# TechWave EP162: Eighty Percent Is Quick, Every Time Is Hard Source: Realpha Blog (blog.getrealpha.com) Original article and charts: https://blog.getrealpha.com/en/blog/techwave-2026-11-16-ep162-ai-ai-speak-colton-gyulay/ > Notes on TechWave EP162 (16 November 2026): Speak COO Colton Gyulay on the last mile of an AI product, matching the level of intelligence to the cost of the task, anchoring price to the thousands a year people spend on human tutoring, how channels differ across Korea, Japan and Taiwan, and how Taiwanese users pushed the company into enterprise. Educational listening notes only — no investment advice, no price targets, and the judgements here may be wrong. Published: 2026-11-17 Locale: en Tags: TechWave, podcast-notes, AI products, language learning, market expansion TL;DR: Getting a product to eighty percent is quick; making it work every time is the moat. And the businesses whose definition of good is hardest to write down are the ones models reach last. ![Night on an apartment balcony, one person in earbuds speaking into a phone, hallway light receding deep into the flat behind them, city streetlights spread out beyond the railing](/covers/techwave-2026-11-16-ep162-ai-ai-speak-colton-gyulay-cover.png) > I do not open the door for a student who has not struggled, nor find words for one who has not yet failed to find his own. Show him one corner, and if he cannot turn it into three, I do not go on. > > —— The Analects, Book VII "Shu Er" (Spring and Autumn period; translation mine) In TechWave EP162, published 16 November 2026, host Harry interviews Colton Gyulay, chief operating officer of the AI language-learning app Speak. Colton says that after Speak soft-launched in Taiwan, the growth curve took about a year to start bending, and that what justified building a local team was the funnel: people signing up, starting trials, converting to paid subscriptions even at small scale, and weekly and monthly retention holding. He also says people spend thousands of dollars a year on human tutoring and classroom learning, so Speak anchors its price against that number. One condition behind all of it: Speak has raised money, which lets it ignore token costs while testing new experiences — room a smaller company does not have. ## A ten-year-old company and a 10-minute phone call First, who is speaking. Speak is an AI language-learning app whose core function is getting people to practise talking. Colton Gyulay is its COO, running global operations and market expansion. He took time off Harvard to co-found a YC-backed gaming hardware startup, went back to finish a computer science degree, did product management on Google's Pixel team, and spent time in healthcare. He mentions in the episode that Speak just passed its ten-year mark, while the modern version of the product is only three or four years old. The conversation starts with the state of the consumer AI market and works through product design, how to choose a new country, how the company uses AI internally, and who it hires now. The second half gave me the most — how the Taiwan market was actually built, and the layer of judgement a user like me cannot see from inside the app. ## Eighty percent is quick, every time is hard Colton says building a prototype or a first version is easier than ever, and that the last mile remains hard: something that works 80 or 90 percent of the time comes together fast, while something that works every time and solves the same customer problem predictably does not. He treats the willingness to spend resources polishing details to the level they need as what separates a handful of products and companies from the rest, because a person opening your product should know what they will get and know it will work. I read that as a platitude the first time through. On the second pass I saw the business inside it: reliability goes by the name trust in this market, and trust can be bought with money and time while it cannot be copied and pasted. ![A completion curve that rises steeply through the first 80% and then flattens out, with the long flat stretch on the right labelled as the last mile.](/figures/last-mile-curve-en.svg) ## Start with the problem, then look at the technology's window Harry asks whether to start from the technology or from the problem. Colton picks the problem, then adds a lens I found useful: ask at the same time what is possible today, or will be possible in three, six or twelve months, that was impossible several months ago. He calls that a way to find opportunities that did not previously exist. The two questions need each other. Ask only about the problem and you rebuild something that was tried and failed three years ago; ask only about the technology and you build something nobody needs. I had not separated those two failure modes before. ![A four-quadrant chart has whether the problem is real on the horizontal axis and whether the technology has only just become possible on the vertical axis, with only the upper-right quadrant highlighted.](/figures/problem-x-tech-window-en.svg) ## Expensive intelligence only where it earns its keep The cost section is the most practical part of the episode. Colton says that when they test a new product experience they do not start by optimising token cost — first they find out whether they can create a step change in what people feel they are getting. Once that exists, they go back through it: this part can be turned down, that part does not need maximum intelligence and