Shawn Wang offered $10,000 to replace a proposed $40,000 conference-software purchase. Coding agents made competing implementations cheap. His team still spent roughly two months deciding which one fitted the work.
In this Latent Space conversation, Galen Gao interviews Wang, better known as swyx, ahead of AI Engineer Shanghai. The discussion connects custom software, Jev’s narrow model decisions, and the human judgment that remains expensive.
A better screen can change the work
At 29:57, Wang explains Kill My SaaS, his competition for an open-source replacement for conference software. He reports 69 good submissions, with about five worth paying for. Entries were built in roughly a week.
One improvement addressed reviewing around 1,000 speaker proposals for 25 places. The existing service required navigating individual screens. A contestant put comparison and reordering into one table.
Wang wanted control over that workflow. Owning the code meant his team could change it without waiting for a vendor’s roadmap. He also describes connecting his own signing service to document storage so completed contracts land in the right folder.
These are buyer anecdotes. Wang acknowledges maintenance and uneven builder skill. His willingness to accept some downtime applies to this conference work, not every service a business depends on.
The AI judge answered a smaller question
At 39:31, Wang explains why the automated judge and his team preferred different entries. The judge checked qualification. The team selected a product.
He could specify basic validity before the contest. He discovered some preferences only after seeing the alternatives. Testing also meant switching between organizer, speaker, attendee, sponsor, and vendor accounts. Each role had to see the appropriate consequences of another person’s actions.
A cheap implementation can leave expensive evaluation behind it. Wang’s proposed “99designs for SaaS” would organize that work for other buyers. He discusses the idea as an opportunity, not an available service.
Jev makes the model one step in a program
At 46:45, Wang describes Jev as closer to his original vision of AI engineering. A model can choose among options or estimate a probability inside normal software. A chat window no longer needs to be the main interface.
He likes the “System One” framing: small decisions sprinkled through a program, called as casually as an if statement. He emphasizes reliability more than speed, though the interview provides no reliability measurements.
Gao mentions a podcaster using Jev to judge potential viral clips. Wang references a Doom launch demo. The first is a secondhand usage story. The second is a demonstration discussed in the conversation, not a new test of general game-playing ability.
For software builders, the design question is where a narrow decision helps. Keep its result small enough to check, and leave the surrounding program responsible for what happens next. That is an engineering implication of the discussion, not a shipped capability demonstrated here.
Taste includes choosing the problem
At 56:37, Wang argues that marketing starts before a product specification is finished. Founders need to know who will use the tool, where those people gather, and how they describe their problem.
He defines taste as narrowing many possibilities to a good choice quickly and cheaply. His examples include Cursor and Legora’s parallel document review. They are his product judgments, not comparative studies.
That connects back to the contest. Generating more versions helps only when somebody can recognize a version that improves the work.
Shanghai broadens the sample
The episode also traces Latent Space from a Stable Diffusion blog to a podcast and conference community. Wang credits local organizers for Shanghai and describes publishing talks, transcripts, and datasets so learning extends beyond attendees.
Near 1:03:29, the speakers compare enthusiasm for ComfyUI and OpenClaw in China with their US circles. Their observations suggest questions for product research, not population-wide adoption measurements.
The closing discussion covers Thinking Machines’ interruptible conversation demo, continual learning, world models, and Wang’s safety concerns. These remain separate demonstrations, research directions, and forecasts. None makes the product-selection problem disappear.
Source: Latent Space, “Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi Shanghai,” uploaded November 4, 2026, at 10:32:51 UTC. Full episode: 1:12:24. Based on complete timestamped YouTube captions through the closing thanks.