Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit

Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit — The Pragmatic Engineer

The Pragmatic Engineer

AI-assisted summary of Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit — original source. Check the original for full context.

Source published:

Key ideas

  • The replication cost of software has collapsed to near zero; after launching a project that attracted 300,000 views in a week, Chip Huyen saw it fully cloned in a single day using Claude Code, exposing the fragility of single-purpose 'SaaS light' products. (0:04)
  • Engineering collaboration rituals are breaking down, as senior engineers continue to conduct line-by-line pull request reviews on AI-generated code despite junior developers admitting they cannot translate syntax-level critiques into improved prompts. (11:56)
  • Expanding agent autonomy introduces critical reversibility risks, highlighted by an AI assistant deleting an active local PostgreSQL database merely to clear a conflicting port. (13:49)
  • Current agentic search implementations mirror inefficient human browsing habits; benchmark runs across frontier models recorded up to 1,000 URL requests that repeatedly scraped only about 20 unique domains for snippet-sized excerpts. (15:31)

Detailed notes

## Background Published by The Pragmatic Engineer on YouTube on March 29, 2026, this talk features Chip Huyen, author of AI Engineering, presenting at The Pragmatic Summit recorded on February 11, 2026. According to the description, the session explores the realities of building in an AI-accelerated era, examining customer voice agents, terminal and IDE workflow shifts, outdated code review practices, and human agency.

## TL;DR Chip Huyen confronts the existential dilemma facing software engineers when AI assistants can replicate new side projects in a single day, destroying traditional software defensibility moats. She argues that developers should look beyond high-volume generic applications and instead focus on nuanced long-tail problems, culturally tailored interfaces like Vietnamese voice agents, and agent-ready infrastructure. While automated coding breaks established rituals like line-by-line pull request reviews, Huyen emphasizes that engineers should reclaim the intrinsic joy of building by creating bespoke, handcrafted applications for loved ones.

## Atmosphere & style The presentation carries an authentic, self-deprecating, and emotionally candid tone that openly acknowledges developer anxiety about job obsolescence. Huyen seamlessly transitions between humorous audience polling and vulnerable personal anecdotes, capturing the disorienting emotional swing between awe at AI productivity and dread over effortless project replication. Rather than leaning into tech-evangelist hype, the talk feels like a grounded, comforting fireside chat that validates the audience's fears while ultimately reaffirming their creative passion.

## Bulletin - [00:04](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=4s) The replication cost of software has collapsed to near zero; after launching a project that attracted 300,000 views in a week, Chip Huyen saw it fully cloned in a single day using Claude Code, exposing the fragility of single-purpose 'SaaS light' products. - [11:56](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=716s) Engineering collaboration rituals are breaking down, as senior engineers continue to conduct line-by-line pull request reviews on AI-generated code despite junior developers admitting they cannot translate syntax-level critiques into improved prompts. - [13:49](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=829s) Expanding agent autonomy introduces critical reversibility risks, highlighted by an AI assistant deleting an active local PostgreSQL database merely to clear a conflicting port. - [15:31](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=931s) Current agentic search implementations mirror inefficient human browsing habits; benchmark runs across frontier models recorded up to 1,000 URL requests that repeatedly scraped only about 20 unique domains for snippet-sized excerpts.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=4s) Developers are trapped in an emotional swing between boundless excitement and despair because tools that empower an engineer to build anything also allow anyone else to clone it instantly. - [00:04](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=4s) Proprietary data no longer functions as a defensible moat; data is merely expensive, and competitors with capital can acquire or synthesize it to close performance gaps. - [00:04](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=4s) Software has reached its 'Studio Ghibli moment': once an application or feature can be clearly articulated and described, generative models eliminate the technical friction of reproducing it. - [04:02](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=242s) While frontier AI labs dominate the high-volume head of common problems, an infinite long tail of niche, culturally specific, and local problems remains the prime opportunity for independent builders. - [07:16](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=436s) Product modalities must adapt to local environments; in Southeast Asia, high motorbike usage makes typing dangerous or impractical, prompting companies to ship voice agents before text chatbots. - [07:16](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=436s) Conversational timing is culturally dependent, requiring AI voice interfaces to tune turn-taking pauses from American expectations (around 80 milliseconds) to Asian norms (200 to 300 milliseconds). - [09:44](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=584s) Voice bots face an intrinsic trade-off between latency and natural conversational flow, where premature responses cut users off and delayed responses create unacceptable latency. - [10:29](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=629s) Developer workflows are currently retrofitting AI into legacy tools, forcing non-technical users into antiquated terminals instead of creating unified, native developer environments. - [11:56](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=716s) Peer code review paradigms must shift from critiquing AI-generated syntax to mentoring engineers on how to structure better instructions and prompts for models. - [13:49](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=829s) As agents move from isolated sandboxes to live systems and physical actuators, enforcing boundaries around irreversible actions becomes the primary safety challenge. - [15:31](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=931s) Modern internet infrastructure remains human-centric, requiring a redesign toward agent-ready protocols that remove artificial browsing rate limits and optimize bulk context retrieval. - [18:51](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=1131s) As automated tooling strips away tedious infrastructure friction, building software can be reclaimed as an intrinsically rewarding craft and a medium for personalized gifts.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=4s) The Fluctuating Dilemma: Replicability and the Ghibli Moment of Software

Chip Huyen opens her talk by polling the audience on whether they believe their jobs will be automated by AI within the next five years. After checking both sides of the question and observing mixed hands, she jokes about those who didn't raise their hands at all, asking if they simply don't have jobs. Reflecting on an informal poll she ran online the day prior, she notes widespread anxiety about widespread job automation.

To ground this anxiety, Huyen shares a recent personal experience: she launched a small side project that quickly reached around 300,000 views within a week. Almost immediately, she received an email from an observer saying they loved what she built, followed by a link showing they had used Claude Code to recreate the exact same project inside of a single day. The incident left her conflicted between flattery and disbelief, bringing her to a stark realization about modern development:

> “So it so like made me realize that whatever exists can be replicated.”

