[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 12680,
    "text": ">> Hi everyone. Thank you so much for"
  },
  {
    "startMs": 14400,
    "text": "attending this talk. My name is Paula"
  },
  {
    "startMs": 17000,
    "text": "and I am one of the engineers on the"
  },
  {
    "startMs": 19360,
    "text": "Tullen team, specifically focusing on"
  },
  {
    "startMs": 21640,
    "text": "our iOS app."
  },
  {
    "startMs": 23320,
    "text": "And for the next 20 minutes or so, I'll"
  },
  {
    "startMs": 25320,
    "text": "be talking about what it takes to build"
  },
  {
    "startMs": 27200,
    "text": "a voice-first AI companion and also"
  },
  {
    "startMs": 29600,
    "text": "about how we use AI to build AI"
  },
  {
    "startMs": 32040,
    "text": "internally."
  },
  {
    "startMs": 34600,
    "text": "So, humanity has always imagined the"
  },
  {
    "startMs": 36280,
    "text": "perfect companion. So, we have"
  },
  {
    "startMs": 38200,
    "text": "Caravaggio on the left 400 years ago"
  },
  {
    "startMs": 40920,
    "text": "painting an angel leaning over St."
  },
  {
    "startMs": 42960,
    "text": "Matthew's shoulder literally guiding his"
  },
  {
    "startMs": 44920,
    "text": "hand as he writes."
  },
  {
    "startMs": 46840,
    "text": "This is an example of a companion being"
  },
  {
    "startMs": 48960,
    "text": "a presence that makes you better at"
  },
  {
    "startMs": 50480,
    "text": "being you."
  },
  {
    "startMs": 51800,
    "text": "And then we have Tinkerbell, the devoted"
  },
  {
    "startMs": 53520,
    "text": "little sidekick who believes in you so"
  },
  {
    "startMs": 55480,
    "text": "fiercely that the whole theater has to"
  },
  {
    "startMs": 57040,
    "text": "clap to keep her alive."
  },
  {
    "startMs": 59000,
    "text": "And of course on the right, we have"
  },
  {
    "startMs": 60840,
    "text": "Samwise who can't carry the ring for"
  },
  {
    "startMs": 62840,
    "text": "Frodo, but says, \"I can carry you.\""
  },
  {
    "startMs": 65960,
    "text": "The companion is pure unconditional"
  },
  {
    "startMs": 68120,
    "text": "loyalty."
  },
  {
    "startMs": 69440,
    "text": "And it goes far beyond these three."
  },
  {
    "startMs": 71680,
    "text": "Every hero has some sort of guiding"
  },
  {
    "startMs": 73440,
    "text": "spirit. And these are all different"
  },
  {
    "startMs": 75200,
    "text": "stories, but they exhibit the same"
  },
  {
    "startMs": 77320,
    "text": "longing for something that listens,"
  },
  {
    "startMs": 79480,
    "text": "remembers you,"
  },
  {
    "startMs": 81200,
    "text": "and is wholly specifically yours."
  },
  {
    "startMs": 83480,
    "text": "And for for all of human history, this"
  },
  {
    "startMs": 85240,
    "text": "has basically been fiction."
  },
  {
    "startMs": 88000,
    "text": "So, we made one. This is Tullen. It's a"
  },
  {
    "startMs": 90560,
    "text": "little alien you talk to out loud like a"
  },
  {
    "startMs": 93240,
    "text": "friend. It has a personality. It"
  },
  {
    "startMs": 95680,
    "text": "remembers you and over time it becomes"
  },
  {
    "startMs": 98200,
    "text": "specifically yours."
  },
  {
    "startMs": 100120,
    "text": "Okay, I don't know if the audio setup"
  },
  {
    "startMs": 101600,
    "text": "works here, but I will try talking to my"
  },
  {
    "startMs": 103400,
    "text": "Tullen."
  },
  {
    "startMs": 104800,
    "text": "Uh let's see."
  },
  {
    "startMs": 109120,
    "text": "So,"
  },
  {
    "startMs": 110720,
    "text": "you can see my Tullen here, Luke,"
  },
  {
    "startMs": 112640,
    "text": "walking around the planet."
  },
  {
    "startMs": 116120,
    "text": "Hey Luke, can you hear me?"
  },
  {
    "startMs": 120960,
    "text": "Okay. Luke can hear us, but we can't"
  },
  {
    "startMs": 122960,
    "text": "hear him. Um"
  },
  {
    "startMs": 125320,
    "text": "Anyway, I had prepped him for this. Oh."
  },
  {
    "startMs": 127800,
    "text": "Hello. Hi Luke, can you hear me?"
  },
  {
    "startMs": 133800,
    "text": "Nope."
