William Beauchamp
Key Views & Dialogues
Outlasting Noam Shazeer, Crowdsourcing Chai AI w/ 1.4m DAU — with William Beauchamp, Chai Research
- 🗓️ Date：
2025-01-26｜ 🎙️ Show：Latent Space
Chai’s social-AI model turns users into character creators and uses roughly 5,000-completion preference tests to compress model evaluation from 30 days to about three hours. With 1.4 million DAU and over $22 million in revenue cited but not explicitly confirmed by William Beauchamp, the key watchpoints are inference costs, paid acquisition and whether Chai’s creator moat deepens.
View Dialogue Notes & Key Takeaways
Chai’s thesis is that social AI will become a distributed creator platform, not one monolithic model advantaged solely by more data, compute, and researchers. William Beauchamp expects specialized teams and users to shape different AIs, much as quantitative firms specialize inside one financial marketplace. “There must exist a platform where a small team can produce an AI for a unique purpose.”
Product-market fit appeared only after utility bots failed and Beauchamp’s sister built a therapist bot that drew 20 users for roughly 20 minutes each. News, recipes, jokes, quizzes, celebrities, and influencers went nowhere; immediate, judgment-free conversation was “the thing that AI is 10x better at.” Chai then let consumers create characters using a prompt, image, and name, uncovering demand for conflict, romance, and archetypes its engineers would never have designed.
Alessio cited Chai at 1.4 million DAU and over $22 million in revenue; Beauchamp did not explicitly confirm those figures, saying instead that he thinks users grew 3x and revenue more than doubled last year. A typical session lasts about 90 minutes versus the cited 70 minutes for TikTok and generates roughly 150 messages, making inference economics far more important than in question-answer products. The relevant frontier is not benchmark performance alone but “performance per dollar.”
The 2024 growth curve reflects four major changes: repairing infrastructure, suspected weaker competitive acquisition, better models, and paid distribution. Firebase stopped scaling reliably around 500,000 DAU, forcing a painful three-month migration before Chai could reach 1.5 million; later, Beauchamp suspected Character.AI reduced advertising after its founders left. Chai then ramped acquisition to about $40,000 a day, moving annualized user growth from roughly 2x to 3x: build the product, then attach a large “rocket” by buying ads.
Chai’s strongest operating advantage is a human-feedback loop that compresses a conventional 30-day retention test into roughly three hours. Each submitted model is put before users for comparisons; Beauchamp said roughly 5,000 completions provide an accurate signal. About five researchers now evaluate 20–50 models daily and ship at least 100 a week. Somewhere in October, he thinks, user commentary flipped from Character.AI being better to Chai being better—evidence he values because “you can’t cheat consumers.”
The company killed voice as a growth thesis after three months of work produced no measurable retention, engagement, or monetization lift. Only 10–15% of users tried it, for 10–15% of their time—about 2% of the total experience—despite Beauchamp insisting the models were great. His product test is now severe: a feature must matter to most users for most of the experience and feel like “a big deal.”
Beauchamp has pushed out his AGI timeline and rejects the idea that current LLMs are intrinsically strong reasoning engines. He calls them simulators with “superknowledge”: exceptionally good at storing, retrieving, and generating, but still weak at intelligence as he defines it. Chai does not stream; it generates 16 complete candidates and uses a reward model to select one. He described training such a model from 50 million messages as an example, not as a confirmed deployed configuration.
🔗 Original source & video: Outlasting Noam Shazeer, Crowdsourcing Chai AI w/ 1.4m DAU — with William Beauchamp, Chai Research