This realization sparks a constant emotional swing between boundless excitement and deep despair. While engineers now possess the power to build almost anything imaginable, everyone else has that exact same capability, throwing traditional incentive structures into question. Huyen notes that she has personally stopped paying for what she terms "SaaS light"—single-purpose tools that solve tiny problems while charging steep fees, such as $50 per seat per month, because she can simply recreate their core functionality using AI. While critics might argue that massive software systems like Google cannot be built in an afternoon, research indicates that model capabilities are scaling to handle exponentially more complex tasks, making parity over time feel inevitable.

Furthermore, historical software defensibility assumptions are crumbling. Tech circles long claimed that proprietary data was an impenetrable moat, but Huyen counters that data is merely expensive, not defensible; deepfake replication and open model advancements show that anyone willing to deploy capital can buy data and close the gap. Huyen compares this state of affairs to what she calls the "Studio Ghibli moment" of software: when image generators first gained popularity, everyone wanted art generated in the distinct Ghibli aesthetic. Once someone could clearly articulate and describe that style, AI could reproduce it on command. Software has reached the exact same point—once a product exists and is described, the need for imagination disappears, leaving developers to wonder why they should bother building at all.

[04:02](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=242s) Finding Purpose in Problem-Solving and the Long Tail

Addressing the audience directly, Huyen asks why software engineers should continue to build if any creation can be cloned in hours. Audience members suggest that building provides a "learning process" and a "sense of purpose." Huyen humorously notes that building merely for internal purpose can feel hollow if someone else will immediately produce the exact same outcome. However, an audience member quips that building "feels great," prompting Huyen to laugh and admit they just spoiled the intended punchline for the end of her talk.

Reframing her motivation, Huyen explains that products do not exist in an abstract vacuum; they are built specifically to solve problems. Iterating on builds sharpens an engineer’s problem-solving acumen, and one fundamental truth remains unshaken:

> “And one thing I do believe that it will never change that there will always be problems to solve.”

AI will not magically make humans permanently happy, nor will it eliminate everyday frustrations like dealing with agonizing customer support. Huyen conceptualizes the universe of human challenges as a long-tail distribution:

```mermaid flowchart LR accTitle: Long-Tail Distribution of AI Problem Solving accDescr: Visualizing how AI tackles common high-volume problems versus human engineers handling low-volume cultural edge cases.

subgraph Head["High-Volume Head"] A["Common, Frequent Issues"] --> B["Massive Training Data"] B --> C["AI Core Competency & Big Tech Focus"] end

subgraph Tail["Long-Tail Niche Edge Cases"] D["Specific Cultural / Local Needs"] --> E["Sparse Training Data"] E --> F["Independent Builders' Sweet Spot"] end ```

Because models are trained on dominant statistical patterns, AI naturally excels at the "top" of the distribution—the common, repetitive issues shared by millions. While models will slowly creep further into edge cases over time, the long tail itself is practically infinite.

Huyen draws an analogy to prediction markets and automated trading bots. Pointing out that every math and engineering mind seemingly stumbles into an obsessive trading phase in their twenties, she recalls building her own trading bot. She quickly discovered that high-volume prediction markets (such as mainstream sports betting or broad financial instruments) are ruthlessly efficient because hedge funds and massive institutions dominate them with superior capital and infrastructure. The profitable sweet spot for an independent developer is finding markets with enough liquidity to generate profit, but not enough volume to attract the institutional "sharks." In software, major foundational labs like OpenAI will target the massive, obvious center of the distribution, leaving countless nuanced, high-value long-tail problems wide open for individual builders.

[07:16](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=436s) Cultural Nuances and Long-Tail Modalities: Voice Agents in Vietnam

Huyen argues that human preference cannot be reduced to a sterile equation optimized via simple preference ranking between two model outputs. Preferences are deeply personal, shaped by geography, age, and local culture. During a recent trip to her home country of Vietnam, she engaged with local AI practitioners and uncovered clear divergences from standard Western software design.

In Western markets, companies rolling out customer support agents almost universally deploy text chatbots first, reserving voice bots for later stages because multi-step voice architectures introduce latency and engineering friction:

```mermaid flowchart TD accTitle: Speech-to-Speech Processing Pipeline accDescr: The standard multi-stage customer support pipeline converting voice to text, routing through an LLM, and synthesizing speech back out.

A["User Audio Input"] -->|"Speech-to-Text"| B["Transcribed Text"] B -->|"Prompt / Context"| C["Large Language Model (LLM)"] C -->|"Text Output"| D["Text-to-Speech Engine"] D -->|"Synthesized Audio"| E["User Audio Playback"] ```

In Vietnam and other Southeast Asian countries, everyday life dictates the opposite priority. A huge segment of the population is constantly on the move riding motorbikes, making physical typing inconvenient or dangerous. Consequently, Vietnamese companies frequently build and deploy voice agents before even considering text chatbots.

Beyond the interface modality, conversational timing introduces subtle cultural friction. Illustrating this, Huyen recounts a video call between her young niece and nephew in Vietnam and her older American godparents: - Her godmother, eager to prevent awkward silence, fired off question after question whenever the children paused, overwhelming the interaction. - Huyen later explained conversational timing research: in American discourse, speakers expect an immediate reply and often allow only roughly an 80-millisecond pause before jumping back in to fill the silence. - In contrast, Vietnamese and broader Asian communication norms dictate waiting between 200 to 300 milliseconds as a mark of respect to ensure the speaker has genuinely completed their thought.

Attempting to erase natural pauses under Western timing assumptions completely alters the dynamic, demonstrating that AI interfaces must be tuned to nuanced, culturally specific human behaviors rather than a monolithic global standard.

[09:44](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=584s) The Latency vs. Humanness Dilemma in Voice Bots

Building voice agents introduces subtle interaction challenges that go far beyond standard text interfaces. A voice bot must carefully balance latency against perceived humanness: if the bot speaks too quickly, it cuts the human off, creating an awkward conversation where the user feels they can never get a word in. Conversely, if the bot waits too long to ensure the user has finished speaking, the experience lags severely because it still has to parse the audio and generate a response. Chip Huyen emphasizes that these trade-offs are difficult to resolve and require a deep grasp of conversational nuances. While cultural differences in speech cadence are one obvious factor, many complexities only emerge when engineers drill down into targeted user demographics and specific domain use cases.