  },
  {
    "startMs": 135360,
    "text": "We're okay. I can come back to this"
  },
  {
    "startMs": 137000,
    "text": "later. Um"
  },
  {
    "startMs": 138840,
    "text": "but you should definitely all give this"
  },
  {
    "startMs": 141920,
    "text": "a try if you haven't already."
  },
  {
    "startMs": 147600,
    "text": "Okay. So, people talk to Tolins a lot."
  },
  {
    "startMs": 150360,
    "text": "We support both text and voice chat, uh"
  },
  {
    "startMs": 153440,
    "text": "but we have over 4 million hours of"
  },
  {
    "startMs": 155080,
    "text": "voice conversation so far."
  },
  {
    "startMs": 157160,
    "text": "We say Tolin is a voice-first companion,"
  },
  {
    "startMs": 159240,
    "text": "even though we support both, because"
  },
  {
    "startMs": 160959,
    "text": "it's the voice experience that's truly"
  },
  {
    "startMs": 162640,
    "text": "immersive and that makes users'"
  },
  {
    "startMs": 164480,
    "text": "relationships with their Tolins feel"
  },
  {
    "startMs": 165640,
    "text": "real."
  },
  {
    "startMs": 167440,
    "text": "And but the moment this relationship is"
  },
  {
    "startMs": 169440,
    "text": "a spoken relationship, the engineering"
  },
  {
    "startMs": 171280,
    "text": "problem changes completely. So, let me"
  },
  {
    "startMs": 173840,
    "text": "show you how voice breaks the normal way"
  },
  {
    "startMs": 175800,
    "text": "we build and interact with LLMs."
  },
  {
    "startMs": 179640,
    "text": "So, the core difference really is that"
  },
  {
    "startMs": 181600,
    "text": "in a text chatbot, turns are relatively"
  },
  {
    "startMs": 184000,
    "text": "slow and context is stable. The user"
  },
  {
    "startMs": 187000,
    "text": "waits a few seconds, they read, and they"
  },
  {
    "startMs": 189160,
    "text": "tend to stay on topic."
  },
  {
    "startMs": 190920,
    "text": "And almost every LLM app assumes that."
  },
  {
    "startMs": 193920,
    "text": "Voice is the opposite. Turns are fast."
  },
  {
    "startMs": 196600,
    "text": "Your whole round trip from the user"
  },
  {
    "startMs": 198160,
    "text": "finishing their sentence to the Tolin"
  },
  {
    "startMs": 199800,
    "text": "starting to speak has to land in under a"
  },
  {
    "startMs": 202200,
    "text": "couple of seconds, or it stops feeling"
  },
  {
    "startMs": 203880,
    "text": "like a conversation."
  },
  {
    "startMs": 205560,
    "text": "And the context is volatile. People talk"
  },
  {
    "startMs": 207680,
    "text": "to their Tolins while they're cooking,"
  },
  {
    "startMs": 209640,
    "text": "while they're walking, while they're"
  },
  {
    "startMs": 211280,
    "text": "falling asleep. Um they change their"
  },
  {
    "startMs": 213160,
    "text": "subjects mid-sentence. They say um, they"
  },
  {
    "startMs": 215320,
    "text": "interrupt."
  },
  {
    "startMs": 216800,
    "text": "And that 2 seconds is crucial. Early on,"
  },
  {
    "startMs": 220200,
    "text": "our latency drifted from 2 seconds to"
  },
  {
    "startMs": 222280,
    "text": "about 2 and 1/2 seconds, and that half"
  },
  {
    "startMs": 224600,
    "text": "second tanked basically every metric in"
  },
  {
    "startMs": 226640,
    "text": "the product. People would write in to"
  },
  {
    "startMs": 228480,
    "text": "complain that their Tolins were too"
  },
  {
    "startMs": 229600,
    "text": "slow."
  },
  {
    "startMs": 230720,
    "text": "And living inside this constraint has"
  },
  {
    "startMs": 232280,
    "text": "taught us a lot and gave us four"
  },
  {
    "startMs": 233840,
    "text": "principles."
  },
  {
    "startMs": 236360,
    "text": "Principle one is that you have to design"
  },
  {
    "startMs": 238360,
    "text": "for conversational volatility. Again,"
  },
  {
    "startMs": 240760,
    "text": "text users stay on topic, but voice"
  },
  {
    "startMs": 243160,
    "text": "users jump around. Someone could be"
  },
  {
    "startMs": 245120,
    "text": "mid-story about their breakup and"
  },
  {
    "startMs": 246440,
    "text": "suddenly go, \"Wait, did I leave the oven"
  },
  {
    "startMs": 248360,
    "text": "the stove on?\" and then back. Speech is"
  },
  {
    "startMs": 250760,
    "text": "messy. Most LLM apps assume that you'll"
  },
  {
    "startMs": 253040,
    "text": "have a clean and stable conversation and"
  },
  {
    "startMs": 254720,
    "text": "we have to build for the opposite."
