Topics Pioneers Insight Method Author
Back

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

·Latent Space

Rendering the long image…

Topics SaaS Valuation Reset Under AI Disruption Autonomous Agents Reshaping Enterprise SaaS System-Call Models for Background Execution Startup Playbook for Vertical AI Applications Chinese Founders Driving Global AI Expansion Frontier AI Regulation and Existential Risk
Pioneers Shawn Wang

Summary

  • Shawn Wang sees coding agents shifting bargaining power from mid-tier SaaS vendors to their customers. Facing a $40,000 software purchase, he offered $10,000 for an open-source replacement; the competition attracted 69 submissions, with roughly five he would pay for, built in a week. His warning to vendors is blunt: “The moment I have access to your software, I can clone it.”
  • For business owners, customization can justify maintaining software themselves. Wang estimates conference software costs around $300,000 per event, potentially approaching $1 million annually across his business. Maintenance “is not zero,” but replacing repetitive workflows and integrating tools can outweigh occasional downtime; he nevertheless concedes that buying established SaaS is sometimes better.
  • Wang’s preferred AI architecture puts small decisions throughout software rather than intelligence inside one chat window. Diogo’s System 1 approach restricts models to choices and probability estimates that sit directly in ordinary code. “Models are for us. We don’t work for the model”: reliability matters more than the speed headline.
  • Technical differentiation needs distribution and customer understanding built into the product. Gao describes Chinese founders who ship quickly but struggle with market fit abroad; Wang argues that marketing cannot repair positioning the night before launch. His example is Diogo’s Doom demo, created by DevRel: “Show, don’t tell.”
  • AIE is turning an engineering identity into conferences, distribution, and reusable technical knowledge. Wang reports more than 7,000 attendees against a 6,000-person target, 1.5 million online viewers, and 30 tracks at its flagship event. Publishing talks, transcripts, datasets, and MCP access extends AIE’s reach beyond attendance, while local partnerships expand the brand internationally.
  • China offers a different adoption map from San Francisco’s “B2B SaaS bubble.” Gao describes a million-scale community built around shared learning, while Wang says he wants to see the differences firsthand, noting that ComfyUI has little use in his circles while OpenClaw has traction in China. He sees substantial founder capability, but the conversation also preserves disagreement about working intensity and the difficulty of presenting Chinese technical work to Western audiences.
  • Wang pairs enthusiasm for interaction models with explicit commercial and safety fears. Full-duplex communication, continual learning, and world models interest him; an AI winter threatens a conference business he acknowledges “relies on hype.” Beyond that cycle, he says lab friends worry about society adapting too slowly, and argues that the default assumption should be “AI doom”—his stated position, rather than a settled forecast.

Deep dive

1. Latent Space began as a way to earn expertise in public

  • Wang spent eight years in trading, investment banking, and hedge funds before moving into tech. A frustrating software engineering role at Two Sigma pushed his energy into writing and speaking outside work. Publishing became a way to “choose your own boss”: prospective collaborators could discover his work rather than his career depending entirely on an assigned manager.

  • His explanation for finding tech more comfortable is economic: “Finance is zero-sum.” Sharing an informational advantage can destroy its value, whereas technical knowledge often becomes more useful when others adopt it. Open source, public explanations, and admitting mistakes therefore made sense to him in a way they had not in finance.

  • Latent Space started as a blog after Stable Diffusion appeared, before ChatGPT. Wang was working at Airbyte and wanted to enter AI without claiming expertise. His analysis attracted a response from Stability AI’s Emad Mostaque—“This is good”—illustrating how substantive public learning could create access without a new degree or formal appointment.

  • Wang rejects credentialism without promising that effort guarantees success. “Just because you worked very hard, you’re not entitled to win. You have to work very hard and be good.” His own assessment remains qualified: he is “a bit more qualified” than four years earlier, with persistence and judgment determining whether the work becomes substantial.

2. AI engineering professionalizes the work around models

  • The defining encounter came at a hackathon: a nontechnical woman asked what she could do after going all in on AI. Wang’s provocative conclusion was that engineers could control “100 AIs at once,” while her immediate opportunity was better prompting. That suggested a distinct profession beyond prestigious model training and research.