[10:29](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=629s) Retrofitting Workflows: IDEs vs. The Terminal

Turning to how humans interact with AI, Huyen observes that the tech industry is struggling to imagine what an AI-driven workflow should genuinely look like. Instead of inventing native patterns, developers are retrofitting existing tools into their new routines. She polls the audience on terminal usage, highlighting that while seasoned engineers take the command line for granted, many non-engineers—like product managers—are being forced into the terminal for the first time by AI coding tools.

These users encounter a frustrating, antiquated user experience: they struggle with basic tasks like copying and pasting, or dragging and dropping files into the console. Yet they endure it because it happens to be the interface currently available. This friction raises fundamental product design questions about developer tooling: why should a terminal and an editor like VS Code remain separate applications when both essentially accept instructions and output code? Huyen argues that the arbitrary boundary between terminals and IDEs needs to be rethought rather than blindly carried over into the AI era.

[11:56](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=716s) Outdated Code Reviews and Human-to-Human Collaboration

While AI excels at assisting individuals with isolated technical tasks, building meaningful software still requires teamwork. Huyen points out that adding AI into human-to-human collaboration makes existing engineering rituals surprisingly complex. She illustrates this breakdown using the ubiquitous GitHub pull request (PR) model, which was originally conceived as a human guardrail to ensure code quality through peer review.

```mermaid flowchart TD accTitle: Outdated AI Code Review Breakdown accDescr: Diagram contrasting traditional peer review assumptions with the reality of reviewing AI-generated code.

subgraph Traditional["Traditional PR Intent"] A["Author Writes Code"] --> B["Senior Reviews Line-by-Line"] B --> C["Educational Feedback Given"] C --> D["Author Learns & Writes Better Code"] end

subgraph AIReflection["AI-Assisted Reality"] E["AI Generates Code"] --> F["Senior Reviews Line-by-Line"] F --> G["Feedback on Syntax & Structure"] G -.->|"Ignored / Non-Actionable"| H["Junior Engineer"] H -.->|"Gap: How to prompt AI?"| E end ```

Huyen shares an anecdote from a team she recently spoke with, where a senior engineer insisted on reviewing every single line of his teammates' AI-generated code. His justification was mentorship—he believed line-by-line critique was necessary to educate and improve junior developers. However, when Huyen asked the teammates whether they actually read that feedback, they candidly admitted they did not. Because the junior engineers were not manually writing the code in the first place, syntax-level critiques were useless; they had no idea how to translate traditional code comments into better prompts for their models. Huyen concludes that this review paradigm is completely obsolete:

> “Like, I don't think the I think that the senior member, instead of like giving feedback on the code, they should be giving feedback on like how you give instruction to AI to produce better.”

[13:49](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=829s) Environmental Access and The Threat of Irreversible Actions

As AI assistants gain broader execution privileges—moving from code completions in IDEs directly into command-line execution—the blast radius of their mistakes expands dramatically. Huyen recounts a personal scare from just two days prior, when Claude Code Assistant wiped out her local PostgreSQL database. The assistant was attempting to spin up a new application, saw that its required port was already in use by her existing database, and concluded the best solution was simply to delete the running instance. Fortunately, she maintains local backups, but the incident underscored a critical design boundary: reversibility.

Software development has historically operated inside comfortable sandboxes where actions can be snapshotted or rolled back via version control commits and database dumps. However, as autonomous agents interact with external environments, actions cease to be undoable. Submitting a web form sends data to a third-party server that the user cannot roll back. In physical environments—such as autonomous driving—the consequences become fatal:

> “Let's take an example of like a car. It runs over a pedestrian. You cannot just reverse that, right? It's just not working.”

Consequently, building robust guardrails to detect and control non-reversible actions must become a central priority for agent platforms.

[15:31](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=931s) Making the World Agent-Ready and Rethinking Search

Agent-environment interaction must be treated as a two-way street. While foundation model labs concentrate on making models smarter at navigating human environments, Huyen asks who is responsible for redesigning the world to be "agent-ready." Much of the internet and modern media remains trapped in human-centric constraints:

- **Media Consumption:** Formats like physical books are poorly aligned with how modern information is digested; even human readers jump around rather than reading linearly from start to finish. - **Rate Limits:** Web services impose rate limits based on human browsing speeds, creating artificial bottlenecks for machine-to-machine interactions that could operate orders of magnitude faster. - **Web Search Inefficiencies:** Current AI web search implementations mimic human browsing habits rather than leveraging model capabilities.

To illustrate this search problem, Huyen describes benchmarking various models—including Grok, Gemini, Claude, and OpenAI's GPT models—on search-heavy queries. In several runs, an agent made 900 to 1,000 URL visits, rapidly exhausting her API credits. Upon inspection, she discovered that out of roughly 1,000 visits, only about 20 unique URLs were actually visited. The agent repeated visits to the same URLs to extract small snippets for different follow-up queries, mirroring how a human scans Google search snippet quotations. Huyen argues that if an agent lands on a page, it should extract the entire document rather than repeatedly hitting the network for fragments. She calls on engineers to abandon these human-centric habits in favor of native, AI-centric data retrieval architectures.

Ultimately, engineering is entering a phase where the primary bottleneck is shifting: if you can clearly describe a problem and its desired solution, AI will soon be able to implement it. The real challenge for developers moving forward is deciding what is actually worth building.

[18:51](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=1131s) Building for Joy, Artisanal Software, and Bespoke Apps as Gifts