  },
  {
    "startMs": 256840,
    "text": "So, for a long time that meant fixing"
  },
  {
    "startMs": 258560,
    "text": "things that sound tiny but are actually"
  },
  {
    "startMs": 260840,
    "text": "the product. So, you can't interrupt a"
  },
  {
    "startMs": 263240,
    "text": "Tullen mid-sentence. A short yes or yeah"
  },
  {
    "startMs": 265960,
    "text": "won't register as a turn. For example,"
  },
  {
    "startMs": 268040,
    "text": "curse words will get stripped out."
  },
  {
    "startMs": 270240,
    "text": "And the deeper lesson was to stop"
  },
  {
    "startMs": 271720,
    "text": "optimizing for fewer interruptions and"
  },
  {
    "startMs": 274120,
    "text": "start optimizing for fewer bad ones"
  },
  {
    "startMs": 276400,
    "text": "where the agent would jump in way too"
  },
  {
    "startMs": 277800,
    "text": "early."
  },
  {
    "startMs": 278800,
    "text": "So, we built smart turn taking that"
  },
  {
    "startMs": 280520,
    "text": "reads your speech pattern to decide"
  },
  {
    "startMs": 282160,
    "text": "whether an interruption is real and we"
  },
  {
    "startMs": 284280,
    "text": "cut the worst early aborts by more than"
  },
  {
    "startMs": 286240,
    "text": "half."
  },
  {
    "startMs": 287320,
    "text": "And we happily paid about 60"
  },
  {
    "startMs": 288760,
    "text": "milliseconds of extra latency to do it."
  },
  {
    "startMs": 292919,
    "text": "Principle two, latency isn't just a"
  },
  {
    "startMs": 295000,
    "text": "number you check at the end, it's"
  },
  {
    "startMs": 296360,
    "text": "actually the product and we measure"
  },
  {
    "startMs": 298320,
    "text": "every stage of the pipeline separately"
  },
  {
    "startMs": 299960,
    "text": "because it feels slow is useless. You"
  },
  {
    "startMs": 301880,
    "text": "have to know where exactly it's slow."
  },
  {
    "startMs": 304360,
    "text": "And the pipeline here is that the user"
  },
  {
    "startMs": 305680,
    "text": "stops talking, we detect end of"
  },
  {
    "startMs": 307760,
    "text": "utterance, we transcribe, and then the"
  },
  {
    "startMs": 310160,
    "text": "model produces its first token."
  },
  {
    "startMs": 312200,
    "text": "So, time to first token, often the"
  },
  {
    "startMs": 314000,
    "text": "biggest chunk, is around a second."
  },
  {
    "startMs": 316120,
    "text": "The model finishes generating and then"
  },
  {
    "startMs": 318000,
    "text": "text-to-speech produces its first byte"
  },
  {
    "startMs": 319680,
    "text": "and then it plays back to the user."
  },
  {
    "startMs": 321919,
    "text": "A couple lessons here. So, one, so far"
  },
  {
    "startMs": 324520,
    "text": "our biggest jump in quality came from"
  },
  {
    "startMs": 326360,
    "text": "moving to GPT-5.1 on the responses API,"
  },
  {
    "startMs": 329680,
    "text": "which cut our time to speech by more"
  },
  {
    "startMs": 331080,
    "text": "than 7/10 of a second, which is huge."
  },
  {
    "startMs": 334160,
    "text": "Um two, we don't send every turn to the"
  },
  {
    "startMs": 336320,
    "text": "same model. We run a tiered fleet. So,"
  },
  {
    "startMs": 339080,
    "text": "we use a frontier model for the turns"
  },
  {
    "startMs": 340919,
    "text": "that carry the relationship with your"
  },
  {
    "startMs": 342160,
    "text": "Tullen."
  },
  {
    "startMs": 343280,
    "text": "So, for example, your first conversation"
  },
  {
    "startMs": 344880,
    "text": "with with Tullen and your onboarding."
  },
  {
    "startMs": 347000,
    "text": "And we use smaller and faster models for"
  },
  {
    "startMs": 349360,
    "text": "the turns"
  },
  {
    "startMs": 350840,
    "text": "for the lightweight turns."