can be swapped out. Ordinary conversation runs on cheaper off-the-shelf models; long-range curriculum planning, and carefully explaining and correcting a learner's mistake, is where they reach for something smarter. ![One entry splits into two paths; the upper path leads to a very small box, and the lower path leads to a much larger box.](/figures/cheap-and-smart-routing-en.svg) Pricing follows the same move. People spend thousands a year on human tutoring, he says, so show users the outcome they are actually getting and anchor the price against that outcome — a better reference point than ten dollars a month. His comparison is customer support, where many vendors charge something like a dollar per resolution, because having a human resolve the same ticket costs more, which makes the trade-off obvious for the buyer. ![Two vastly different-height pillars stand side by side: one very tall pillar labeled human tutors and classes, and one barely above the ground labeled monthly subscription fee.](/figures/pricing-anchor-en.svg) ## Chess was solved thirty years ago and more people play it than ever If AI translates this well, why learn a language? Colton answers in two layers. For transactional interactions, near-real-time translation serves people adequately. But people have a deep need for genuine connection, and learning to speak someone's language is the shortest route to it, along with an understanding and appreciation of the culture behind it. Then he brings up chess. Computers solved it thirty years ago, and yet more people play today than ever before, and they play better, because the tools that help them train got good. His inference is that language learning follows the same path: give everyone one-on-one instruction at a fraction of the cost of a human tutor and more people will succeed, which pulls more people in — where previously most people who tried a language gave up without much progress. He also cites Bloom's 2 Sigma Problem, the finding that one-on-one human tutoring runs about two standard deviations better than traditional group classrooms on improvement and success, and says the opportunity there goes well beyond languages. ![A small cluster of dots appears on the left, while a much larger and denser cluster appears on the right, separated by a dashed line marking thirty years ago and today.](/figures/chess-more-players-en.svg) ## What worked in Korea hit a wall in Japan This section lands hardest for a Taiwanese listener. Speak's first market was Korea, then Japan and Taiwan. Colton says the core message held across all three — the best way to learn is by speaking, and this product lets you speak without fear of mistakes — while the channels and the execution changed. Advertising on big established channels like Meta and Google worked faster out of the box in Japan. X is far bigger there; he thinks Japan may be the world's largest X audience, and that same channel never worked for them in Korea, where they barely tried it. The reverse also happened. Influencer and key-opinion-leader marketing worked in Korea from the start and failed repeatedly in Japan. His phrase for it was banging their heads against the wall, and they had to rethink partners and content types from the ground up before the channel moved. His read on Japan is a market that leans on institutional signals, where earned media and TV placements do more to build trust, while in Korea and Taiwan a good Instagram or YouTube placement matters more. ![A single circle in the center stands for one core message, with channels branching out to the left and right, and the marks for what works and what hits a wall are exactly reversed on the two sides.](/figures/same-message-different-channels-en.svg) The Taiwan story is my favourite in the episode. Taiwanese relatives of a friend saw their early ads and said, "That was so Taiwanese" — which he calls the greatest compliment they could have received, because it meant the message connected with problems people here actually feel. In practice, local teams get a high degree of trust and autonomy to develop genuinely local concepts and build campaigns around them; headquarters campaigns do not get shipped over wholesale. ## Taiwanese users pushed the company into enterprise The causal chain here is the part of the episode I did not expect. Taiwan was the first market where Speak felt a strong pull toward workplace English, with users concentrated on work-related goals, so that is where they first invested in interview and job curricula. That led the whole company toward businesses and schools — and Colton credits Taiwan's users with convincing them the opportunity was real. There is one more layer. Enterprise, school and higher-ed buyers care about proving that learning works and showing measurable outcomes, and that demand rewrote the consumer product: they began investing in measuring and demonstrating efficacy, so learners can see what they can do now that they could not before. Colton says they want people hooked on making learning progress, using that in place of gamification and streaks running on a dopamine loop. A set of fussier buyers pushed the product in a healthier direction. ![A row of evenly sized check-in dots on the left becomes a staircase rising step by step on the right, while a buyer-requirement box above points to the space between them with an arrow.](/figures/streak-to-progress-en.svg) ## He is running an internal campaign against long AI documents Engineering changed the most. Colton