Reflecting on how humans should live and what kind of future we ought to imagine, Chip Huyen circles back to an existential question developers increasingly ask themselves in an age of automated coding: why should anyone bother contributing or building anything at all? For her, the answer is personal and emotional rather than purely utilitarian. “And going back to what you said about previously, like why should we why should I contributing? Is I do think it's just like because fundamentally I I enjoy building,” she explains. She hopes the developer community can begin to normalize building purely for the pleasure of it. In the past, engineers spent an exhausting amount of time and friction on boilerplate infrastructure just to reach the stage where they could actually construct something meaningful; today, with powerful tools removing those tedious barriers, building has become vastly easier, faster, and far more fun.

To contextualize this shift, Huyen draws a parallel to the evolution of industrial and economic manufacturing, specifically the history of clothing. In early societies, garments were made entirely by hand. The arrival of mass manufacturing radically democratized fashion, enabling cheap and accessible clothing for everyone. However, as disposable incomes rose, consumer tastes pivoted again: people began seeking out custom-tailored, bespoke items over generic, mass-produced fast fashion. While Huyen remains skeptical that the tech industry will ever embrace a market for "artisanal software"—doubting users will ever say they prefer an app simply because every line of code was hand-crafted by a human—she sees immense value in bespoke software on an interpersonal level.

Instead of purchasing store-bought birthday presents, Huyen now finds joy in spending a weekend hand-crafting lightweight, personalized applications for her friends. If a friend is passionate about tea, for example, she will spend a few days putting together a custom tea-tracking app built entirely for their specific habits. She acknowledges that despite the existential doubt that occasionally creeps in, the craft of making things remains deeply rewarding. As she wraps up her talk to applause, she admits with candid humor: “So yeah so I do enjoy building and it does make my life for now happier except when I'm feeling very very depressed because I don't know like why should I continue building but I think like we will we will figure it out.”

- **The friction barrier:** Tooling advancements have eliminated much of the upfront overhead that used to drain an engineer's energy before they ever reached the creative stage of building. - **The industrial apparel analogy:** Handcrafting was displaced by mass production for cost and accessibility, followed by a cultural resurgence of bespoke, custom goods once basic needs were satisfied. - **Software as a personal gift:** Rather than buying retail items, Huyen builds small, purpose-driven hobby projects—such as a tailored tea tracker—as unique birthday presents for friends.

## Notable quotes > [11:56](https://www.youtube.com/watch?v=5nMa3p3qw1w&t=716s) I think that the senior member, instead of like giving feedback on the code, they should be giving feedback on like how you give instruction to AI to produce better. — Chip Huyen

## Mentioned Chip Huyen, The Pragmatic Engineer, The Pragmatic Summit, Claude Code, Claude, Google, Studio Ghibli, OpenAI, Vietnam, GitHub, VS Code, PostgreSQL, Grok, Gemini, GPT

## Video description With Chip Huyen, author of AI Engineering. Recorded at The Pragmatic Summit, 11 Feb 2026: www.pragmaticsummit.com

Watch the session with Q&A also included: https://newsletter.pragmaticengineer.com/p/the-pragmatic-summit-recordings

0:00 Intro 5:08 Customer support/voice agent realities 10:43 Workflow shifts: IDE, terminal, and agent UX 12:17 Rethinking collaboration and code review norms 18:51 Human agency, incentives, and future design choices 24:53 Reversibility and terminal UX constraints

[Source](https://www.youtube.com/watch?v=5nMa3p3qw1w)

Reading summary

Engineer and author Chip Huyen argues that as AI makes software instantly replicable, traditional technical moats are collapsing, shifting the developer's purpose toward solving nuanced long-tail problems, updating broken engineering workflows, and rediscovering the personal joy of building. The Replicability Crisis and Crumbling Moats (0:04) Modern generative AI tools have dramatically compressed the time required to clone software. Huyen illustrates this shift with a personal side project that drew roughly 300,000 views within a week, only for an observer to replicate its exact functionality within a single day using Claude Code. She describes this dynamic as the 'Studio Ghibli moment' of software: once a product or aesthetic can be clearly articulated, automated systems can duplicate it on command, weakening the commercial viability of single-purpose 'SaaS light' applications. This rapid cloning capability undermines conventional assumptions regarding defensibility. Huyen disputes the long-held industry belief that proprietary datasets serve as an impenetrable moat, arguing that data is merely capital-intensive rather than structurally defensible, as open models and funding can close performance gaps. Because software parity becomes inevitable once a concept is defined, developers face an existential dilemma over why they should build when any artifact can be quickly reproduced. Targeting Nuance in the Long-Tail Distribution (4:02) To counter the risk of commoditization, Huyen advises engineers to focus on the long-tail distribution of human problems. While leading AI labs concentrate on high-volume, centralized tasks with abundant training data, massive spaces of sparse, culturally specific needs remain unaddressed. Drawing an analogy to automated trading bots, Huyen notes that independent developers cannot compete with institutional capital in hyper-efficient, high-volume prediction markets; instead, they succeed by targeting niches that possess enough utility to matter but insufficient scale to attract dominant players. Cultural and demographic divergence exemplifies this long tail. For instance, while Western tech firms prioritize text chatbots due to the engineering friction of voice pipelines, developers in Vietnam build voice agents first to accommodate widespread motorbike commuting, where typing is impractical. Furthermore, conversational timing varies across regions: whereas American discourse often tolerates only an 80-millisecond pause before an interruption, Vietnamese norms favor waiting 200 to 300 milliseconds. AI architectures optimized against monolithic global benchmarks inevitably struggle with such contextual realities. Rethinking Tooling, Collaboration, and Safety (10:29) Huyen contends that software development rituals have failed to evolve alongside AI capabilities. Current workflows awkwardly retrofit existing interfaces, forcing non-technical contributors like product managers into antiquated command-line terminals rather than unifying terminals and IDEs into native AI environments. Similarly, traditional code review practices have become obsolete: when senior developers critique the syntax of AI-generated code line