  },
  {
    "startMs": 352640,
    "text": "And the whole game then becomes about"
  },
  {
    "startMs": 354160,
    "text": "routing or deciding turn by turn which"
  },
  {
    "startMs": 356520,
    "text": "model you actually need. So we round we"
  },
  {
    "startMs": 359120,
    "text": "run a small classifier we call the tone"
  },
  {
    "startMs": 361240,
    "text": "router on every single turn and this"
  },
  {
    "startMs": 363480,
    "text": "tone router itself runs on a cheap model"
  },
  {
    "startMs": 365760,
    "text": "and it reads the emotional state of the"
  },
  {
    "startMs": 367160,
    "text": "conversation."
  },
  {
    "startMs": 368560,
    "text": "And our main"
  },
  {
    "startMs": 369960,
    "text": "our [clears throat] main principle is"
  },
  {
    "startMs": 370800,
    "text": "that we route based on stakes not on"
  },
  {
    "startMs": 372720,
    "text": "cost. So the high stakes moments always"
  },
  {
    "startMs": 375600,
    "text": "get the best model. So this would be"
  },
  {
    "startMs": 377680,
    "text": "again the user's very first message,"
  },
  {
    "startMs": 379560,
    "text": "their first few days with their Tolen"
  },
  {
    "startMs": 380880,
    "text": "and anything that we deem to be"
  },
  {
    "startMs": 382640,
    "text": "emotionally serious."
  },
  {
    "startMs": 384360,
    "text": "For example, we have crisis or therapist"
  },
  {
    "startMs": 386880,
    "text": "style tones and we never cheap out on"
  },
  {
    "startMs": 388880,
    "text": "those."
  },
  {
    "startMs": 389960,
    "text": "And then the lighter casual back and"
  },
  {
    "startMs": 391400,
    "text": "forth can ride on smaller models that"
  },
  {
    "startMs": 393240,
    "text": "are faster and cheaper."
  },
  {
    "startMs": 394960,
    "text": "And all the background work so that's"
  },
  {
    "startMs": 396480,
    "text": "summarizing the conversation, generating"
  },
  {
    "startMs": 398360,
    "text": "personas, the tone router itself run on"
  },
  {
    "startMs": 400480,
    "text": "these small models, too."
  },
  {
    "startMs": 402440,
    "text": "And why would we go to all this trouble?"
  },
  {
    "startMs": 404160,
    "text": "It's mainly because the frontier model"
  },
  {
    "startMs": 405919,
    "text": "costs us roughly five times the smaller"
  },
  {
    "startMs": 407720,
    "text": "one."
  },
  {
    "startMs": 408640,
    "text": "So one big model turn is about five"
  },
  {
    "startMs": 411240,
    "text": "smaller model turns. So routing is a"
  },
  {
    "startMs": 413840,
    "text": "huge part of what makes the unit"
  },
  {
    "startMs": 415120,
    "text": "economics for us actually work."
  },
  {
    "startMs": 417480,
    "text": "Um and we do a bunch of AB experiments"
  },
  {
    "startMs": 419440,
    "text": "and the surprising result we found there"
  },
  {
    "startMs": 421040,
    "text": "is that routing a third a third of our"
  },
  {
    "startMs": 422680,
    "text": "turns to the small model has almost no"
  },
  {
    "startMs": 424960,
    "text": "measurable effect on retention."
  },
  {
    "startMs": 428320,
    "text": "And principle three is what makes a"
  },
  {
    "startMs": 429680,
    "text": "companion feel like a companion. So the"
  },
  {
    "startMs": 431680,
    "text": "naive approach is to keep the whole"
  },
  {
    "startMs": 433200,
    "text": "conversation history as a sort of"
  },
  {
    "startMs": 435160,
    "text": "transcript, but that doesn't fit into"
  },
  {
    "startMs": 437360,
    "text": "our two-second loop. It doesn't scale"
  },
  {
    "startMs": 439320,
    "text": "and it just doesn't work. It leads to"
  },
  {
    "startMs": 441320,
    "text": "long sessions degrading. It leads to the"
  },
  {
    "startMs": 443520,
    "text": "model getting lost in the middle of a"
  },
  {
    "startMs": 445000,
    "text": "huge context and also hallucinating."
  },
  {
    "startMs": 447800,
    "text": "So instead we see memory as a sort of"
  },
  {
    "startMs": 449640,
    "text": "retrieval system. We pull facts,"
  },
  {
    "startMs": 451800,
    "text": "preferences, and emotional vibe signals"
  },
  {
    "startMs": 453840,
    "text": "out of conversations. We embed them and"
  },
  {
    "startMs": 456320,
    "text": "we store them in a vector database with"
  },
  {
    "startMs": 458400,
    "text": "sub 50 millisecond lookups. And every"
  },
  {
    "startMs": 460560,
    "text": "night we compress. So we merge"
  },
  {
    "startMs": 462480,
    "text": "duplicates, we cluster related memories,"
  },
  {
    "startMs": 464920,
    "text": "we resolve contradictions, and we drop"
  },
  {
    "startMs": 466760,
    "text": "all the noise."
