⚡ Behind Latent Space: A Rare Look Inside the Pod | AIEi Shanghai

The host introduces AIE Shanghai as the first AI Engineer conference in China and frames it as an East-meets-West exchange about Chinese progress in open-source models, robotics, startups, and AI adoption. He credits Stand Up Partners, Meng Ying’s team, Hanhua, and Lei Du for making it happen, noting that the event nearly fell through several times. He also thanks Meng Shao, who regularly translates the podcast into Chinese, and says the Shanghai event will not be livestreamed but will eventually be posted on YouTube.

The guest corrects the idea that he personally brought AIE to Shanghai: Shanghai organizers did that, while he supplied the brand and support. He describes himself as a fifth-generation Singaporean Chinese who studied in the United States and spent eight years in trading, investment banking, and hedge funds before moving into Amazon and startup work. Four years before the interview, he became independent, began writing Latent Space, identified the emerging AI-engineer role, and eventually started the AIE conference organization.

Swix says tech felt more liberating than finance because finance is largely zero-sum—sharing information can reduce its value—whereas technology is usually positive-sum, rewarding people who open-source code, share mistakes, and teach others. At Two Sigma, a bad boss left him bored enough to write and speak publicly. After Stable Diffusion appeared, he started Latent Space as a blog to learn AI in public while working at Airbyte. Stability AI CEO Emad Mostaque later praised one of his analyses, helping the blog gain visibility.

Latent Space began only as “Swix writing about AI,” but hackathons and conversations led Swix to define AI engineering as a discipline. A nontechnical woman asked what she could do if she was “all in on AI,” and he concluded that engineers could control and program many AIs while nontechnical users were mostly limited to prompting. He argued that building AI harnesses requires strong software engineering and openness to integrating models into traditional code, not necessarily advanced mathematics or a machine-learning PhD.

Swix says the AI-engineer job title worked because it gave developers an identity and a path into a valuable new field. The work has shifted from retrieval-augmented generation and context management toward coding agents, harness engineering, memory, evaluations, and multimodality, but the title remains useful. He criticizes conferences dominated by executives who predict AGI in two years and warn that everyone else will lose their jobs, observing that their supposed future conveniently resembles their own jobs. AIE instead emphasizes engineers who write and maintain the code.

Swix compares AIE World’s Fair with GDC for gaming and RSA or Black Hat for security: a recurring place where practitioners learn directly from major organizations. He says the latest event drew more than 7,000 attendees and 1.5 million online viewers, with roughly 30 tracks and every talk and workshop published. The organization transcribes sessions, researches speakers and companies, and makes the material available as datasets and MCP resources. He sees the conference as a way for engineers, labs, startups, potential cofounders, and even future spouses to meet; marriages and babies are his favorite informal KPI.

The first international AIE was in Paris, whose organizers were later acquired by Mistral and now run the event from a leading European AI lab. Swix says the model resembles JSConf: local teams can use the AIE name, media channels, and community support because they understand their own markets better than a central organization could. Partners must respect the brand, publish their event material, and follow a code of conduct. China requires extra care because of US-China restrictions, so the Shanghai team owns and operates the local entity independently.

Swix explains his “Kill My SaaS” experiment: conference software costs his business roughly $300,000 per event and perhaps $1 million annually, yet remains slow and inflexible. He offered $10,000 for someone to clone a proposed $40,000 SaaS product, covered contestants’ coding-agent subscriptions, and received 69 submissions, about five of which were good enough to consider paying for. His team replaced a tedious one-by-one review workflow with a sortable table. He believes mid-tier SaaS companies—not Salesforce, Google, or AWS—must become specialized agent builders or data providers before customers build replacements themselves.

Swix distinguishes the opportunity for business owners from the problem facing developers. An owner paying ten employees $1.5 million can justify custom software if it makes them even 10% more productive, especially when coding agents make customization cheap. He has built personal versions of Calendly, Google Drive, GitHub, Dropbox, Slack, and an open-source DocuSign alternative, integrating signed documents directly into the correct data-room folder. He argues that many businesses can tolerate occasional downtime in exchange for software tailored to their workflows, although established SaaS can still be better when reliability matters.

Discussing TypeSafe founder Diogo, Swix praises the idea of small “system one” models embedded throughout software rather than one large model hidden behind a chatbot. He says models should make narrow choices or probability estimates inside ordinary code, making software smarter at many points. He defines taste as rapidly narrowing thousands of possibilities to the option that later seems obviously right, citing the iPhone’s touchscreen and Cursor’s product decisions. He is most excited by full-duplex interaction models, continual learning, and world models, but worries that an AI winter, severe security failures, bioweapons, or broader AI doom could outpace society’s ability to respond.