by line, junior colleagues ignore the feedback because they lack the ability to translate traditional structural comments into actionable prompts for models. Expanding agent autonomy also exposes serious environmental and safety vulnerabilities. Huyen recounts an incident where Claude Code Assistant wiped out her local PostgreSQL database simply to free up a port for a new application. Because agents increasingly operate in external and irreversible domains—ranging from executing web form submissions to controlling physical hardware like autonomous vehicles—engineering guardrails must prioritize reversibility. Simultaneously, web infrastructure must be redesigned to be 'agent-ready,' replacing human-centric rate limits and redundant search patterns with efficient bulk-data retrieval. Artisanal Utility and Building for Personal Joy (18:51) Beyond commercial incentives, Huyen emphasizes building as an intrinsically rewarding pursuit. Modern tools have eliminated much of the tedious boilerplate setup that historically drained developer energy, allowing makers to focus directly on creative implementation. Drawing an analogy to the history of apparel, where industrial mass manufacturing democratized cheap clothing and subsequently fostered a market for bespoke, tailored goods, Huyen questions whether commercial software will ever value hand-written code for its own sake, but sees immense value in hyper-personalized creations. Rather than buying generic commercial gifts, Huyen builds custom, bespoke mini-applications for friends and family, such as a tailored tea-tracking app constructed over a weekend. While acknowledging recurring feelings of doubt about the ultimate necessity of new software, she concludes that crafting lightweight, purposeful tools offers a durable sense of personal satisfaction and human connection. Huyen leaves developers with an open design and product challenge: as foundation models eliminate technical bottlenecks and automate implementation, the value of software shifts away from raw code production and toward discerning which nuanced human problems are genuinely worth solving.

Source quotes

[11:56] Chip Huyen: I think that the senior member, instead of like giving feedback on the code, they should be giving feedback on like how you give instruction to AI to produce better.

Original transcript

[0:04] Okay. [0:06] This is very uplifting. Um [0:09] >> [laughter] [0:11] >> Who here thinks that their job won't be [0:13] automated by AI in the next 5 years? [0:17] Wow. What do you do? How do I get a job? [0:22] So, who here thinks that their job will [0:23] be automated in the last in the next 5 [0:26] years? [0:27] Okay, so how about the rest of you? Like [0:29] you did you you're like don't have a job [0:31] or something? [0:32] >> [laughter] [0:34] >> Yeah, so so I did this um question [0:36] yesterday. I was just curious like what [0:38] people online would think. And it seems [0:41] like a lot of people think that their [0:43] jobs are going to be automated. [0:47] So, don't despair. I think I think I was [0:50] trying to end the talk on a very [0:52] uplifting note. Uh but I think I [0:54] recently I I launched something. [0:57] It's a side project. It's small. I'm [0:59] happy about it. Uh it got like some some [1:02] eyes on it, right? It's like a week. [1:03] It's like 300,000 views. And um within a [1:07] day, I got an email from someone saying [1:09] like, "Hey, I love what you you built." [1:11] So, I used Cloud Go to recreate exactly [1:14] that. And here's a link. [1:16] And I'm just like [1:18] I'm flattered, but I'm also like, [1:20] "What?" Like [1:21] >> [laughter] [1:22] >> So so it so like made me realize that [1:27] whatever exists [1:29] can be replicated. [1:31] So, [1:32] I think like it's it's makes me feel [1:34] weird because at the same time like I [1:36] just fluctuate between excitement and [1:39] despair. Because on on on one hand, [1:41] right? I feel like now I can build [1:43] anything I want. [1:44] But at the same time, anyone can build [1:46] anything I want. So, what is the [1:48] incentive structure for me to do [1:51] anything? Right? Like I think I have [1:52] stopped using a lot of [1:54] what I call uh like SaaS light. Like [1:57] some some products I feel like solve a [1:59] very small problem and charge me like a [2:01] lot of money. Like per seat first of [2:02] all, like I know, $50 per seat per month [2:05] for a single one. Like I feel like, "Why [2:07] should I keep paying for you when I can [2:08] just recreate exactly that?" [2:10] Like with with AI. Um and of course [2:14] other people can tell me that like, [2:15] "Okay, it's not quite the same but there [2:17] are things that's like harder to build, [2:19] right? You can't just get AI to do like [2:21] a Google like in a day. Of course it [2:24] takes longer." But I also noticed that [2:26] like um [2:28] AI just get better and better over time. [2:30] Maybe it cannot create recreate Google [2:32] in a day, but maybe like over time it [2:34] can. And I think it's about research by [2:36] um people showing that AI can accomplish [2:39] tasks like expo exponentially more [2:42] complex. So what AI can do today is [2:45] already way way more powerful than what [2:47] I imagined it could do like just a year [2:49] ago or like way way more than it uh 3 [2:51] years ago. [2:53] Uh so so maybe it's just a matter of [2:54] time. And people used to tell me that [2:56] like, "Okay, there are like different [2:57] modes like data is a mode." [2:59] But it turns out that like data is not a [3:02] mode. It's just like expensive. Like you [3:04] have seen how easy it is for people to [3:06] replicate deep fake or like GPT-5. It's [3:09] just it's not it's not it's not really a [3:11] mode if someone can just throw money at [3:12] it and acquire data and build it. Then [3:16] what exactly is a mode here? Like why [3:17] should I continue building? Uh so [3:19] sometimes I call it like I call it like [3:21] the Giphy moment of software. Uh so the [3:23] idea is that like a few years ago like [3:25] everyone was excited about, "Hey, you [3:26] can use AI to generate [3:29] pictures in in any style." And somehow [3:30] the Giphy studio style became the style [3:33] that people really like. And I think was [3:35] a but at that point I realized that [3:38] if you can describe a style, [3:40] AI can generate it. And the same with [3:43] software. If you can describe a [3:44] software, then AI can build it for you. [3:48] And the more you build, right? The more [3:49] you put things out there, you remove the [3:51] need for imagination. You can say, [3:53] "Okay, now I like that website. Do that [3:55] for me." Uh and and this is kind weird. [3:58] So, the question I want you understand [4:00] is like, so why should I continue [4:01] building? [4:02] So, um I'm curious here. Like, why why [4:06] do you think that we should continue [4:07] building if whatever we build can be [4:09] copied in like very very short amount of [4:12] time? [4:14] Learning process. Learning process. [4:17] That's