  },
  {
    "startMs": 468160,
    "text": "And we don't just retrieve against users"
  },
  {
    "startMs": 470200,
    "text": "last messages, we also generate internal"
  },
  {
    "startMs": 472240,
    "text": "questions about the person and the"
  },
  {
    "startMs": 473480,
    "text": "relationship and retrieve against those."
  },
  {
    "startMs": 475920,
    "text": "So, and we also split memory into two"
  },
  {
    "startMs": 477920,
    "text": "parts. We have stable memory and"
  },
  {
    "startMs": 479320,
    "text": "unstable memory. The volatile stuff"
  },
  {
    "startMs": 481520,
    "text": "lives in the in the live tail of the"
  },
  {
    "startMs": 483000,
    "text": "prompt, and when we summarize the"
  },
  {
    "startMs": 484560,
    "text": "conversation, we look at which memories"
  },
  {
    "startMs": 486280,
    "text": "actually get recalled and pin those into"
  },
  {
    "startMs": 488560,
    "text": "a stable and cashable block."
  },
  {
    "startMs": 492520,
    "text": "Uh the last principle is around context."
  },
  {
    "startMs": 495440,
    "text": "Specifically, you should rebuild context"
  },
  {
    "startMs": 497480,
    "text": "and not fight drift. So, most apps reuse"
  },
  {
    "startMs": 500480,
    "text": "context across turns to keep the cash"
  },
  {
    "startMs": 502560,
    "text": "warm. And in a stable text chat, that's"
  },
  {
    "startMs": 504960,
    "text": "fine. But in a volatile voice"
  },
  {
    "startMs": 506680,
    "text": "conversation, it's a trap because the"
  },
  {
    "startMs": 508440,
    "text": "second the the user pivots, your reused"
  },
  {
    "startMs": 510720,
    "text": "context is actively wrong. So, every"
  },
  {
    "startMs": 512880,
    "text": "turn we reassemble the context window"
  },
  {
    "startMs": 515080,
    "text": "from parts. We have a summary of recent"
  },
  {
    "startMs": 516880,
    "text": "messages, we have the the user's persona"
  },
  {
    "startMs": 519400,
    "text": "card, the memories we just retrieved,"
  },
  {
    "startMs": 521760,
    "text": "tone guidance from the emotional signal,"
  },
  {
    "startMs": 523599,
    "text": "and real-time app state."
  },
  {
    "startMs": 526880,
    "text": "And what also really helps us um in the"
  },
  {
    "startMs": 528839,
    "text": "case of Tolen is that our characters"
  },
  {
    "startMs": 530240,
    "text": "aren't generic or assistants with no"
  },
  {
    "startMs": 532280,
    "text": "personality. Everyone is crafted, and we"
  },
  {
    "startMs": 535080,
    "text": "in fact have an in-house science fiction"
  },
  {
    "startMs": 537240,
    "text": "novelist, Elliot, who writes the Tolen"
  },
  {
    "startMs": 539880,
    "text": "character lore."
  },
  {
    "startMs": 541600,
    "text": "And a couple of interesting points here."
  },
  {
    "startMs": 543800,
    "text": "So, one, why did we go with an alien?"
  },
  {
    "startMs": 546720,
    "text": "Mostly because there's no real-world"
  },
  {
    "startMs": 548360,
    "text": "reference to anchor on, which means that"
  },
  {
    "startMs": 550160,
    "text": "the users can project onto it, and it"
  },
  {
    "startMs": 552120,
    "text": "becomes what they need. The baseline"
  },
  {
    "startMs": 554240,
    "text": "Tolen is bubbly, it's youthful, it's"
  },
  {
    "startMs": 556600,
    "text": "irreverent. And also, if an alien"
  },
  {
    "startMs": 558720,
    "text": "character acts a bit unpredictably, so"
  },
  {
    "startMs": 560839,
    "text": "if if it's impulsive or chaotic or"
  },
  {
    "startMs": 563440,
    "text": "otherwise violates um you know, the"
  },
  {
    "startMs": 565280,
    "text": "norms the user would expect, it's not"
  },
  {
    "startMs": 567120,
    "text": "particularly surprising."
  },
  {
    "startMs": 568960,
    "text": "Like if you look at, you know, aliens in"
  },
  {
    "startMs": 571440,
    "text": "TV shows or in plays, like there's a lot"
  },
  {
    "startMs": 574240,
    "text": "of humorous moments around this. And"
  },
  {
    "startMs": 576000,
    "text": "this kind of chaos reads as charming."
