Behind Latent Space: swyx and Galen Gao on AI Engineers, SaaS, and Shanghai

Shawn Wang, known as swyx, introduces a conversation ahead of AI Engineer’s first Shanghai event. He credits the local organizers and hopes to share more of the progress of Chinese AI builders with a global audience. The event is not livestreamed, but the talks are intended for YouTube.

Wang thanks a Chinese listener who regularly translates episodes and points to progress in open models, robotics and startups. He says the interview also explores differences in community scale and local adoption.

Galen Gao, a longtime Latent Space listener, asks Wang about his background and the conference. Wang corrects the premise that he personally brought it to Shanghai: local partners are leading the event, and he is grateful for their work.

Wang grew up in Singapore, moved to the United States for university and spent roughly eight years in finance before moving into technology. He later became independent, started Latent Space and helped establish the AI Engineer community.

He describes running the conference and podcast while advising startups. His personal startup vehicle was acquired by Cognition, where he worked before stepping back into an advisory role so he could focus on the growing conference organization. He says his mix of technical work, writing and speaking is unusual and not necessarily a career he recommends to everyone.

Gao asks about learning in public. Wang says that after leaving finance he found it liberating to work in a field where sharing knowledge can increase its value. In his experience finance is often more guarded, while technologists share code, discoveries and mistakes.

Wang contrasts his view of sharing in technology with his experience of finance, where knowledge can lose a competitive advantage when shared. The contrast helped him understand why writing and public work felt more natural to him.

At an earlier job he was bored and put energy into writing and speaking outside work. Learning publicly gave him more options over whom to work with. When Stable Diffusion appeared, he became fascinated by AI and began writing to learn, even though he did not consider himself an expert.

Latent Space began as a blog. Wang says persistence matters, but hard work does not itself guarantee success: judgment and work on substantive problems matter too. He sees learning in public as a way to change fields without waiting for a perfect credential.

He recalls that the Stable Diffusion posts eventually reached people working on the technology. He also cautions against treating credentials as the only route into a field: people can pivot, though they still need to develop skill and judgment.

At a hackathon, someone asked what a nontechnical person could do in AI. Wang concluded that software engineers could have a distinct role building systems around models. He wanted a name and community for people applying AI, rather than only for researchers training the models themselves.

The hackathon conversation prompted him to think about professionalization. Model training was prestigious, but he believed the engineering work of putting models into products and building harnesses deserved recognition in its own right.

He describes harness engineering as combining models with ordinary software: important work that does not necessarily require a research PhD or supercomputing background. The label AI engineer offered developers a professional identity and companies a role to hire for.

Gao describes a Chinese community that grew around Stable Diffusion and ComfyUI, with people sharing documentation and use cases. Wang says a community can form around the person someone aspires to become, not just their current activities.

The two discuss naming. Wang says people considered labels such as LLM engineer and applied AI engineer, but a title must connect with the identity and career people want. In his account, a clear role helped developers and employers find each other.

Wang thinks the job title has remained useful even as the fashionable work changes: early attention went to retrieval and context, while coding agents, harnesses, memory and evaluation have become prominent. He also sees related roles in architecture, design and product.

For him, an engineer is someone responsible when the system fails, including the person who must wake up and fix it. He wants the conference to give practitioners a platform rather than rely only on executives making broad predictions.

Wang criticizes generic conference predictions that overlook the people who build and operate systems. He borrows from developer conferences where practitioners share specific failures, decisions and lessons with other engineers.

Wang compares the AI Engineer World’s Fair to established industry gatherings such as the Game Developers Conference. Labs often work privately; an engineering conference can bring people together to explain what they built and learn from one another.

He also values chance meetings across organizations and countries. He says the event exceeded its attendance target, with more than 7,000 people in person and 1.5 million online views, and that talks and workshops are published for people who cannot attend.

He says an event can lead to hiring, cofounder introductions and friendships that are hard to predict in advance. Publishing sessions online can extend the resource to people who lack the money or ability to travel.

The team transcribes and catalogs conference talks and speaker information, making the dataset and an MCP interface available through AI Engineer’s data site. Wang hopes people can query the material and learn from the accumulated engineering experience.

On international events, Wang describes a partner model inspired by JSConf. Local organizers know their region better; they can use the brand and distribution while publishing their work and maintaining a code of conduct. Paris was the first partner event, followed by other regions.

Under the partner arrangement, organizers agree on brand use, publish their material and provide a way to address misconduct. Wang describes it as a balance between local autonomy and responsibility for a shared community.

He says sharing the brand creates risk, so choosing partners matters. For Shanghai, local leadership is especially important because of regulatory and cross-border constraints. Gao and Wang welcome the opportunity for Chinese and international engineers to exchange ideas.

The conversation moves to Kill My SaaS. Wang says a conference can cost millions to run and that software subscriptions add up to a substantial bill, even when individual employees have good reasons to request each tool.

Wang says subscription spending can appear reasonable one tool at a time but become large in aggregate. In his account, the existing products did not always match the team’s workflow, and requested features could take months to arrive.

For one proposed roughly $40,000 subscription, Wang offered a $10,000 prize to build an open-source alternative that his team could customize with coding agents. He says the contest produced 69 valid submissions, with about five he considered strong enough to pay for.

Wang argues that some midmarket SaaS products could face pressure if customers can build substitutes. He speculates that vendors may need deep domain expertise or distinctive data. These are his views on competitive risk, not a claim that every subscription can be replaced safely.