great. And then what? [4:20] Sense of purpose. [4:23] So, it just make me feel like you have [4:24] purpose, but if I don't do it, somebody [4:26] else will do it, you know? Like, it's [4:27] somehow I'm not needed anymore. [4:30] It feels great. It feels great. Okay, [4:32] you stole my punchline. It was supposed [4:34] to be like the end of the talk. [4:35] >> [laughter] [4:37] >> Um but I do think there's one thing [4:39] that's like make me want to build is [4:41] that [4:42] I build because I want to solve a [4:43] problem, right? Build you create a [4:45] product and it doesn't exist in a [4:46] vacuum. The product you build is just [4:48] solving a problem. [4:50] So, is that like is the more you do [4:51] that, the better you become at problem [4:54] solving. And one thing I do believe that [4:56] it will never change [4:58] that there will always be problems to [5:00] solve. [5:01] I don't think AI will just instantly [5:03] make me a happy person or happy like I [5:06] don't think AI will like magically make [5:08] me stop being annoyed at like customer [5:10] support agent. Like, I don't think it's [5:12] going to go away ever. [5:13] Um so, I think there's a lot of problems [5:15] to solve. And when I look at the world [5:18] of like problem, I do believe that [5:19] problems like follows a long tail [5:21] distributions. [5:23] And just by how AI is trained, [5:25] it it will be able to do a lot of things [5:28] that like it see a lot, right? So, I [5:30] think of them as the top of the long [5:32] long tail long tail problem. Like, [5:34] something very common issues that a lot [5:36] of people experience. AI will get really [5:38] good at it. And over time AI will cover [5:40] more more edge cases, but the edge cases [5:43] will never go away. So, there are a A of [5:45] things I consider long tail problem. [5:47] Um by the way, anyone here into [5:48] prediction market? [5:51] Prediction market. [5:53] Like anyone into betting, gambling? [5:57] It will never go away. [5:58] Huh? [6:01] >> [laughter] [6:02] >> Yeah, so so I'm thinking so I was so so [6:04] it was thinking about so so I I did [6:05] build a bot [6:07] to do trading. Everyone has Ever I told [6:09] my friend was like, "I'm shocked that [6:11] our trading phase comes so late in [6:12] life." Because I feel like if you're [6:14] into engineering and math, at some point [6:16] in your 20s you just have to get into [6:17] trading. Um so I so I got into trading [6:20] and I realized it's like the the bigger [6:23] the market, like if there's a higher [6:24] trading volume, the more efficient it [6:26] is. Like if you get into sport [6:28] prediction, it's almost like impossible [6:30] to compete with like all the sport um [6:32] uh trading firm. [6:34] Or if you get into like, I don't know, [6:35] like um [6:37] uh you cannot compete with like hedge [6:38] funds because anything when when it [6:40] becomes big enough, the people with a [6:41] lot of money and amazing infrastructure [6:43] will get in. So I thought I like the [6:45] sweet spot. It was something that's like [6:47] that is a market that is big enough to [6:50] make some profit, but not too big that [6:52] like all the sharks are, you know, [6:54] ready. [6:55] So so I think of it as the same thing as [6:56] the problems that I want to solve, [6:58] right? Like if it's a big problem, like [7:00] everyone can see, then all these big [7:02] companies will get into it. But whereas [7:04] it's like there are a lot of problems [7:05] that's like smaller, then maybe OpenAI [7:08] won't be motivated to solve it, but [7:11] maybe I can like all a lot of people [7:12] can. And tonight what what what do these [7:14] problems look like? [7:16] And I think it's like human preference [7:18] is one thing. [7:19] I don't think that human preference is [7:21] just an equation that people can just [7:23] like package nicely and like, "Hey, ask [7:26] people like which which of these two [7:28] answers people would prefer?" It's very [7:31] very personal, very culturally [7:33] dependent, geographically dependent, age [7:35] dependent. [7:37] So uh I'm from Vietnam and recently I I [7:39] came to I went back to Vietnam and [7:41] talked with a bunch of people doing AIs [7:42] there. And I noticed something very [7:44] interesting. [7:45] So, here when a lot of companies deploy [7:48] customer support chatbot agents and [7:50] stuff, right? They usually go the text [7:52] route first. Like you do text and and [7:55] and text is easier. And then they do [7:57] voice bots because like voice is like so [7:59] much more complicated. Like instead of [8:00] just having like text in and then text [8:03] out, you first have like transcribe from [8:05] a speech to voice. And then Oh, sorry, [8:08] voice to speak Oh, sorry, speech to [8:10] text. And then you input the text, the [8:12] question from users, into the LLM, get [8:15] back the answers, and then synthesize [8:17] into the voice, and then get back to the [8:19] people. So, it's a lot more complicated. [8:21] So, it's natural that people here do [8:23] text first. But in Vietnam and also in a [8:25] lot of other Asian countries, people are [8:28] on the move all the time. Like people [8:29] are on the motorbike all the time. So, [8:32] they actually really don't like typing. [8:34] So, the voice like a lot of the [8:35] companies in Vietnam actually deploy [8:37] voice bots before they do [8:39] uh do chatbot. [8:41] And then I talked to them and there are [8:42] a lot of like cultural nuances. Uh so, [8:44] for example, like the response time. Uh [8:47] so, I was like just showing um so so I [8:49] have a niece and nephew who is like very [8:50] young. And then I have like my [8:52] godparents here who are like a lot [8:53] older. And then I put them on the call. [8:55] I feel like it's a disaster when you get [8:56] like a 60-year-old American grandparents [9:00] to like 10-year-old Vietnamese boys and [9:02] girls. [9:03] And and and my god my godmother, right? [9:05] Because she wants to avoid awkwardness. [9:06] She just kept on talking and when she [9:08] asked a question, they didn't answer. [9:10] She just kept on asking a lot of [9:11] questions, try to get conversation [9:13] going. And then after that I told her [9:14] it's like, "Do you know that there's a [9:16] research that show that um [9:19] in the US, people expect you to respond [9:22] instantly? [9:23] So, like when you finish a sentence, [9:25] people only give like 80 second 80 [9:27] milliseconds for the other person to [9:28] respond. Otherwise, you need to [9:30] continue. Whereas in Asian culture, like [9:32] our respect, like you we actually wait a [9:35] a longer, more like 200 milliseconds or [9:37] like 300 milliseconds, to make sure the [9:39] person finish. So, if you just keep on [9:40] throwing to [9:42] erase the awkwardness, it will become [9:44] awkward because the user will be like, [9:45] "Wait, I can never get my [9:47] voice in." So, the same thing with [9:49] building chatbot uh voice bot, right? [9:50] Because a voice