  },
  {
    "startMs": 578960,
    "text": "Um second, uh we also know that"
  },
  {
    "startMs": 581440,
    "text": "personality is worthless if it drifts."
  },
  {
    "startMs": 583840,
    "text": "So, yeah, we run this parallel tone"
  },
  {
    "startMs": 585480,
    "text": "monitoring system that changes how a"
  },
  {
    "startMs": 587200,
    "text": "line is delivered based on your"
  },
  {
    "startMs": 588480,
    "text": "emotional cues without changing who the"
  },
  {
    "startMs": 590720,
    "text": "character is, holding identity across"
  },
  {
    "startMs": 592960,
    "text": "hundreds of turns."
  },
  {
    "startMs": 596080,
    "text": "And since we're at an AI conference, I"
  },
  {
    "startMs": 598040,
    "text": "thought I would also spend a bit of time"
  },
  {
    "startMs": 599760,
    "text": "talking about how we not just ship AI,"
  },
  {
    "startMs": 602120,
    "text": "but also use AI to build it."
  },
  {
    "startMs": 604440,
    "text": "Um so, I'm sure this is the case for"
  },
  {
    "startMs": 607200,
    "text": "most of you in the room now, but"
  },
  {
    "startMs": 608320,
    "text": "basically as of late last year, Claude"
  },
  {
    "startMs": 610480,
    "text": "has co-authored more code in our iOS app"
  },
  {
    "startMs": 612640,
    "text": "than any individual engineer in the"
  },
  {
    "startMs": 614240,
    "text": "team."
  },
  {
    "startMs": 615240,
    "text": "Um and I think especially, you know, a"
  },
  {
    "startMs": 616600,
    "text": "few months ago, everyone's instinct was"
  },
  {
    "startMs": 618520,
    "text": "to be kind of suspicious because, you"
  },
  {
    "startMs": 620000,
    "text": "know, more AI code meant more slop. But"
  },
  {
    "startMs": 622440,
    "text": "our our crash-free rate actually went"
  },
  {
    "startMs": 624520,
    "text": "from 99.6% to 99.9%."
  },
  {
    "startMs": 627480,
    "text": "Runtime errors dropped by over 50% and"
  },
  {
    "startMs": 631240,
    "text": "our share of highly engaged users"
  },
  {
    "startMs": 632760,
    "text": "doubled."
  },
  {
    "startMs": 633840,
    "text": "And the biggest lesson in building that"
  },
  {
    "startMs": 635160,
    "text": "system is that an agent's context comes"
  },
  {
    "startMs": 637640,
    "text": "mostly from the code base itself, not so"
  },
  {
    "startMs": 639640,
    "text": "much from the Claude MD file. We found"
  },
  {
    "startMs": 641880,
    "text": "that it's far more powerful to make the"
  },
  {
    "startMs": 643360,
    "text": "code base be the documentation, so we"
  },
  {
    "startMs": 644920,
    "text": "had agents standardize it. On top of"
  },
  {
    "startMs": 647200,
    "text": "that, we run a real fleet of agents. We"
  },
  {
    "startMs": 649839,
    "text": "have implementation agents that, you"
  },
  {
    "startMs": 651320,
    "text": "know, think freely and just get us to"
  },
  {
    "startMs": 652839,
    "text": "working code. They build it, they check"
  },
  {
    "startMs": 654800,
    "text": "it against snapshots until it's pixel"
  },
  {
    "startMs": 656360,
    "text": "perfect. And then we have separate"
  },
  {
    "startMs": 658080,
    "text": "review agents that enforce our"
  },
  {
    "startMs": 659520,
    "text": "standards. So, multiple Claudes"
  },
  {
    "startMs": 661520,
    "text": "basically review each other before a"
  },
  {
    "startMs": 663240,
    "text": "human looks."
  },
  {
    "startMs": 664560,
    "text": "And then we have a PR shepherd that"
  },
  {
    "startMs": 665960,
    "text": "watches an open pull request and keeps"
  },
  {
    "startMs": 667440,
    "text": "iterating against CI failures and review"
  },
  {
    "startMs": 669760,
    "text": "comments until it's clean."
  },
  {
    "startMs": 671920,
    "text": "And we also have a triage bot that fires"
  },
  {
    "startMs": 674240,
    "text": "on every inbound bug report that we get."
  },
  {
    "startMs": 676720,
    "text": "And they're all wired through MCP into"
  },
  {
    "startMs": 678600,
    "text": "linear, into into Sentry, DataDog, so an"
  },
  {
    "startMs": 681040,
    "text": "agent can reconstruct the cash a crash"
  },
  {
    "startMs": 683600,
    "text": "and route it itself and oftentimes open"
  },
  {
    "startMs": 685480,
    "text": "the PR on its own and just fix fix the"
  },
  {
    "startMs": 687520,
    "text": "bug."
