The Staircase to an AI Tutor: Connor Wick on Speak, Moats, and Voice-First Learning

Jacob Effron introduces Connor Wick as CEO of Speak, an English-language learning platform backed by OpenAI that launched in South Korea in 2019, grew to more than 10 million users and customers in over 40 countries, and most recently raised at a $500 million valuation. Connor's entrepreneurial path began in high school with a flashcard app that sold, and Jacob asks how he would run that app today. Connor says the obvious modern move would be to build a generative-AI layer on top of the flashcards, because the app had already accumulated structured knowledge pairs.

The flashcard app had several million users and produced hundreds of millions of decks and billions of cards, giving it a graph of linked front-and-back knowledge. Connor remembers wanting to aggregate that into a graph that could generate and teach anything—an omniscient tutor. He now says that with hindsight, crawling the entire internet and compressing it into a large model was probably the better data strategy, though structured flashcard data would still be valuable for learning because it is centered on what a person is trying to learn.

Connor's crash course in AI predates the Transformer: he describes paying attention to RNNs and convolutional neural nets around 2015 and crashing a Berkeley course. He and his cofounders explored computer vision ideas such as cameras on city vehicles for automated meter maids—which they rejected as bad for the world—custom clothing measurements, medical imaging, and deep-learning weather prediction. They ultimately gravitated to speech because they wanted to build technology that felt like a relationship with a persona behind it.

Speak is a full fluency solution focused on speaking and real conversations rather than grammar, vocabulary memorization, or flashcards alone. The pedagogy teaches high-frequency chunks of words, has learners say them until automatic, and then practices them in simulated conversations tied to the learner's own motivation—for example, Spanish for a trip to Mexico City. The experience is meant to be highly individualized to the learner's interests, level, and reasons for learning.

Connor says Speak was deliberately long-term oriented: the team expected models to improve over 5 to 10 years, surpass humans on various tasks, and eventually replace the human in the learning process. That North Star forced product decisions to align with a coherent vision instead of short-term optimization. Connor describes the roadmap as a flight of stairs, with the product meaningfully evolving every year or two; early unlocks included accurate speech recognition, phoneme recognition, and basic language understanding.

Jacob asks whether startups should build around today's model shortcomings or wait for improvements. Connor says builders need a deep technical intuition for how models work today and in the future, but also must understand and articulate the user problem; progress on that problem can be worthwhile even if chunks of technology are swapped later. Speak's head start, he argues, comes partly from its long work on learning methodology, engagement, and motivation—value outside the core model.

In the short to medium term, Speak builds in-house models for specialized tasks where general models are not yet good enough. Its speech recognition is tuned for accented speech, for understanding what learners are trying to say, for detecting specific mistakes, and for streaming results quickly and reliably to the client. Its phoneme-recognition system uses Speak's data to detect pronunciation and prosodic errors; Connor expects multimodal speech-to-speech models may eventually subsume this work, but not tomorrow.

Connor calls specialized models a big investment but only one of many. He compares the current moment to the 1980s personal-computer market, when Apple used Intel processors instead of building its own because specialized vendors spent enormous sums doing it better. In his analogy, the modern 'AI firmware' is what Speak calls ML scaffolding: the technically difficult layer that orchestrates models, backend systems, product behavior, and data.

That scaffolding includes evaluation, model chaining, inference infrastructure, fine-tuning decisions, data collection, and specialized systems too hard to buy from vendors. Speak is also building a real representation of a language and a knowledge graph of where each learner is proficient and where they are not. Connor says at least 50% of product development time now touches these systems, and he considers this layer a bigger long-term technological moat than modeling alone.

Connor finds it silly that Speak still does prompt optimization such as telling a model to 'pretend you're very friendly,' though he expects that need to diminish as models get smarter. He is more focused on interface design: any tooltip, config panel, or user-education screen is a sign that the team has not designed the experience intuitively enough. Speak is inventing audio-first paradigms, and its onboarding now asks users to press a microphone and say why they are learning English—leaving many users unsure what to say, which language to use, or how long to speak.

Speak uses the word 'hybrid' for its interface goal: users should be able to talk, type, or tap at any point. Connor does not think speech is always better; it is better some of the time, while a keyboard can be faster in other cases. He also notes that ChatGPT has made conversational and open-ended AI interfaces more familiar, which is already shifting user expectations.

Beyond pull-based querying, Connor sees an interface unlock in a background GPU that thinks about the user and acts proactively—perhaps reminding a traveler in Paris to practice ordering at a café. A more concrete example is running compute-heavy analysis overnight when Korean users are asleep, then preparing distilled lessons or analysis for the next day. The mental model, he says, is technology that thinks about you when you are not using it, which he expects to be a profound shift.

Connor resists making structured curriculum and open-ended personalization mutually exclusive. Language learning has a right sequence at the core: about 100 words account for 20% of usage, and the next 500 words account for 80%, so beginners need a certain set of high-frequency vocabulary first. Within that first concentric circle, however, the ordering and even the specific 500 words can be made bespoke to the learner. He still expects humans to shape curriculum strategy and methodology at a high level for a long time, while the ML team takes on more of the actual creation—requiring cross-functional understanding of how Speak teaches.

Connor cites Neal Stephenson's The Diamond Age as a major inspiration: an AI-powered primer that can teach a girl anything and can take a fully unique, creative direction for an individual. That image captures the kind of personalized learning Speak often discusses internally. The appealing moment is when a learner ventures off the standard path into a way of learning that is weird or unique to them.