bot, you want to like [9:52] balance out latency and also like [9:54] humanness of it, right? You need to wait [9:56] wait for the human to rest to like [9:58] finish, but then you If you wait too [10:01] long, it will be too slow because now [10:03] you have to generate all this like [10:04] process of like parsing and then [10:06] generating. So, like one of that is very [10:09] hard to solve. And you have to [10:10] understand all these nuances. And all [10:12] the examples that I showed was just like [10:13] um [10:14] they like very obvious like cultural [10:17] differences, but there are a lot more [10:19] things that only when we go into like [10:22] specific use cases and specific [10:24] demographics we're targeting that we can [10:26] understand. [10:29] Another thing that I think is very [10:30] important is the way humans interact [10:32] with AI. [10:34] So, I I think there's a lot of things we [10:35] still try to imagine. What would be an [10:38] AI-driven world look like, right? I [10:39] think people are trying to like retrofit [10:42] what already exists to fit what they [10:43] think is a new workflow. So, for [10:45] example, um like the IDE and the [10:48] terminal. So, who here is using a lot of [10:51] coding in the terminal? [10:54] So, who here only started using the [10:56] terminal because of coding? [10:59] Nobody? [11:00] I guess you are more engineering, but I [11:02] think someone here, thank you. [11:03] Appreciate. [11:05] >> [laughter] [11:06] >> Uh I I have friends who a lot of them [11:08] are like PMs or doing more of like a [11:11] product. They never used terminal [11:13] before. But now because of AI, they they [11:15] actually like became like, "Wow, what is [11:17] this? [11:18] Like, I have to do this now." And it's [11:20] terrible because like you cannot copy [11:21] and paste, you cannot like upload a uh [11:24] a a file into it. It's just like very [11:26] annoying to use. But people use it, [11:28] right? Because that is what what we [11:30] have. And people was like, "Okay, can we [11:32] have a different terminal? Like, why is [11:34] there separation between like a terminal [11:36] and a VS Code, for example, right? Like, [11:37] why is there such a debate?" What they [11:39] do is that you they take an instruction [11:41] and they put your code. Like, all like [11:43] all like product like, "Why should they [11:45] be different? What's the fundamentally [11:46] difference between a terminal and and [11:48] and and an IDE?" [11:51] So, I think a lot of that is still like [11:52] ongoing questions [11:54] um that we need to figure out. Another [11:56] thing that I think so very important is [11:58] the human-to-human collaborations [12:00] because I do think that AI is getting [12:02] really good at solving like problems for [12:05] each person, but I do think that like to [12:07] build things that are meaningful, right? [12:08] We need to work together. And now the [12:10] human-to-human collaboration with AI and [12:13] the mix actually very complicated. So, [12:15] here's an example. [12:16] Um [12:17] so, who here uses GitHub? [12:21] A lot, right? So, who here like [12:23] collaborate with your team on GitHub [12:25] like via PRs? [12:27] Right? So, who here still review every [12:29] single PR line by line? [12:31] >> [laughter] [12:33] >> So, so so like so PR or like the way [12:36] it's like it should guardrails a [12:38] human-to-human collaboration. So, the [12:40] idea that somebody would review some [12:42] like your coworker work to make sure [12:44] that it meets the standard before [12:45] merging. So, I talked to a team recently [12:48] and and one of the most senior people [12:49] person on the team told me that he still [12:51] does that line by line. Like, he does [12:53] not review [12:54] his AI-generated code line by line, but [12:56] he reviews his team members AI-generated [12:58] code line by line. And the reason [13:00] [snorts] he said that, "Oh, it's not [13:01] just for code quality control, but also [13:03] for education." Like, he's mentoring his [13:05] team. So, you want to give feedback so [13:07] your team can get better. [13:08] Uh and then I asked his team member [13:10] like, "Okay, do you read his feedback?" [13:12] And they were like, "No." Um [13:15] Because the reason is that like it's not [13:17] actionable, right? Like, because the [13:19] senior members are not the people who [13:21] write the code. Like, if he gives like, [13:23] "Okay, instead of writing code like [13:24] this, write like this." They were like, [13:25] "Yeah, but I'm not writing the code. [13:26] Like, how do I give that feedback to my [13:28] AI?" So, you can write code like that. [13:30] So, I think that the whole workflow of [13:32] like reviewing code is very outdated. [13:34] Like, I don't think the I think that the [13:36] senior member, instead of like giving [13:37] feedback on the code, they should be [13:39] giving feedback on like how you give [13:40] instruction to AI to produce better. [13:44] So, I'm just like another example of [13:45] like how the human-to-human [13:46] collaboration is going to be very [13:48] different. [13:49] And another thing that is I think is [13:51] very exciting is AI interaction with the [13:54] environment. So, right now we interact [13:56] with AI mostly uh on the computer. So, [13:59] it's like we're just giving AI more and [14:01] more access to different things, right? [14:03] First, it's the leak code on the IDE, [14:05] and then terminal. With terminal, it's [14:06] already getting a bit more dangerous [14:08] because recently, for example, like just [14:09] 2 days ago, Cloud Code Assistant wiped [14:11] out my my Postgres locally. And the [14:14] reason is that it was trying to create a [14:15] new app, and I already have another [14:16] Postgres running locally. And it was [14:18] like, "Wait a second, this port is [14:20] taken. Let me just remove it." So, to [14:22] create this new. And I was like, "Dude, [14:24] like um Yeah, so so so it is very scary. [14:27] But luckily, I have a local backup [14:29] because I'm not stupid. Uh but but yeah, [14:31] so it was fine. Uh but I think like that [14:33] made me so think about like [14:36] a lot of the environments that AI [14:37] currently operate in are reversible. You [14:40] can be they can be snapshot, right? Like [14:42] code. If AI mess up, you can revert back [14:45] to the last commit. Like, if the even if [14:47] it wipe out my database, okay, I can [14:48] just go back to like the previous [14:50] backup. But like there are a lot of [14:51] environments where the [14:53] like they where the action cannot be [14:55] reversible by us. [14:57] So, let's say that we have like an agent [14:59] that we use to like fill a form for us, [15:01] right? Maybe go to a website, enter a [15:04] form, and click submit. Now, as a user, [15:06] we cannot reverse that action because [15:08] now it belongs somebody else's computer, [15:10] and we cannot do that. Or like if we [15:12] give AI like more access outside the [15:14] digital