says code is agent-written now and engineers mainly review and improve pull requests, with that step often agent-assisted too. Other functions move slower; the approach is to hand out tools and budget, give people time for hackathons inside teams, and pair more technical teams with less technical ones so the builders act as champions. He is blunt about the cost. He reads more very long documents than ever before, calls that output AI slop, and has a sort of internal campaign going against it. Early in the year the push was to apply AI everywhere; now it is to use it for the right things — you stand behind the output, use your brain to make the right decision, and stop taking tool output at face value. On budget he says he has no amazing answer: there is a monthly baseline threshold, set high enough that most people never reach it, and hitting it triggers a conversation — what are you working on, what problem are you solving, is this a good use of money? If it is, here is more budget, keep building. ## "Now that AI is here, anyone can build this — what moat is left?" This is where I get stuck most often looking at consumer apps, and Colton offers a cut I had not considered. Some business and product problems are easier to articulate and solve, he says, and those are the products and opportunities that get commoditised fastest. Where it is difficult to define what good looks like, the business may stay durable much longer, even as models get more intelligent. He uses his own hard problem as the example: long-range curriculum planning. What makes a great curriculum, what maximises someone's progress toward a goal over months, what makes a single lesson enjoyable — there is no quantitative right or wrong, much of it is qualitative, so even asking the strongest frontier model to plan every lesson and unit produces work that, by their internal experience, still is not good. He files this under problems that need taste or higher-level judgement. ![A downward-sloping curve rests on a horizontal spectrum: the left end is easy to describe and solve, while the right end is where it is unclear what good means.](/figures/hard-to-define-good-en.svg) The half-sentence that follows matters more. Internal effort has shifted toward building evaluations for the specific problems the company wants to solve, because good evals let you keep testing models, scale things up or down, and keep a check on output quality. So when I listen to an AI company's story now, I add one question: do they have their own evals, and can they say what good looks like? Where they cannot, they do not know how good their own output is either. ## "A company says it cracked a new market — can I believe that?" Also a common question, and Colton's sequence is borrowed easily. What tells them a market deserves more money is how well the product fits that market and those users, with acquisition sitting behind it. Before growth arrived, they watched the funnel: sign-ups, trials starting, conversions to paid even at small scale, weekly and monthly retention strong enough. Taiwan went exactly that way — about a year between the soft launch and the growth curve bending, with those numbers carrying the investment decision in between. So when an outside story says growth in some market is fast, it is worth asking whether that growth was bought or grew on its own. Colton says they use media mix modelling to attribute which users came from which paid activities, which pulls the organic share out separately. The other thing worth asking is sequence: after Korea worked, Japan's influencer channel still hit the wall several times over, and the channel had to be rebuilt. A flat first year in a new market can mean failure, and it can mean normal. Retention tells those apart; volume of noise does not. ![Two stacked columns of equal height stand side by side: the left column is mostly paid acquisition, while the right column is mostly organic growth.](/figures/bought-vs-organic-growth-en.svg) The mistake I have made runs the other way: a product becomes visible everywhere around me and I assume it is established everywhere. After this episode I ask how long it has been here first. ## Further reading - TechWave EP162 (16 November 2026), hosted by Harry, with Speak COO Colton Gyulay - Speak's own site and app feature pages, where AI Tutor Call and the curriculum categories can be seen directly - Benjamin Bloom's 1984 paper introducing the 2 Sigma Problem, with the comparison between one-on-one tutoring and group classrooms ## One thing to take away Work whose definition of good is hard to write down gets automated last. The whole episode circles that: the last mile of a product, which problems deserve expensive models, why frontier models cannot plan a curriculum. Each answer lands in the same place — tools become useful once a person can say what good means, and where nobody can, the tools simply generate slop faster. Here is something I have tried. Take one thing you handed over recently, with no need for any connection to investing: a report, a dinner you cooked for friends, a letter of apology, the way you sit with a child doing homework. Write three lines, each one a visible sign that would show up if the thing had gone well — and the sign has to be checkable by a third person, rather than known only to you. Three lines means you hold your own eval; go back to it a week later and mark yourself against it. Come up short of three, and I take it as a sign I am doing that thing on feel.