estimating probabilities, while the software serves the people using it.

Wang describes a community he helps run that started three years ago after the group noticed many Chinese founders wanted to expand internationally. He says these founders were generally good at building or shipping products quickly, but they had a large gap in go-to-market and marketing and in understanding the market they were entering.

The community responds with meetups, camps, workshops and advisers. Wang says some advisory work is paid in equity and some is free. He says he has worked with about 35 to 36 companies, which gives him a sample from which he believes he can now see some universal patterns, although he notes that the sample size is still limited.

Wang argues that taste determines which product wins. He offers the iPhone as an example: in 1990, smartphones could do many things, but Steve Jobs insisted on a touchscreen and a flat virtualized screen. He says that required about 10 years of R&D after General Magic tried and failed to do the same thing. He says every phone now looks that way, which he considers taste.

Asked about AI products, Wang says he thinks everything Cursor does reflects taste, since it was early with VS Code and is now building Crockpot, and that competitors copying a product is a sign of taste. He also points to specialized tools such as Harvey in legal document review as examples of products that succeed in narrow areas.

Galen Gao says people in San Francisco claim to work 996 but still rest on weekends, while in China the schedule is genuinely 996. Wang responds that it is normal in startups and in big tech in China, and says he finds the pressure intense.

Wang says he works that hard because he has no job security as a founder, whereas employees cannot legally be asked to work such hours. He adds that he thinks the leading Chinese technology companies would be strong competitors in the US if they operated there.

Wang says he personally sees a possible AI winter, but he thinks the current period is probably the longest AI summer in history. He notes that AI has about 50 years of history with recurring summers and winters, and he worries about what that would mean for a conference business that depends on hype.

More broadly, he says his friends at the labs are very worried about safety, and that society may not be able to adjust quickly enough. He says this concern is not about fake news or job loss but something worse. He says the current framing is cybersecurity, while the extreme example is bioweapons. The podcast's default assumption, he says, is AI doom, because in human history a more advanced civilization has tended to displace a less advanced one, and AI is evolving roughly 1,000 times faster than humans.