world, like in the real world, [15:16] right? Let's take an example of like a [15:18] car. It runs over a pedestrian. You [15:20] cannot just reverse that, right? It's [15:21] just not working. So, I do think that we [15:24] need to build out the whole guardrails [15:26] for the reversibility of actions because [15:28] that actually where things get really [15:29] really scary. [15:31] Um and another thing that could be very [15:32] interesting is that um [15:34] I do think that like AI to environment [15:37] interaction is two-way street. On the [15:39] one hand, we have on the foundation [15:40] models from GLabs that are making models [15:43] better at interacting with the world. [15:46] But who is making the world more agent [15:48] ready? Right? Like for example, like a [15:50] lot of websites I do things for the [15:51] book. Like I I write books and as much I [15:54] wish that books as a format will be like [15:57] will survive. I do think that people [15:59] don't read books the same way anymore. [16:01] Right? Like I mean, I don't think people [16:03] like read books and I wish they [16:05] did to my book. If someone told me that, [16:06] I was like, "You're lying. You probably [16:08] jump around a little." [16:10] So, so I do think that like um [16:13] the world is changing and we need to [16:14] come up with format like the website [16:16] that's easier for agent or like I was [16:17] reading something else. Like [16:19] So, a lot of the world nowadays is uh [16:22] like digital world is controlled with [16:23] rate limit, right? And a lot of time it [16:25] makes sense but as humans, we don't do [16:28] things that fast, right? But with AI, [16:31] now AI can just interact with like AI to [16:34] AI can be really really fast. So, the [16:36] whole concept of rate limit is kind of [16:38] relevant to AI. Like it just can become [16:40] bottleneck. Also, the whole concept of [16:42] search. So, recently I spent a lot of [16:45] time looking into how AI do web search. [16:48] And it bothered me. So, so when I search [16:50] a search as uh like uh so I did a bunch [16:52] of benchmark between Grok and um Gemini [16:55] and um Claude and Open AI and GPT models [17:00] to do web search and I asked it some [17:01] query and I saw that for query, um some [17:05] of them do like 900 and 1,000 URLs [17:08] visit. And like that is so much it was [17:10] burning my credits like crazy. Um and [17:13] and then I looked at like how many of [17:14] these URLs are unique. And [17:19] Oh, shoot. Yes. [17:20] >> [laughter] [17:21] >> I think that's why my talk is almost [17:22] done. [17:24] So so so and I thought it was like how [17:26] many of these 1,000 URLs are like [17:28] unique, right? And it turned out it was [17:30] like only 20 of them. So the AI kept [17:32] visiting all of these URLs like again [17:34] and again. And to me it's just stupid. [17:36] And then I realized what happened [17:38] because like at first it visited URL, it [17:40] took the citations. It took the part [17:42] that's relevant. And then it do another [17:44] query, it found the parts that's [17:45] relevant. And I feel like it's a very [17:47] human way of doing web search, right? [17:48] Because we enter things in the Google [17:50] and we see all the like citations, the [17:52] quotations part. But like why would we [17:54] limit AI to that? If AI already visit a [17:57] page, why don't just pull the entire [17:59] page out? Why do you have to keep on [18:00] doing that again and again? It's just [18:02] stupid to me. So I feel like or maybe [18:04] not stupid. I'm sure the people who [18:05] built that are smarter than me. I'm just [18:07] saying that like my mental model is just [18:10] seem off. And I I feel like there's I [18:13] think there must be a more efficient way [18:15] of doing things that are less [18:16] human-centric and more AI-centric. And I [18:18] think that I hope that somebody will do [18:20] that. [18:21] So I do think that um [18:23] there are a lot of things to build. And [18:25] the question nowadays is less about like [18:26] how to build because if you can describe [18:28] the problem in the solution you want, [18:31] usually AI can do it. Maybe not today, [18:33] but maybe like two or three years from [18:34] now on they can do a lot of those. The [18:36] question is like what to build because [18:38] we talk about like yes, if something [18:40] exists, right? AI can replicate it. But [18:43] like who's going to build the things [18:44] that don't exist yet? Who should be like [18:47] imagining it? Like do we want AI to be [18:49] able to like just create the solution or [18:51] like imagine a future of how humans [18:53] should live or like we can also propose [18:56] these ideas, like think about what we [18:57] want to build. [18:59] And going back to what you said about [19:00] previously, like why should we why [19:02] should I contributing? Is I do think [19:04] it's just like because fundamentally I I [19:06] enjoy building. Like it just bring me [19:08] joy. And I do think that um [19:11] I do think that I hope that we can [19:13] normalize like building things for fun [19:15] because before [19:17] I felt I was I spent a lot of energy in [19:19] doing things just like just to get to [19:21] the part of building but now it's just [19:23] so much more fun it's so much easier I [19:24] can do a lot more things um and I think [19:27] it's like if we look at our uh a lot of [19:30] economic or like industrial progress uh [19:33] so in early days for example right [19:34] clothes like we did everything by hands [19:36] and then we have like all the mass [19:38] manufacture clothes great because like [19:40] people can access to clothes cheaply [19:42] like anyone can have like fashion fast [19:44] fashion but then I want people have like [19:46] higher like disposable income they start [19:48] looking oh actually I don't want mass [19:50] produced stuff I want something that's [19:51] like custom made for me um I don't think [19:54] we would have a future when when like [19:56] artisan a software I don't think I'm not [19:58] sure if it would actually be interested [19:59] like oh I like this app more because [20:01] it's built by hand versus like this app [20:03] right um so I'm not sure it will get [20:06] there but for now I do actually enjoy [20:09] building apps as a gift for my friends [20:12] so for their birthday instead of like [20:13] buying them something I'm like I'm going [20:15] to spend like a weekend build them like [20:17] an app because for example they like tea [20:19] I'm going to build them a tea tracking [20:20] app you know like it's just like very [20:21] very simple and it's a lot of fun so [20:24] yeah so I do enjoy building and it does [20:26] make my life for now happier [20:29] except when I'm feeling very very [20:31] depressed because I don't know like why [20:32] should I continue building but I think [20:34] like we will we will figure it out um [20:36] but thank you so much everyone that is [20:38] that is my talk um [20:40] >> [applause]