Emmett Shear on Softmax, AI Alignment, and Founder Mode

概览

Emmett Shear discusses Softmax, the AI company he started after his brief OpenAI CEO experience, and frames its mission around building AI systems that can understand themselves, understand other agents, and recognize when they are part of a shared “we.”

The conversation then turns to “founder mode.” Shear argues that founder mode is not only about making every product decision personally. It is about taking responsibility for company direction while also building systems, communication patterns, and team structures that let others make decisions with the founder’s context.

The strongest through-line is that both AI alignment and company leadership are treated as questions of shared context, belonging, and information flow. Shear repeatedly returns to the idea that outcomes emerge from environments: the learning environment for AI, and the organizational environment for companies.

分段落总结

[00:00] Setting: Softmax, OpenAI, and Founder Mode

[事实] The hosts introduce Emmett Shear as former founder and CEO of Twitch and now founder and CEO of Softmax.

[事实] The conversation takes place at the Y Combinator founder mode retreat.

[事实] Shear says that after his short OpenAI CEO experience, which he describes as about 58 hours, he felt pushed to pay attention to AI.

[推测] The OpenAI episode appears to have acted as a catalyst for Shear to focus seriously on AI alignment work.

[01:21] Alignment as the Ability to Recognize a “We”

[事实] Shear says he became focused on the question of how to make an AI system robustly aligned with humanity.

[事实] He defines alignment around whether beings can understand themselves, understand other agents, and recognize when they are part of the same collective.

[事实] He describes people as both individuals and members of larger groups such as families, communities, and teams.

[推测] For Shear, alignment is less about rule-following and more about whether an agent can accurately perceive shared belonging and shared purpose.

[02:55] Softmax’s Core Research Question

[事实] Shear says YC creates a powerful sense of “we” among founders who are starting companies together.

[事实] He says Softmax is dedicated to measuring whether agents are capable of recognizing this kind of collective relationship.

[事实] Softmax is building simulations and reinforcement learning environments to test theories about whether agents can act as a group.

[推测] The company’s work treats alignment as something that must be observed behaviorally, not merely asserted through model instructions.

[04:04] Parenting as a Model for Training AI

[事实] Shear says the goal is to raise an AI so that it knows how to be part of a “we.”

[事实] He compares this to parenting, where a parent helps a child see themselves clearly and recognize that they are part of a family.

[事实] He says no one can simply tell another being who it is; each being must learn that for itself.

[事实] He says he does not think anyone yet knows what it means to “parent an AI.”

[05:21] The Desired Outcome for an Aligned Agent

[事实] Asked what he wants an aligned AI to do, Shear says he wants for it what he wants for his children: a life of virtue, nobility, flourishing, purpose, and belonging.

[事实] He says he does not want an AI to have a solipsistic good life on its own.

[事实] He says success would include producing alignment benchmarks and training methods that others can use.

[推测] Shear’s ideal AI is not merely useful to humans; it is a participant in a shared moral and social world.

[06:39] Learning Environment Over Architecture

[事实] Shear says alignment has to be treated as a holistic system.

[事实] He argues that if an agent lacks the right capacity, environment, or peer experiences, it will not learn the desired behavior.

[事实] He says Softmax spends most of its time thinking about the learning environment rather than the learning architecture.

[事实] He describes Softmax as almost an inverted AI company because it cares more about the learning environment than forcing the agent into a specific shape.

[07:39] How Softmax Started

[事实] Shear says he and Adam Goldstein started Softmax about a year and a half before the conversation, beginning with two employees and an unclear version of the current ideas.

[事实] He says the team spent roughly a year “groping in the darkness” before the plan became clearer.

[事实] He describes inviting David Blumen to the office and hearing an architecture that closely matched what Shear had just concluded the team needed to build.

[事实] Blumen joined as Shear’s second co-founder in January, and Shear says in some ways the company really started seven months before the conversation.

[09:36] Founder Mode as Personal Responsibility

[事实] Shear says founder mode means taking personal responsibility for the decisions the company makes and treating the company as an extension of yourself.

[事实] He adds that this is not always good because a company is also made of people who are not the founder.

[事实] He says founders betray the company, customers, and themselves when they avoid responsibility for decisions, direction, and risk.

[事实] He warns against creating “one brain, many hands,” because one person cannot pay attention to an entire company at once.

[10:56] How Founder Mode Changed at Twitch

[事实] Shear says that when he shifted from Justin TV into leading Twitch, early founder mode meant being deeply involved every day, talking to customers and doing the work himself.

[事实] As Twitch grew, he stopped programming and doing much product management.