The contest was intended to produce something the team could own and adapt. Wang argues that access to coding agents changes the economics of some custom tools, while recognizing that this is an experiment based on his own organization’s needs.

His main interest is the small-business owner: a custom tool can integrate with a specific workflow and save staff time. Gao asks about maintenance. Wang acknowledges it is not free, but thinks agents and customization can make a trade-off worthwhile for some noncritical systems.

The maintenance trade-off depends on the service. Wang says he may accept downtime for a noncritical internal tool in exchange for customization; that judgment should not be generalized to systems with demanding reliability requirements.

One example was reviewing roughly 1,000 conference submissions for 25 places. The existing tool required repeated individual clicks; one contestant created a sortable table suited to the team’s workflow. Wang says he simplified the example for the audio discussion.

He stresses that entries varied considerably in quality and that a proven SaaS product can still be the better choice. A useful winner needed good human-facing design and an understanding of what conference organizers actually do.

Gao asks why the AI judge and the team selected different winners. Wang explains that the automated judge was designed to determine whether a submission qualified, not to identify the best product. The team learned what it wanted by seeing many different solutions.

Wang compares that process to a logo-design contest: many ideas can reveal options a buyer did not know to request. He imagines a service that runs similar software contests for businesses, handling specifications, evaluation and payment.

The logo contest analogy is about the diversity of candidates, not a guarantee that every entry is good. Wang says having an audience helped him attract submissions and that running such a contest requires administration most businesses may not want to do.

The work of evaluating entries was substantial. The conference software had to support organizers, speakers, attendees, sponsors and vendors, and testing meant creating accounts and moving between those roles. Wang says the experiment took about two months to evaluate alongside other work.

He describes self-hosted alternatives for scheduling, storage and other tools, including linking an open-source signature workflow to his own file system. The benefit for him is integration and control over his own workflow; he does not claim everyone should make the same choice.

For example, Wang describes signed documents flowing directly into the right folder in a system he controls. He sees opportunities in removing repeated copy-and-paste between tools, while noting that his own choices rely on having money and people to build them.

On Jev and TypeSafe, Wang describes a different approach to AI engineering: smaller, fast models make decisions or estimate probabilities within the path of ordinary code, instead of leaving all intelligence in a separate chat window. He sees reliability as important alongside speed.

Wang contrasts an AI feature embedded in many points of an application with an application that has a single chat interface. He interprets system-one models as a way to make narrow choices in ordinary program execution.

He attributes some of Diogo Almeida’s ability to communicate the idea to a nontraditional background, technical credibility and compelling demonstrations. Wang says small models were not a new idea, but the system-one framing made this approach easier to understand.

Gao describes helping Chinese founders expand internationally through community events and advice on market fit and go-to-market. Wang says technical founders can benefit from involving people who understand customers and distribution before the product is fully specified.

Gao says many technically strong founders need help understanding local demand and reaching customers. The community she describes offers workshops, meetups and advisors to help founders bridge that gap.

Wang holds writing meetups and advises companies, often informally and sometimes for equity. He values the chance to test his advice across many situations, although equity can remain illiquid for a long time and successful examples are not guaranteed.

Working with multiple companies gives Wang more examples against which to test a theory. He says the same advice does not fit every context, and the larger sample helps him recognize when a recommendation is unlikely to transfer.

Gao observes that technical founders may build impressive products while missing what customers in a particular local market need. Wang says strong technology is valuable, but marketing should not be left until the night before a launch.

He advises founders to identify their customer, where that person spends time, the words they use for their problems and how the product improves their life. For technical products, a concrete demonstration can communicate a benefit more effectively than a slide full of claims.

He suggests answering customer questions early: who will use the product, what problem do they describe, and where can the company reach them? A go-to-market partner can help build those answers into the product rather than simply promoting it at the end.

Asked to define taste, Wang suggests it is the ability to narrow a large field of possibilities quickly to choices that later seem obviously good. He uses product design examples and cites Legora’s parallel legal-document review as a workflow he finds distinctive.

Wang illustrates taste as selecting from many possible product directions and having the discipline to say no to the rest. He says a good choice can look obvious in retrospect even when it was difficult to identify at the start.

Gao asks what Wang hopes to learn in Shanghai. Wang wants to understand differences in adoption and community scale, including ComfyUI and OpenClaw. Both speakers discuss the intensity of startup work and Wang expresses admiration for Chinese founders; these are their personal impressions.

The speakers compare perceptions of work culture and AI communities in China and the United States. Their comments are anecdotal and reflect their own experiences rather than a comprehensive comparison of either country.

Wang wants more Chinese technical work to reach international audiences. He notes that language and presentation can limit the reach of a technical announcement and hopes the conference can support exchanges focused on technology.

Asked about promising research, Wang highlights interaction models that support continuous two-way communication, where a system can listen, respond and adjust without rigid turns. He is also interested in continual learning and world models.

Wang describes full-duplex interaction as a system that can listen while it speaks, accept an interruption and adapt immediately. He also sees potential in models that continue learning through interaction rather than only in distinct training phases.

His concerns range from an eventual slowdown in AI investment, which would affect a conference business, to more serious safety risks. He voices a pessimistic view of advanced AI and mentions security and biological threats; these are his risk judgments, not established predictions.

For recommendations, Wang points to his list of AI newsletters and podcasts and describes an experiment comparing model answers to common recommendation questions. Gao hopes the Shanghai event will include infrastructure, safety and payments, as well as conversations between researchers and builders across regions.

They close by wishing the local team success. Wang says that organizing the first event is often the hardest step, and both hope the exchange can continue.