  • Building harnesses, in his framing, does not require advanced mathematics or supercomputing experience; it requires strong software engineering and an understanding of how AI fits into conventional code. Alternatives such as LLM engineer and applied AI engineer circulated, but “AI engineer” offered an identity people wanted and employers could hire. “You give them a job title that pays money.”

  • Wang says the role has changed less than its fashionable specialties: RAG and context management gave way to coding agents, harnesses, memory, and other disciplines. An AI architect increasingly sets companywide practice; an engineer owns implementation and failure. The practical test is responsibility: “The person who wakes up at 4 a.m. to go and fix it.”

3. AIE combines an industry gathering with a public technical reference

  • Wang’s model is gaming’s GDC, where secretive AAA studios explain difficult engineering once a year. AI labs similarly concentrate resources and expertise, while their own developer events are selective and customer focused. A shared conference lets engineers learn across labs, including the many beyond OpenAI and Anthropic, without making famous executives the entire program.

  • His objection to executive-heavy AI events is their interchangeable forecasting: AGI is two years away, jobs disappear, and attendees must prepare for something resembling the speaker’s own role. “Conveniently, whatever is next looks like their job.” AIE instead emphasizes practitioners who can explain implementation, failures, and technical choices to peers.

  • Wang reports a 6,000-person target becoming more than 7,000 attendees, with 1.5 million watching talks online and 30 tracks at the flagship World’s Fair. Every talk and workshop is published. Transcripts, downloadable datasets, and MCP access at ai.engineer/data let users consume the material through models; researched speaker and organization profiles aim toward “the Wikipedia of AI.”

  • Latent Space also experiments with measuring model recommendations. At aeo.latent.space, Wang compares answers across roughly 200 questions spanning podcasts, databases, banks, and other categories. He says Latent Space ranks between first and third for engineering podcasts, interpreting answers as “probably a pretty decent proxy” for the preferences reflected in labs’ dataset choices.

4. Shanghai’s organizers own the local execution

  • Wang immediately corrects the premise that he brought the conference to China: “Shanghai brought AIE to Shanghai.” He credits Han Hua, Lei Du, Meng Ying, and Standard Partners. He wanted to do something in China but says he lacks the citizenship, contacts, and relationships; he sees AIE as a way to surface high-quality conversations between Chinese and US AI communities.

  • International expansion follows JSConf’s partnership model: local teams borrow the brand, YouTube channel, and distribution. Paris applied first; Wang says its organizers were subsequently acquired by Mistral. Partners agree on brand stewardship and revenue sharing, publication of their content, and a code of conduct. Sharing the name creates reputational risk, although he says a bad local event has not happened so far.

  • China adds cross-border restrictions to that model. “I’m McDonald’s, and you’re McDonald’s China” is Wang’s shorthand for the franchise relationship. He also describes conferences as places where people can encounter counterparts despite restrictions on formal meetings involving some entity-listed companies. His intended scope is explicit: “We won’t do politics; we’ll only do tech.”

5. Kill My SaaS turns a customer’s expense into a vendor’s threat

  • Wang puts conference budgets between $3 million and $10 million, with roughly $300,000 per event going to software—about 10% of the budget and margin he says he could otherwise retain. Annual SaaS spending might reach $1 million. The frustration is not just price: employees still face slow interfaces, missing features, and requests placed on a roadmap six months away. “We are an AI conference company,” he says, operating in 2026.

  • Instead of approving one $40,000 purchase, he offered $10,000 to clone the product as open source and customize it afterward with coding agents. The competition produced 69 submissions; most were usable, and approximately five were good enough that he would pay for them. The striking development window was one week, although evaluation took considerably longer.

  • Gao identifies the competitive reversal: customers themselves become competitors to their software suppliers. Wang extends it to models navigating and reproducing interfaces: “The moment I have access to your software, I can clone it.” He imagines customers abandoning a SaaS vendor within a year, while distinguishing his physical conference business from software that agents can reproduce directly.

  • Wang’s proposed escape routes remain qualified: “Nobody knows the right answer.” He thinks vendors may need to become leading agents within a vertical, citing Harvey and OpenEvidence, or own specialized data and become data suppliers. Software between those positions faces customers increasingly capable of building what they need; he is more interested in helping those business owners than protecting incumbent vendors.