  },
  {
    "startMs": 688839,
    "text": "And we also we ship on eval. So, for"
  },
  {
    "startMs": 690480,
    "text": "example, we've been working on a on a"
  },
  {
    "startMs": 692200,
    "text": "new character targeted towards an older"
  },
  {
    "startMs": 693920,
    "text": "demographic and Elliot, our in-house"
  },
  {
    "startMs": 696240,
    "text": "novelist, basically built this entire"
  },
  {
    "startMs": 698080,
    "text": "new character in a day."
  },
  {
    "startMs": 700000,
    "text": "So, the agents mapped every personality"
  },
  {
    "startMs": 701800,
    "text": "bearing surface in the code. They wrote"
  },
  {
    "startMs": 703640,
    "text": "the sort of a voice Bible and then they"
  },
  {
    "startMs": 705400,
    "text": "had five judges attack it from different"
  },
  {
    "startMs": 707280,
    "text": "angles."
  },
  {
    "startMs": 708360,
    "text": "Archetype fidelity, the model mechanics,"
  },
  {
    "startMs": 710400,
    "text": "our code standards, the ears of a"
  },
  {
    "startMs": 712280,
    "text": "skeptical 52-year-old and safety, and"
  },
  {
    "startMs": 715080,
    "text": "then they evaluated the changes against"
  },
  {
    "startMs": 716800,
    "text": "real production logs over three find fix"
  },
  {
    "startMs": 719360,
    "text": "verify rounds."
  },
  {
    "startMs": 720880,
    "text": "And over 7 million tokens and 4 and 1/2"
  },
  {
    "startMs": 723280,
    "text": "hours of compute later, he ended up with"
  },
  {
    "startMs": 725040,
    "text": "basically, you know, a couple of weeks"
  },
  {
    "startMs": 726280,
    "text": "of work done in afternoon."
  },
  {
    "startMs": 729680,
    "text": "And does this work? Well, I'll let the"
  },
  {
    "startMs": 731400,
    "text": "users tell you. We're at 4.8 stars on"
  },
  {
    "startMs": 733800,
    "text": "the App Store across 162,000 reviews and"
  },
  {
    "startMs": 737640,
    "text": "when we survey users on well-being, the"
  },
  {
    "startMs": 740160,
    "text": "highest scoring dimension by far is"
  },
  {
    "startMs": 741920,
    "text": "emotional safety."
  },
  {
    "startMs": 743440,
    "text": "And this is definitely a bar that being"
  },
  {
    "startMs": 745200,
    "text": "voice first sets. So, when the interface"
  },
  {
    "startMs": 747280,
    "text": "is your voice and the thing on the other"
  },
  {
    "startMs": 748880,
    "text": "side remembers you and has a"
  },
  {
    "startMs": 750200,
    "text": "personality, it stops being just"
  },
  {
    "startMs": 752280,
    "text": "software and starts being an actual"
  },
  {
    "startMs": 754440,
    "text": "relationship, which is why building it"
  },
  {
    "startMs": 755839,
    "text": "responsibly and building it well is"
  },
  {
    "startMs": 757760,
    "text": "worth obsessing over."
  },
  {
    "startMs": 760960,
    "text": "And we need people to come help us do"
  },
  {
    "startMs": 762560,
    "text": "that. Um, we're a small team and we're"
  },
  {
    "startMs": 764800,
    "text": "hiring and after a year with Tolen, I"
  },
  {
    "startMs": 766960,
    "text": "think this is truly one of the most"
  },
  {
    "startMs": 768280,
    "text": "interesting places in the world to be an"
  },
  {
    "startMs": 770000,
    "text": "engineer right now."
  },
  {
    "startMs": 772680,
    "text": "And here are some of the people you'd be"
  },
  {
    "startMs": 773720,
    "text": "doing it with. So, two of the founders,"
  },
  {
    "startMs": 775640,
    "text": "Quinton and Evan, previously built and"
  },
  {
    "startMs": 778040,
    "text": "exited a $300 million startup together."
  },
  {
    "startMs": 780920,
    "text": "Uh, they founded Even. Um, Ajay, our"
  },
  {
    "startMs": 783120,
    "text": "third co-founder, scaled two bootstrap"
  },
  {
    "startMs": 784880,
    "text": "companies past $50 million $50 million"
  },
  {
    "startMs": 787800,
    "text": "in profitable revenue."