Jacob asks whether model inference costs constrain Speak's product ideas. Connor says the company does not feel especially constrained for its paying subscribers, and it has no free tier—though a free tier would create a different challenge. If costs did constrain a valuable feature, Speak would be inclined to build it anyway and eat the short-term cost because it expects inference costs to fall. Lower costs, he thinks, should expand demand more than they compress revenue.

Speak thinks about pricing at both extremes. One goal is to make Speak radically accessible to hundreds of millions of people, since software can deliver language practice at a marginal cost far below traditional tutoring. The other is a premium consumer product that can charge dramatically more than today's app because millions of people already pay hundreds of dollars per month for offline tutoring or classes. Connor says Speak does not need to price as a dollar-for-dollar substitute, but it must be differentiated enough to avoid being competed down in price.

Connor calls evaluation one of the most underrated and difficult parts of AI work. For open-ended LLM tasks—and even for speech—the most important step is distilling the evaluation, because a perfect eval basically distills the problem you are optimizing and makes optimization almost straightforward. Speech evaluation is not just word error rate: Speak must decide whether it is catching individual mistakes, handling substantially unintelligible words, and whether a model should understand words humans would not. Once the eval framework is clear, Connor says, it drives execution clarity.

When a new model such as GPT-4o arrives, Speak runs a process rather than assuming the hype is correct. It has close provider relationships, internal tools, and a playbook covering roughly 40 major tasks, the different eval loops for each, and which evaluations require humans. The company also tracks product experiments and guardrail metrics, because shipping to a subset of users and watching behavior is part of the evaluation. Connor says this playbook reduced organizational thrash every time a model changes.

Jacob asks whether AI mostly benefits incumbents such as Duolingo. Connor says AI is a sustaining technology when it simply improves the same problem an incumbent already solves, but it can be highly disruptive when it fully automates a loop—for example, replacing a fleet of human customer-support agents rather than making their managers more efficient. In language learning, he argues Duolingo and Speak solve different problems: Duolingo built a casual, brain-training-like app that made language learning feel productive, and most of its subscribers were not learning a language before. Speak focuses on English learners who have spent 10+ years studying but lack access to human speakers for conversational fluency.

Real-time translation and accent-changing tools might obviate some casual language needs, but Connor is skeptical that they remove the core motivation for Speak's users. Even the best translators have latency and imperfection because languages order words differently. More importantly, Speak's users want human connection—professional, cultural, or social—with more people than those who speak their native language. A universal 'Babel fish' might help a tourist, but it does not solve the fundamental reason someone wants fluency.

Connor is excited rather than threatened by GPT-4o's speech-to-speech capabilities, even if the market read Duolingo's stock drop as a warning. He expects ChatGPT to teach more people that AI can be used for language practice, and serious learners will then look for a specialized, effective solution for a habit that takes months or years. Speak wants to own the category the way Airbnb owned house sharing and Uber owned ride sharing. More people using AI to learn languages also means less confusion when they encounter Speak's first prompt.

On audio-only versus multimodal models, Connor expects room for specialized audio models in niches such as security-sensitive speech, on-prem deployment, or unusual vocabularies, and startups may take risks that large players will not. He praises ElevenLabs for unleashing creative audio applications while noting that OpenAI has to be more conservative. What Speak is waiting for is better multimodal, multilingual audio: a continuous speech-to-speech model that avoids the lossy text bottleneck, preserves tone, emotion, and confidence, and lowers latency. He also wants stronger reasoning and reliable task completion for curriculum planning, though he says Speak is less constrained by reasoning than many higher-stakes domains because language learning can be useful with today's technology.

Connor says language learning has an advantage: Speak can build a fully useful, disruptive product with the current technology and little need for higher-stakes reasoning or perfect hallucination control. That is not true in domains where a human still needs to be in the loop because consistency and reasoning are not good enough. Speak is also building an enterprise version in South Korea, where companies such as Samsung and SK want to offer Speak to employees and English is a professional skill. That opens adjacent opportunities such as learning to give a good presentation in English, not just learning English.

Speak sees three major expansion sectors: schools, businesses and professional skills, and personal learning. Personal learning is almost invisible but massive—reading books, listening to podcasts, watching YouTube, and reading articles are all learning-adjacent behaviors, and Connor expects personal learning to become one of the biggest changes in human activity. He compares it to the early internet and search engines, where behavior existed before the product category was understood. A specialized premium consumer product is one opportunity, but he expects the space to be crowded and competitive.

Zooming out, Connor argues that software ate much of the world but did not fundamentally change education quality: Chromebooks, digital quizzes, digital flashcards, and recorded video lectures mostly digitized old methods. The best way to learn 2,000 years ago—Alexander the Great learning from Socrates—is still essentially one-to-one tutoring, and AI can finally change that. For other subjects, he thinks the bar is higher because current alternatives are better than the status quo in language learning; math may require more reasoning and selling to schools is more of a product-and-market problem than a technology problem. He warns that people are overhyped about short-term change and underhyped about a decade of change, and he hopes the Transformer does not become a local maximum that crowds out other research. He still expects education to be one of the biggest areas of AI-driven change.

At the end, Connor says the AI world is overhyped on funding without much true product-market fit or real usage, and underhyped on research beyond the Transformer. His biggest recurring surprise is that every exciting new capability looks like a panacea but is never as good as expected, and changing user behavior is really hard; even tying human-level transcription into GPT-4 for open-ended lessons was good but not a game-changer. He has changed his mind on building all models in-house, recognizing that some models are too expensive to build and alternatives to Transformers may matter. Speak is hiring across roles and points listeners to speak.com and its careers page.