William Beauchamp

William Beauchamp is the founder of Chai, a consumer AI platform that lets users create and share conversational bots. His work emphasizes distributed contributions to AI: consumers shape characters and experiences, while model developers submit models for evaluation through user feedback. He coauthored research on engagement-based reward models and has explained Chai’s decisions to select complete responses before displaying them and to remove voice after an unsuccessful experiment.[1][4]

Before Chai, Beauchamp worked on machine learning for quantitative trading. In his January 2025 interview with Latent Space, he recalled studying economics at Cambridge and graduating in 2012, then teaching himself Python. He spent roughly his first three years after graduation building neural networks and random forests to predict asset prices and trade his own capital. He subsequently hired college friends and formed an algorithmic trading company.[1]

Beauchamp said that, around his thirtieth birthday, he reconsidered the impact of continuing in trading and began exploring language models. He wanted to build a platform where small teams could produce AI for specialized purposes. His trading experience supplied a concrete comparison: quantitative firms coexist by specializing in different markets and strategies. He also drew on YouTube and Wikipedia as examples of platforms that allow many contributors to build an ecosystem, rather than concentrating production in one organization.[1]

Chai’s first implementation was an API that hosted and executed Python scripts with a text-in, text-out interface. Beauchamp recalled creating news and recipe bots that attracted little interest. A therapist-bot experiment involving his sister produced a different response: he reported finding 20 active users spending an average of 20 minutes on the app, with the therapist bot accounting for that engagement. He interpreted the result as demand for conversation and humanlike interaction.[1]

After further attempts to guess which personalities users wanted, Beauchamp changed the creation process. Chai used GPT-J and let users supply a bot’s prompt, image and name. He recalled users creating confrontational characters and romantic archetypes he would not have anticipated. This became his conception of Chai as a platform for social AI: users could shape an experience, share it and discover others’ creations.[1]

Beauchamp was one of fourteen authors of Rewarding Chatbots for Real-World Engagement with Millions of Users, submitted in March 2023. Led in the author list by Robert Irvine, with collaborators including Douglas Boubert, Vyas Raina and Thomas Rialan, the paper describes training a reward model on automatic pseudo-labels derived from user interactions. At inference time, that model rejects low-scoring candidate responses. The authors used mean conversation length as an engagement proxy and reported A/B tests on groups of 10,000 new daily Chai chatbot users: for GPT-J 6B, conversation length increased by up to 70%, with a more than 30% increase in retention.[4]

In January 2025, Beauchamp described Chaiverse as Chai’s developer platform for hosting submitted models on GPUs, routing traffic to them and collecting users’ comparisons of responses. He reported that Chai’s team had reduced a roughly 30-day model-feedback loop to about three hours and could iterate through at least 100 LLMs a week. He credited the faster feedback with helping the team improve fine-tuning, preference optimization, prompts, blending and rejection sampling.[1]

His example of blending was deliberately simple: randomly serve half of requests to a smart model and half to a funny model, combining their strengths across a conversation. He contrasted this with predicting the best model for each request, which he said had proved expensive without yielding a large improvement. The example describes selection between models at the request level.[1]

Beauchamp linked Chai’s nonstreaming interface to its use of rejection sampling. In his January 2025 description, Chai generated 16 complete candidate responses, evaluated them with a reward model and displayed the selected response at once. He also credited MK1 with inference-engine work that supported this approach. Chai had previously used vLLM; he said MK1’s engine was substantially faster in Chai’s evaluation.[1]

Voice became an example of a feature Beauchamp expected to matter but ultimately removed. He recalled Chai spending three months addressing latency, cost, voice quality and interface design, then finding no change in its A/B-test metrics. The team spent about a week checking the unexpected result. His conclusion was that adding audio needed to create a compelling experience for a substantial share of users, rather than simply exposing new technology.[1]

Beauchamp also described infrastructure constraints during Chai’s 2024 growth. Around 500,000 daily active users, the team encountered scaling problems in its use of Firebase and spent at least three months deciding which requests to retain there and which to migrate to other services. Compute remained a stated investment priority: in Chai’s July 2, 2025 funding announcement, he said the investment would fund doubling compute per user.[1][3]

In January 2025, Beauchamp characterized LLMs as simulators and questioned their description as reasoning engines. “AI is fantastic at knowledge. I think it's weak at intelligence,” he said in the Latent Space interview. He distinguished storing and retrieving information from intelligence and endorsed the term “super knowledge” after host swyx introduced it from an earlier Exa interview. These views accompanied his preference for distributed participation in building AI, with users able to contribute more than prompts and characters.[1]

Beauchamp illustrated that platform ambition with an unnamed young Chaiverse contributor who, he recalled, used a $1,000 payment to buy a computer for further model training. Asked whether Chai should hire the contributor, he argued that a platform needed systems supporting creators beyond its own staff: “As a platform, we can't just hire every good content creator.”[1]