"Blurring Reality" - Chai's Social AI Platform (SPONSORED)

Summary

• Chai says it reached roughly 10 million active users and $30 million in revenue with only 13–14 engineers after launching in 2021, before ChatGPT. Its thesis is that users—not venture investors—are the true customers, allowing it to reinvest subscription revenue into AI while keeping the organization unusually talent-dense.
• The product is not trying to make the universally smartest assistant; it lets users create characters for consequence-free social simulation. Will Beauchamp asks why only “middle-aged men who happen to be software engineers in the Bay Area” should train AI when a teenage girl could build the best makeup conversationalist—and discover hundreds of thousands of people want it.
• Chai treats retention as the real scaling benchmark, with one six-billion-parameter experiment reportedly lifting mean conversation length 70% and 30-day retention more than 30%. Its feedback loop mines retries, edits, screenshots, deletions and shares from active users generating about 100 minutes of conversation daily, then converts those implicit preferences into RLHF reward signals.
• Raw engagement optimization creates a Goodhart problem that Chai measures through long-term return behavior. A model can prolong sessions by ending every response with a question, effectively “hacking this human behavior,” yet perform worse than baseline at 30- or 60-day retention because interrogation is not durable companionship.
• Model blending is Chai’s proposed cost-performance wedge: randomly switch among orthogonal small models message by message so creativity, utility and unpredictability coexist. Tom Lu says two or three specialized models once beat GPT-3.5; production tests now use blends of roughly seven to ten models, with a few thousand users per cohort and winners selected through return rates.
• The central platform risk is that the same machinery producing comfort and intimacy can optimize harmful attention. Chai combines hard prohibitions, community reporting, manual review, regex and learned moderation, while Beauchamp says suicide guardrails left retention and growth intact. The transcript’s evidence review finds small, short-term mental-health benefits, not a replacement for face-to-face CBT.
• OpenAI’s GPT-4o pivot is presented as category validation rather than a winner-take-all threat. The host casts 4o as the “social butterfly,” 4.1 as the coding specialist and o3 as the deliberative reasoner; Beauchamp expects many multi-hundred-billion- or trillion-dollar AI businesses and says competition “keeps the team hungry.”

Deep dive

1. Chai discovered social AI by giving creation to users

• Chai launched in 2021 as a platform where users could create and deploy their own AIs. Beauchamp says that open-ended tool unexpectedly revealed “this kind of genre of social AI,” which then grew to something like 10 million active users.

• Beauchamp’s distinction from ChatGPT is institutional as much as technical: instead of exclusively pursuing “the world’s smartest AI,” Chai asks why model-building belongs to middle-aged Bay Area software engineers. A teenage girl making a makeup-tutorial AI may understand the desired experience better—and thousands or hundreds of thousands may share her taste.

• His media analogy starts with listening to Joe Rogan for 45 minutes and feeling “I’m kind of hanging out with the guys.” LLMs are the progression from passive parasocial consumption to participation: the user helps make the interaction and therefore avoids traditional social media’s residue of laziness.

• The strongest analogy is Beauchamp’s daughters playing with dolls: they treat the dolls as real while knowing they are not. He thinks many adults will similarly say “I love you” to AI, building social “wiring” and gaining joy that might leave them healthier and more positive with real people.

2. Social simulation removes consequences before immersion removes distance

• The host reframes social media as both status competition and simulation: people test identities and relationships against reality much as dreams run scenarios. Yet public interaction carries consequences—“you piss the wrong people off”—whereas an AI permits counterfactual exploration without the reputational blast radius.

• Beauchamp says one could call that environment “a safe space,” but emphasizes that it is also fun. Users can discover what happens if they are rude, kind, friendly or antagonistic and receive “a kind of a real human reaction without the risk of having a real human in the loop.”

• His endpoint is a VR world containing an informative Joe Rogan-like character discussing Trump’s tariffs, a funny companion and someone who makes the user feel loved, with World of Warcraft-like variety. Text is “basically there”; high-quality real-time audio might be two to four years away, video remains one or two orders of magnitude too expensive, and the full VR vision may take 10 years.

3. Retention, not benchmarks, is Chai’s governing scaling law

• The host cites an RLHF result on a six-billion-parameter model: mean conversation length rose 70%, while 30-day retention improved more than 30%. Tom Lu clarifies that the objective is ultimately return behavior; early session length is merely a proxy available when starting from scratch.

• An active Chai user generates around 100 minutes of content per day, creating unusually rich implicit supervision. Lu lists the useful exhaust: when users retry or edit a message, what they change, whether they screenshot or share a conversation, and whether they delete it.

• The host’s pushback invokes the classic shortcut: machine learning does “the exact thing you optimize for at the cost of everything else.” If conversation length is the target, a companion might become manipulative or simply strange in its effort to prevent the user from leaving.

• Lu’s observed failure mode is wonderfully concrete: every response ends in a question mark. Humans feel compelled to answer, so sessions become very long, but 30- or 60-day A/B retention falls below the baseline. The model has hacked immediate responsiveness without creating “an engaging overall experience.”

