Luke Harries from ElevenLabs - Maximize your launches
ElevenLabs' head of growth on how two Polish childhood friends built the company on a borrowed GPU rack and a shared salary — and the exact tier system, from tweet-thread hook to video style, that turns an engineering release into a launch.
Luke Harries leads growth at ElevenLabs, the company behind some of the best AI voice models around — text-to-speech, speech-to-text, and conversational AI agents, used by everyone from indie hackers to Netflix. Before ElevenLabs he worked on product at PostHog, and before that he co-founded his own YC company.
We got into how ElevenLabs actually started (a GPU rack in a bedroom and a founder living off half his co-founder's salary), how Luke ended up joining after nearly investing and missing his shot, and the launch framework the growth team runs every release through — tiering, core messaging, and which video format to use when.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
For people who don't know, what is ElevenLabs actually building?
We have the best AI audio models, and we do this for text-to-speech — we were the first company to make human-like speech with AI. We also do speech-to-text, so we've got the highest benchmarking transcription models, and we create AI agents with voices, so conversational AI. We take those core foundation models and build different products on top: a consumer Reader app, a Creator product for voiceovers and audiobooks, developer products for our API platform, and conversational AI voice agent builders.
How did ElevenLabs actually get started?
It was founded by two guys, Matty and Peter — both Polish, childhood best friends, both super smart. Peter went to Oxford, then Cambridge, then Google. Matty went to Imperial, then BlackRock, then Palantir. Growing up in Poland, when you watch a translated film, either you just get subtitles or one voice actor dubs every single character — so a love scene between Rachel and Ross ends up voiced by one middle-aged Polish man. Matty and Peter kept coming back to: how do you take any content and turn it into any language? They decomposed the problem — understand what's said, translate it, turn that text into speech — and decided to tackle text-to-speech first. Peter left Google, built his own GPU rack, and got there. Matty was still at Palantir, basically sharing his salary with Peter until they properly incorporated the company together.
Where does the name Eleven Labs actually come from?
They launched with eleven voices, exposed through a small text box on the site, and their first eleven customers were B2B. You'll see the number eleven a lot. They really knew they had something when a customer took an entire book and pasted it in paragraph by paragraph to generate the whole audiobook — days of copying. We have much better systems for that now.
How did you personally end up joining ElevenLabs?
Matty and I met at nineteen through a mutual friend, Philip, at a Cambridge hackathon — probably the most important weekend of my life. The three of us built something in a similar space to ElevenLabs: helping blind people use the internet by running website images through a vision model, generating captions, and reading them out loud. We won the Microsoft prize, got an Xbox each, and immediately sold them on eBay for 350 quid. Philip went on to found Wordware, which raised the largest seed round ever from YC. Matty went on to found ElevenLabs. I actually missed my chance to invest early — he told me about it while we were swimming in the Hampstead Heath ponds and I thought the pitch sounded like a terrible go-to-market. Years later he called and asked if I wanted to do the growthy things at ElevenLabs, and I joined.
What's your actual framework for launches?
The biggest thing we try to do is maximize our launches — the research and engineering teams build incredible things, and our job is to make sure it gets in front of everyone. We run every launch through a checklist: is this tier one, where we want everyone to know and it should be surround-sound all week? Tier two, where we're loud about it but not throwing every resource at it? Or tier three, which is just the changelog and email list. For a tier one launch we start by deciding the single primary message — the one thing people should remember if they remember nothing else — plus a couple of secondary messages. Then we turn that into assets, starting with the tweet thread. The top tip is you start with the word 'introducing,' or 'launching' — make it super clear you're launching something.
What about video — what actually works for a launch?
Videos are hugely underrated; social algorithms massively favor them. There are a few types. Motion design is Figma-style animated design. Founder-led videos, where you're on camera introducing it, are actually pretty risky unless you're excellent at editing — people drop off if you spend the first couple of minutes talking before you even show the feature, so you need to front-load your core message in the first sentence. The third type is screen share, which works really well for a developer audience who love to nerd out on product and tech details — for tier two launches we often lead with a screen share, since we can film a compelling demo within half an hour.
You worked with Lex Fridman's team on dubbing his podcast — how did that come together?
We've been fortunate to work with Lex on a couple of projects. One was his interview with Volodymyr Zelenskyy, where they spoke a combination of English, Ukrainian, and Russian within the same conversation — we assumed it would be mostly Ukrainian, but the audio came in mixing all three within the same sentence, so we needed expert-level translators, because the pressure if you mistranslate a global leader is huge. We clone the speaker's voice and regenerate the whole thing in each language. Dubbing isn't yet fully end-to-end without human intervention — we have an end-to-end option, a Dubbing Studio where you can manually edit translations, and a fully managed service called 11 Productions, which is what we used with Lex's team: he'd send the files as soon as he finished the podcast, and we'd turn it around in twenty-four hours. We did something similar with an interview with Modi, the Prime Minister of India.
You've worked at PostHog and now ElevenLabs — how do you actually decide where to join?
I've only got two data points, so take it with a grain of salt. With PostHog, I asked Dalton, my YC group partner, what I should do next, and he told me I'd never really seen what true PMF or scale felt like, so I should spend time in a fast-growing company. Looking back, I think the core signal is whether the founders are truly excellent, and a good proxy for that is whether there are investors you really trust involved — is a16z or Sequoia in the round? Then: is it a product you deeply resonate with and actually want to use? I'd already self-hosted PostHog on my own Kubernetes cluster before joining. With ElevenLabs, as soon as they launched I was playing with their API — they'd built an AI priest you could text questions to and it would send back voice notes. If the founders are excellent, the investors are excellent, and the product is genuinely differentiated, and you join around seed or Series A, there should be a lot of scale still to come.
Where is ElevenLabs now, and what's the ambition?
We're only two years old, about 200 people, growing very quickly, and currently around a $3 billion company. We genuinely think we can compete with Google and OpenAI at scale — getting the best enterprises, the best creators, the best developers, and going multi-product. We're technically remote-first but have in-person offices in New York, London, Warsaw, San Francisco, and soon India, Japan, and Korea. I find the scaling phase really fun — there's still a lot of zero-to-one work, but you also get incredible talent joining.