How to talk to statues — Joe Reeve, ElevenLabs

Joe Reeve, from the growth organization at ElevenLabs, presents a fast-paced, largely conversational session at AI Engineer Europe 2026 about building the viral "Talk to a Statue" app and broader implications for voice AI, vibe coding, and multimodal interaction design.

The statue app was built in two hours on a Sunday using Cursor and a single one-shot prompt. The pipeline works as follows: a user points their phone at a statue, OpenAI Deep Research identifies the object and generates historical context plus a voice description, the ElevenLabs Voice Design API generates a matching voice from that natural-language description, an ElevenLabs agent is instantiated, and a conversation begins — the entire flow completes in approximately 30 seconds. Joe posted the demo on a Tuesday and it received 50,000 impressions; he then reposted emphasizing vibe coding and it went to 1.5 million impressions the following day. The virality attracted museums (including the Science Museum and V&A), auction houses like Bonhams and Christie's, and travel-adjacent platforms like Tripadvisor competitors, all wanting similar experiences built for their collections. A museum CEO tracked down Joe's WhatsApp number to ask how a single developer had outpaced a 10-person team working for a year.

The session emphasizes that the technical "hard part" is not the back-end scaling — ElevenLabs' voice design and agents APIs are designed to absorb heavy volume — but rather the curation layer: working with museum curators to build canonical, accurate content rather than relying on live web research. The V&A's public collection API is cited as an example of existing institutional data that can be integrated. A philosophically interesting challenge is raised: what voice should an inanimate object have? A carved object with Chinese-origin stone, shaped in Vietnam, residing in a British museum for 200 years raises genuinely difficult questions about provenance and sonic identity that ElevenLabs is exploring in collaboration with former British Museum Americas head Jago.

The audience Q&A covers multiple forward-looking topics. On voice as an interface, Joe argues that the binary between pure voice and visual interfaces is a key unsolved problem — users want high-bandwidth voice input but high-density visual output, not symmetric voice-in voice-out. He describes wanting to speak to a product manager agent that then dispatches a coding agent, rather than speaking directly to the tool. The group discusses the difficulty of interrupting voice agents, the rudeness perceived in terse audio responses, and the lack of backchanneling ("uh-huh") cues in current agent interactions. Joe proposes an "async transcript analysis" architecture for agents to signal intent to interrupt without blocking the main conversation stream.

On vibe coding, Joe argues that the real consumer breakout moment has not yet arrived — current tools like Lovable still target B2B SaaS builders. He draws an analogy to the Facebook Instant Games API era, where he bought a Fruit Ninja clone for £15, instrumented it with the social graph API, and woke up to 15 million users in Vietnam. He believes a similar social-primitives moment for vibe coding is coming, and that vibe coding events are already producing people who have never considered that apps are made by humans.

On content creation, Joe describes his video production process: editing on a phone using CapCut in 20–25 minutes, using a DJI Bluetooth lapel mic (£200), front-loading hooks in the first 6–12 seconds (matching his analytics on median view duration), and generating background music with ElevenLabs' commercially licensed AI music API before or after the spoken narrative depending on the desired vibe.