Thariq Shihipar on Closing a Startup and Leaving Anthropic

SOTA Wives hosts Sarah Chieng and Vivian Midha Shen welcome Thariq Shihipar, a former Anthropic engineer and YC founder, as the show’s third member. He joins just after leaving Anthropic and talks about his startup, the lab and what might come next.

Shihipar says he planned to take a quiet break after his last day, but news of his departure spread online within hours. Messages from company leaders followed, prompting him to explain the move publicly earlier than planned.

He joined Y Combinator’s winter 2020 batch. He remembers the program beginning in person before COVID disrupted demo day, which became a much more limited online process involving a slide and investor messages.

Asked to compare YC then and now, he is cautious about judging from the outside. He recalls earlier founders saying that financing norms had already changed by his cohort, and says recent founders he met were tackling difficult problems.

His company, One More Multiverse, built a platform for playing tabletop role-playing games online. He says the small team raised a substantial Series A during the pandemic gaming boom before it had established comparable traction.

The group jokes about collecting prestigious institutional badges. Shihipar says the temptation to optimize for acceptance or funding can be confused with building a durable product, though a strange early idea can also become the start of something useful.

He defines over-raising as taking capital far ahead of demonstrated traction. In his view, a high financing benchmark can make a later acquisition harder and give founders a misleading feeling that the work is already done.

Ultimately he shut down the startup and returned the remaining money. Before joining Anthropic, he spent time exploring AI, including work with South Park Commons and Goodfire and building a spreadsheet agent.

After roughly a year and a half at Anthropic, he says he wants time to rest, play World of Warcraft and visit his parents. The conversation presents this as his immediate plan, not an announcement of a new company.

The hosts and Shihipar compare growing up in Toronto and the Bay Area, joking about the bonds formed by people who move into San Francisco’s technology community. They note that the city contains many communities beyond the version visible on tech social media.

They discuss the pressure of competitive high schools and imagine what kind of education they might choose for future children. The speakers acknowledge that these are hypothetical personal preferences, not a universal prescription for families.

Shihipar says he hopes future children can feel comfortable in their surroundings without constant self-examination. The hosts connect that thought to the intense introspection they see in some achievement-focused technology circles.

He distinguishes arriving in San Francisco while running a remote company from later joining the local tech scene. In his experience, friendships and professional networks overlap heavily, making it difficult to draw a clean boundary between work and personal life.

Shihipar remembers an early AI-interpretability post receiving around 500 likes and initially finding the online community exciting. After growing his own audience, he came to see some of the surrounding social scene as more ordinary than its reputation suggested.

The group jokes about meeting anonymous online personalities and discovering that a polished persona differs from the person behind it. They extend the observation to creators whose on-camera identities can differ from everyday life.

Shihipar says his early widely shared Claude Code content came from experimenting with creative uses, including video. He could see potential in a rough workflow, then spent substantial effort making the work and explaining it publicly.

His advice on content is to keep doing the underlying work. Interesting posts often come after experimentation, repetition and tedious details; if a creator stops engaging with the work itself, the material becomes harder to sustain.

He says Anthropic appealed to him because he believed its coding models were particularly strong and expected coding to become a major AI application. That was his personal assessment at the time, not a universal model ranking.

Asked whether people idolize frontier labs, he says the attention can be unhealthy and describes large labs as magnets for talent. He includes himself in that pattern, while emphasizing that the people inside institutions are still ordinary colleagues.

The hosts ask whether startups can survive when major labs launch competing products. A YC investor notes that Anthropic itself was once a startup facing powerful incumbents and argues that there can still be room for new products.

The group describes startups as requiring momentum and repeated searches for product–market fit. They caution, through their own experience, that excessive analysis of a daunting competitive landscape can discourage founders before they test an idea.

Shihipar points to smaller AI products that executed a focused idea well. He believes dedicated teams can still find opportunities, while acknowledging that future competitive outcomes are uncertain.

Although his own startup ended, he says working on it taught him about prompting and using AI. A host suggests that the experience may have helped him contribute at Anthropic even without the startup producing the outcome its founders hoped for.

He says he was already experimenting with Claude Code before joining the lab, including workflows around databases and events. He rejects the idea that an institution automatically turns everyone into a genius: people also arrive with their own ideas and experience.

On social-media promotion, he says a repost from a large account is not a substitute for useful content. He sometimes shares friends’ work he likes and sometimes gives candid feedback when he does not think a post is ready.

The speakers discuss moving between labs with different missions and business incentives. Shihipar advises against tying one’s whole identity to an employer or institution; the team, role and opportunity to contribute matter as well.

They jokingly compare AI labs to fictional school houses. Shihipar describes the people he worked with at Anthropic as hardworking and earnest; the analogy is playful shorthand, not an objective account of entire organizations.

The conversation moves to friendships and relationships across competing companies. Shihipar compares employment to playing for a sports team: compete seriously while respecting people on the other side and remembering that careers can change.

One host says venture investors at competing firms can also be friends, despite an outside perception of constant rivalry. The group sees shared experience as one reason professional competitors may understand and learn from one another.

They briefly joke about online discussion of espionage and “honeypots.” None of the speakers claims firsthand knowledge of a specific case, and their speculation is not evidence of an actual incident.

The episode ends with a lighthearted hypothetical about how they might respond to an apocalypse. The hosts thank Shihipar and say they look forward to seeing what he chooses to do next.