6. Customization still requires maintenance, skill, and human evaluation

  • Gao’s maintenance objection gets a direct concession: “It’s not zero.” Wang’s argument concerns mid-tier tools rather than AWS, Google, or systems requiring 99.999% uptime. For a noncritical internal database, he would sometimes accept a day’s outage in exchange for customization. Coding agents help with repairs; they do not make maintenance disappear.

  • Wang frames the productivity calculation from a business owner’s perspective. A developer might see $200 as a lot, but he says he pays $1.5 million for ten employees: if they work 10% better, that is $150,000 in value; he also posits that they might become five times more productive, or $7.5 million in value. That helps explain why he is willing to spend on agents and customized tools.

  • His concrete workflow involved 1,000 submissions for 25 spots at AIE Code the following month. Existing software required opening records one by one. A contestant supplied a table where organizers could review and reorder everything together—a view the incumbent lacked. The benefit was removing repetitive interaction rather than merely recreating existing screens.

  • Wang cautions against hiring an untrained person who looks vaguely technical and expecting excellent software immediately. There was a large quality gap between average and winning submissions, especially in human usability. “Sometimes buying a tried-and-tested SaaS is going to be better.” His own organization chooses to work closer to the edge because it wants to be AI-forward.

  • The AI judge assessed valid submissions, not the team’s eventual favorite: “We didn’t know what we wanted.” Seeing roughly 70 alternatives clarified preferences. Evaluation took about two months, partly because Wang was busy; testing also required switching among organizer, speaker, attendee, sponsor, and vendor accounts to verify that their interactions worked together.

7. Integration weakens the case for buying disconnected tools

  • Wang says he now has his own versions of Calendly, Google Drive, GitHub, Dropbox, and Slack. He explicitly conditions that achievement on having money and people he can assign. The broader change is software for an “audience of one, or an audience of one company”: it can be worthwhile without becoming a product sold to everyone else.

  • His sharpest integration example is an open-source DocuSign alternative connected to his own drive: signed documents automatically reach the correct folder. Gao asks whether DocuSign’s installed network is a moat; Wang rejects that argument—“The network is the email address.” He says electronic signatures are legally enforceable and DocuSign has no patent on them. He describes the company as worth $10 billion, “or maybe seven now,” while arguing that the workflow can be reproduced.

  • Wang’s formulation is memorable: “Inside every one million copy-pastes, I think there’s a fantastic startup and product to be built.” He also sees room for a “99designs for SaaS,” handling specifications, contests, evaluation, and payment. His logo competition drew 400 ideas; he would pay someone to administer a similar process for software, rather than repeat the substantial work himself.

8. System 1 models put intelligence directly into ordinary code

  • Wang presents Diogo’s approach as closer to his original vision of AI engineering than a large harness pleading with a model for the right answer. “Models are for us. We don’t work for the model.” Narrowing the model’s job to making choices or estimating probabilities lets developers place intelligence directly in critical application logic.

  • The architectural distinction is between a conventional application with one chatbot window and software that is “at every point smarter.” System 1 models should become as casual to use as an if statement. Wang says speed is almost secondary to reliability, connecting that priority to the TypeSafe name rather than treating latency as the whole proposition.

  • He explicitly concedes that Diogo did not invent the argument for smaller models: thousands have tried “faster, cheaper, better.” What differentiated this attempt, in Wang’s account, was OpenAI credibility from work on InstructGPT, a clear System 1 versus System 2 explanation, and technical showmanship. A nontraditional background also gave Diogo sympathy for ordinary developers outside the “big models cult.”

9. Advisory work tests ideas across companies while GTM starts before launch

  • Wang remains a Cognition advisor after spending six months as an employee following his personal startup vehicle’s acquihire. Running a global AIE organization required him to step back and act as CEO. He describes equity advisory relationships as an unusual career he would not recommend broadly, rather than a standard route every technical communicator should follow.

  • Working across 36 companies gives him a larger sample for developer marketing advice than a single employer could. He can recognize where an approach has failed elsewhere instead of assuming universality. He says most people’s sample size is five, while getting to 50 could make someone one of the top people in a field. The financial caveat is equally direct: “You cannot eat equity,” and companies may remain private for years even when their paper valuations become very large.