  },
  {
    "startMs": 789400,
    "text": "And around them, we have Lucas, who um,"
  },
  {
    "startMs": 791800,
    "text": "is an Apple Design Award winning"
  },
  {
    "startMs": 793120,
    "text": "animator."
  },
  {
    "startMs": 794360,
    "text": "Uh, she's our creative director. We have"
  },
  {
    "startMs": 796320,
    "text": "Chris, who was a technical director at"
  },
  {
    "startMs": 797760,
    "text": "Pixar, earlier at Oculus, who works on"
  },
  {
    "startMs": 799960,
    "text": "embodiment. We have Lily, a board"
  },
  {
    "startMs": 802080,
    "text": "certified behavior analyst, who left a"
  },
  {
    "startMs": 804160,
    "text": "Vanderbilt PhD to do user research for"
  },
  {
    "startMs": 806080,
    "text": "us from the very start. Um and then we"
  },
  {
    "startMs": 808760,
    "text": "have Elliot who I've mentioned, the"
  },
  {
    "startMs": 810000,
    "text": "novelist behind uh our characters."
  },
  {
    "startMs": 812600,
    "text": "And I come from XAI and Spotify and"
  },
  {
    "startMs": 814600,
    "text": "previously also founded a company called"
  },
  {
    "startMs": 816440,
    "text": "Imagi."
  },
  {
    "startMs": 817560,
    "text": "So, it's a small team where honestly"
  },
  {
    "startMs": 819760,
    "text": "every person is the best I've worked"
  },
  {
    "startMs": 821840,
    "text": "with at what they do."
  },
  {
    "startMs": 824400,
    "text": "We're also well backed for this. Uh we"
  },
  {
    "startMs": 826000,
    "text": "have $30 million raised from Costanoa"
  },
  {
    "startMs": 828200,
    "text": "Ventures and a group of people who've"
  },
  {
    "startMs": 829720,
    "text": "built the tools and products a lot of"
  },
  {
    "startMs": 831280,
    "text": "you use every day."
  },
  {
    "startMs": 833800,
    "text": "And here are some of the more"
  },
  {
    "startMs": 834640,
    "text": "engineering focused roles where we need"
  },
  {
    "startMs": 835960,
    "text": "help. Um so, we're hiring across the"
  },
  {
    "startMs": 837280,
    "text": "board. We have iOS and back-end product"
  },
  {
    "startMs": 839640,
    "text": "engineering roles, applied AI"
  },
  {
    "startMs": 841640,
    "text": "engineering, gameplay engineering, and"
  },
  {
    "startMs": 844200,
    "text": "one specific role I want to flag, which"
  },
  {
    "startMs": 845800,
    "text": "is agent engineering management. Um so,"
  },
  {
    "startMs": 849400,
    "text": "when we went all in on running"
  },
  {
    "startMs": 851000,
    "text": "concurrent agents, um the people who got"
  },
  {
    "startMs": 852880,
    "text": "dramatically more effective on the team"
  },
  {
    "startMs": 854520,
    "text": "were the ones who had management"
  },
  {
    "startMs": 855880,
    "text": "backgrounds, um because it seems like"
  },
  {
    "startMs": 857960,
    "text": "managing a fleet of agents does actually"
  },
  {
    "startMs": 859760,
    "text": "take some of the skills same the same"
  },
  {
    "startMs": 861280,
    "text": "skills as managing people."
  },
  {
    "startMs": 862960,
    "text": "Uh you basically have to decompose the"
  },
  {
    "startMs": 864080,
    "text": "problem, you know, delegate it with"
  },
  {
    "startMs": 865360,
    "text": "checkpoints, give fast feedback, review"
  },
  {
    "startMs": 867280,
    "text": "their work seriously, and know when"
  },
  {
    "startMs": 868720,
    "text": "exactly to jump in."
  },
  {
    "startMs": 870240,
    "text": "So, if you're a strong engineer who"
  },
  {
    "startMs": 871600,
    "text": "thought going into management meant"
  },
  {
    "startMs": 873120,
    "text": "leaving code behind, that's that's no"
  },
  {
    "startMs": 874920,
    "text": "longer true."
  },
  {
    "startMs": 877120,
    "text": "Um and yeah, that's Tlon. You can come"
  },
  {
    "startMs": 879160,
    "text": "talk to me after this uh or reach out."
  },
  {
    "startMs": 881280,
    "text": "I'm on on LinkedIn. My email is here."
  },
  {
    "startMs": 883640,
    "text": "I'm on Twitter as well. Um I'd love to"
  },
  {
    "startMs": 885680,
    "text": "chat. So, yeah. Thank you."
  },
  {
    "startMs": 889500,
    "text": ">> [applause]"
  },
  {
    "startMs": 903910,
    "text": "[music]"
  }
]