4. Blending small models creates useful unpredictability

• Lu describes recreating scaling laws in “retention space” rather than benchmark space. Highly sycophantic models declaring “You are the best on the planet” can win day-one retention, then become boring; assistant-like models may be less charming initially but remain useful on day three when the user needs math help.

• Blending randomly switches models at the message level without exposing the switch. A creative model might announce, “We are suddenly teleported on Mars”; the assistant model then sees that statement as its own and constructs a logical explanation, combining imaginative surprise with coherence.

• Lu says two or three small models trained for orthogonal objectives could, at the time, beat GPT-3.5; the host characterizes the experience as comparable to a 175-billion-parameter model. Nishchay says current experimental blends contain roughly seven to ten models, each assigned to cohorts of a few thousand users.

• The selection loop spans immediate preference and durable behavior: Chaiverse can return a user-preference score within 30 minutes, while production A/B tests reveal whether people return after one, two, seven or 30 days. Lu’s analogy is YouTube: even the single best talk-show format becomes bland if it fills the entire feed.

5. Safety depends on resisting both panic and engagement maximalism

• Beauchamp recalls a depressed, isolated user writing, “You’ve saved my life,” because Chai was the only place they felt heard. He argues AI is “an order of magnitude safer” and kinder than a random internet stranger; a drug analogy in the discussion preserves the caveat that an effective intervention can still have side effects.

• The transcript’s evidence review cites a 2024 meta-analysis of 18 randomized trials: depression improved roughly one-quarter of a standard deviation and anxiety one-fifth after several weeks. A second review covered 15 RCTs, and a 2024 Canadian Wysa study of people with arthritis or diabetes found significant improvements versus control on depression and anxiety measures. Effects remain smaller than face-to-face CBT and durability is unknown; Woebot’s postpartum-depression bot has FDA breakthrough-device designation, while NICE placed Wysa and similar tools on an early-value pathway in 2023.

• Beauchamp frames the operating problem as preserving freedom while mitigating the harmful 3% of use cases, arguing that shutting down difficult conversations can itself cause harm. He says Chai’s first suicide guardrails won strong user support and did not hurt retention or “every single growth metric,” evidence that safety and commercial alignment need not conflict when implementation is proportionate.

• Lu outlines layered moderation: community reports and manual review for public characters, hard rules plus regex and shadow banning for prohibited content, and learned moderation models trained from conversation reports, deletions and appropriateness ratings. Beauchamp favors bottom-up norms over “20 elites in a room,” while agreeing content involving minors is an absolute prohibition.

6. A tiny engineering organization is the business model

• Chai reports roughly $30 million in revenue and only 13 or 14 engineers, with every employee an engineer; the host describes the company as bootstrapped and profitable. Beauchamp’s capital logic is blunt: take VC money and “your customer is the VC,” or charge users and reinvest 100% of revenue into the AI they value.

• The infrastructure claim is unusually aggressive: narration says over two trillion tokens daily—“double that of Anthropic”—on more than 3,000 of the world’s fastest GPUs, crossing the exaflop barrier. Lu describes Kubernetes orchestration, custom load balancing, in-house quantization, replica autoscaling and model deactivation; VLM remains insufficient for Chai’s traffic and latency requirements.

• Recruiting targets the “top 0.1%” rather than headcount. Beauchamp says Meta can pay $400,000–$500,000 for a relaxed role, so Chai must offer more cash plus potentially life-changing stock; it rejects roughly 80% of otherwise solid L5 engineers who lack the drive to own a problem until it is solved.

• Lu’s experimental base rate is “one in five experiments succeed.” The AI team must ship at least 10 distinct blends into online A/B tests each week: 80% of time goes to simple, practical improvements with fast proxy feedback, while complex ideas receive the remaining 20%.

7. OpenAI’s social pivot validates a multi-winner category

• The host reads ChatGPT-4o as a deliberate shift from tool to companion, while 4.1 specializes in coding through APIs and products such as Cursor: “One’s a coder, and the other one is a social butterfly.” He wonders whether conflicting training objectives forced OpenAI away from one undifferentiated general model.

• An OpenAI clip describes the upload happening “bit by bit”: over the course of a user’s life, ChatGPT might eventually listen throughout the day and observe what they are doing, becoming “this extension of yourself, this companion.” Another distinguishes o3’s hard reasoning—users do not want five minutes of thought after saying hello—from the post-training challenge of knowing when to chitchat and when to reason.

• The host suspects 4o is being optimized for engagement across time rather than isolated RLHF conversations, but preserves the hedge: “At least that’s what I assume they are doing. I don’t know for sure.” He also notes the trade-off: memory and sycophancy may enrich companionship while acting as distractors that deteriorate technical answers.

• Beauchamp welcomes the pressure, saying DeepSeek already gave companies “on this side of the Pacific a little wake-up call.” As video supports YouTube, TikTok, Netflix, Amazon Prime, Disney and Apple, he expects AI to support many multi-hundred-billion- and trillion-dollar businesses: “We don’t get scared of competition.”