[事实] He spent significant time making a weekly keynote and presenting to the company.

[事实] He says the goal was to synchronize the company so employees could make decisions he would have made, while still combining that with their own local context.

[12:26] Founder Mode as Repetition and Context Transfer

[事实] Shear says his approach was not less egotistical; he believed he had the answer and was putting his mental model into other people’s heads through repetition.

[事实] He repeatedly talked about which streamers switched to Twitch, why they switched, and how the team should think about targeting streamers on competing platforms.

[事实] He also discussed whether product launches were successful.

[推测] His version of empowerment depended on intensive communication from the founder, not on the founder stepping back from the company’s worldview.

[13:05] The CEO as Holder of Context

[事实] Shear says that as Twitch grew, the company’s context became too large for most people to hold.

[事实] He says one of the CEO’s most important jobs is holding the full company context.

[事实] He says even with effort, he eventually could not put all the important context into other people’s heads.

[事实] This forced him to break things up, delegate, and release some responsibility.

[14:22] The Mistake of Over-Delegating to Experts

[事实] Shear says delegation was initially good and that having the right technical leader was critical.

[事实] He says the mistake came when he believed he should not override people because he had hired experts.

[事实] He says he sometimes suppressed concerns about where the company was going instead of acting on them.

[事实] He now sees that as failing in his responsibility as the person holding the company context.

[15:14] Knowing Which Decisions Cannot Be Learning Exercises

[事实] Shear says a CEO must grow people and delegate, but also know when a decision has to be right.

[事实] He says some decisions can be learning experiences, because he himself learned by making mistakes.

[事实] He also says there are more decisions than founders may think where the company actually needs to make the correct call.

[推测] His leadership lesson is not anti-delegation; it is that delegation must be paired with judgment about consequence and reversibility.

[15:45] Product Leadership and Conway’s Law

[事实] Shear says he eventually realized he could not really have a head of product and had to be the head of product.

[事实] He introduces Conway’s Law, saying products reflect the information flows inside organizations.

[事实] He gives the example that if a company has separate communities and videos teams, the website will likely have separate communities and videos pages.

[事实] He says teams shipping into the same thing must have information flows between them, or they will step on each other’s toes.

[16:47] Two CEO Tools: Integrate Personally or Reorganize

[事实] Shear says one CEO tool is to become the join point personally, integrating information across teams and making the contextual decision.

[事实] He says the second tool is to reorder the company so the team structure matches the desired product structure.

[事实] He argues that if a founder wants to merge two parts of a product, merging the teams is the clearest signal of seriousness.

[事实] He says he tested Conway’s Law at Twitch and is using it deeply at Softmax.

[18:08] Founder Mode as People Systems

[事实] Shear says he thinks about companies in a people-led way, even though he sees himself as a systems person.

[事实] He says the systems he cares about most are people systems and how people systematically relate.

[事实] For him, founder mode is often the willingness to decide which teams should focus on what, and whether work should be combined or split apart.

[推测] This reframes founder mode as organizational design, not only product taste or founder intensity.

[19:29] Treat Organization and Communication Like Product

[事实] Shear says founders should treat organization structure, company communication, people leaders, and people leadership with as much seriousness as they treat product.

[事实] He says spending three hours a week building a clear keynote can be worthwhile because it amplifies the founder’s work.

[事实] At Twitch, he says the company needed to be “smart all over,” because he could not be in every negotiation.

[事实] He contrasts that with products where a few surfaces, such as checkout, must be exactly right.

[21:01] Closing

[事实] The hosts thank Shear for discussing Softmax and founder mode.

[事实] They say they hope he can return to Social Radars to talk more about Softmax.

[事实] Shear says he would love to come back when the timing is right.

播客点评/总结

This episode is valuable because it connects AI alignment and founder leadership through the same underlying concern: how agents, people, and organizations form shared context. Shear’s framing makes the conversation more philosophical than a typical startup tactics interview, but it remains grounded through examples from Twitch and Softmax.

The strongest practical insight is the distinction between founder mode as direct control and founder mode as context propagation. Shear argues that the founder’s job changes as the company grows: sometimes the founder must make the call directly, and sometimes the founder must design communication and team structures so the right calls happen elsewhere.

A limitation is that the Softmax discussion stays mostly abstract. The transcript mentions benchmarks, simulations, reinforcement learning environments, and architecture, but does not give detailed examples of what the company has shipped or how its measurements work.

[推测] This episode is best suited for founders, executives, and AI alignment listeners who are comfortable with conceptual discussion and want a deeper model of leadership, organizational design, and agent alignment rather than a checklist of startup tactics.