  • Gao’s community helps Chinese founders bridge fast product development and unfamiliar overseas markets. Wang’s first caveat is that a founder needs a genuinely strong product: if they are the global number one at something, marketing can add the “画龙点睛”—the finishing touch. He warns that some founders mistakenly believe they already have that core.

  • Wang’s prescription is to involve GTM before specifications harden: decide which customers matter, where they gather, what language they use for problems, and how the product improves their lives. Reviewing a launch post the previous night cannot repair those decisions.

  • “Show, don’t tell” is the operational rule. Wang says the Doom demo associated with Diogo’s launch came from DevRel rather than Diogo himself, making it an example of GTM contributing to the product’s reception. He wants founders to demonstrate benefits directly instead of submitting a slide deck of claims.

10. Taste narrows possibilities before the right answer looks obvious

  • Wang defines taste as selecting quickly and cheaply from a large possibility space, reaching an outcome that looks obviously good in retrospect. Brute force can test 1,000 options; a tasteful person eliminates most early, sometimes without being able to explain fully why they are wrong.

  • His analogy is the touchscreen phone: many form factors were possible, but Steve Jobs committed to a flat screen and rejected alternatives. Wang cites General Magic’s failed attempt and a roughly ten-year R&D journey. The point is sustained commitment to a product shape before hindsight makes it seem inevitable.

  • For AI products, Wang points to Cursor’s choices and says copying is evidence: “You know you have taste when other people copy what you just shipped.” His legal-software example is Legora, with its massively parallel document-review process. The resulting user reaction—“I don’t want to do my work anywhere else”—is the test he emphasizes.

11. China’s adoption patterns challenge San Francisco’s default assumptions

  • Gao traces WaytoAGI to Stable Diffusion enthusiasts, shared learning documents on Lark/Feishu, and ComfyUI workflows. He describes the community as having around one million users, then says several million users are learning together; he calls it China’s largest AI community. Wang characterizes a million-person community as medium-sized, and Gao corrects him: a million is already top-tier in China.

  • Wang wants to see those differences firsthand. San Francisco, he says, has a “B2B SaaS bubble”; his circles favor Canva and Fal.ai rather than ComfyUI. OpenClaw has more traction in China than ComfyUI, but Wang says he has not personally experienced the Chinese adoption energy people describe to him.

  • Gao pushes back when Wang contrasts China’s working intensity with San Francisco: he has heard American founders and engineers work similarly hard. Wang distinguishes founders with substantial equity from employees and maintains that even highly competitive US startups work less intensely. His admiration is categorical but conditional: leading Chinese founders, if based in America, “would kill Amazon” and Facebook.

  • Distribution also filters what Western audiences notice. Wang says Alibaba’s Qwen roadmap was buried in a broader corporate presentation that his audience did not see, leaving him to translate it. AIE could present such work directly in a familiar format. Gao hopes Shanghai also surfaces agent infrastructure—particularly safety and payments—that he finds less discussed in China.

12. Interaction models excite Wang while winter and safety remain material fears

  • Beyond System 1 models, Wang highlights interaction models associated with Thinking Machines Lab: full-duplex communication in which both sides listen while speaking, interrupt, and adjust immediately. He connects that to continual learning during use rather than a finished pretraining/post-training cycle. World models complete his list of areas that currently interest him.

  • His immediate business fear is an AI winter. He calls the present period possibly the longest AI summer in human history, while noting that AI has been around for about 50 years and that optimism has repeatedly reversed. For a conference company that “relies on hype,” the question becomes personal: “When is it going to happen, and am I going to go bankrupt?”

  • The broader concern is society failing to adjust fast enough. Wang says friends at labs are really worried about risks beyond misinformation and lost employment. Cybersecurity is today’s visible expression; bioweapons are an extreme example, framed as accidentally inventing “COVID-2.” He worries there may be something worse that people do not yet understand.

  • Wang’s strongest position is that “the default assumption for most people should be AI doom.” He invokes encounters between more and less advanced civilizations and says AI is evolving roughly 1,000 times faster than humans. That is the analogy underlying his concern: capabilities could advance beyond society’s ability to determine how to respond.