YouTube Digest

English

Based on "The One Rule That Separates Real AI Agents from Demos | Yutori, Abhishek Das" from EO Watch the original video

Beyond the Demo: The Uncompromising Rule for Real AI Agents

In the burgeoning landscape of artificial intelligence, the promise of autonomous agents capable of navigating the web on our behalf is a compelling vision. Yet, for many, this promise often crumbles upon the first interaction. You've likely encountered them: the "hundred different agent products" claiming to "do anything on the web," only to falter on the simplest tasks. This pervasive unreliability is a significant pain point for Abhishek Das, co-founder and co-CEO of Yutori. Das is on a mission to redefine what an AI agent truly means, guided by a singular, uncompromising rule: "If it's not good enough to work on the first try, it's not good enough."

The Pitfalls of Imperfect Automation

The current state of AI agents, while exciting in concept, often falls short in execution. Das explains the fundamental mathematical challenge inherent in complex workflows: "If we think of a 10-step, 20-step or 50-step workflow, even if the accuracy at each step is like 90%, the 10% error rate compounds very quickly." This means a seemingly robust 90% accuracy per step can quickly lead to an overall success rate that is "quite low" for any multi-stage task. The technology, he asserts, is simply "not there yet to do long horizon workflows."

What truly irks Das, however, is not just the technical limitation, but the industry's growing "tolerance for non-determinism and low reliability in shipping products." He pushes back vehemently against this "normalization of slop," especially for agentic products. For Das, a product that requires multiple attempts to succeed isn't a product; it's a frustrating prototype. He firmly believes that "if it's not good enough to work on the first try, it's not good enough." This "first-try imperative" is the bedrock of Yutori's philosophy.

A Vision Forged in Curiosity and Code

Abhishek Das’s journey to co-founding Yutori is rooted in a deep-seated curiosity and a rebellious spirit. Hailing from a family of doctors, his aversion to blood quickly steered him away from medicine. Instead, he found his calling in engineering, captivated by "the scientific process and method" – the whole life cycle of coming up with hypotheses, designing experiments to validate or invalidate them, drawing conclusions, and then formulating the next set of hypotheses. "I think that is a very neat sort of process and method. I found that really inspiring," he recounts.

His time at IIT Rurki was transformative. Initially enrolled in electrical engineering, he soon realized his true passion lay elsewhere. "That was sort of the first major sort of rebellious streak in me," he recalls, describing how he decided to stop paying as much attention to electrical engineering and instead spent his time learning programming and building software. The vibrant programming club and culture at IIT Rurki, particularly a group called SDS Labs, became his crucible. There, a small group of "10-15 coders" were constantly "tinkering and hack building a ton of applications for the internet for the rest of the campus."

Watching users interact with his creations – "building something from scratch and putting it out there and seeing how users interact with I think that was like a dopamine hit that kept like sort of fueling this." This early exposure to user-centric development, fueled by late nights and shared obsession with like-minded peers, ignited a lifelong ambition to build something of his own. He had considered it at the end of his undergrad and PhD, but "for various reasons didn’t end up doing it. So, it was just a matter of time." He wanted to "push on that vision that I care about as opposed to working on somebody else’s vision."

That vision is nothing less than a complete reimagining of our digital interactions. "Over the last two or three decades, web browsers by and large have stayed the same," Das observes. "There is an opportunity now to reimagine what that experience looks and feels like." Yutori envisions a future where we interact with "AI assistants that take actions and complete tasks on users behalf on the web," often proactively in the background. Das believes "digital agents will become a reality" sooner than physical ones, allowing us to interact with the web at a "slightly higher level of abstraction."

This isn't about AI replacing humans, but about "humans and agents working together to overall improve productivity." By delegating "all the mundane stuff to AI assistants, AI agents on our behalf," we can "focus on tasks and stuff that's more meaningful, that's more interesting to us." Crucially, it’s also about accessibility. "My parents for example no longer have to learn every new website and how to operate it," he illustrates. "If they can just tell an assistant that this is what I want to do on this particular website and it does it for them reliably, then that’s awesome."

Engineering Trust: Yutori's Blueprint for Reliability

The path to reliable agents, however, is fraught with challenges. The very nature of agents – making a "sequence of decisions" – means that even small errors can derail an entire workflow. Yutori’s solution isn't just about preventing errors, but about gracefully recovering from them. "Being able to recognize when it makes mistakes and backtrack from that to then go down a different branch is really, really important."

This commitment translates into rigorous "evals and guardrails." Every single production query that a user runs goes through a "fairly comprehensive set of evals" that lets Yutori "quickly identify where these agents are doing well versus not, which domains need more work and so on." The web's dynamic nature means agents will never be trained on every single website out there. "It is very natural to expect models to also make mistakes," Das admits. The key, he reiterates, is the agent's ability to "recognize and then backtrack and correct itself to do the right thing" when faced with an unfamiliar scenario.

Das’s disdain for "the normalization of slop and non-determinism and poor reliability" is a core tenet of Yutori’s development philosophy. They are not just building features; they are engineering trust.

The Craft of "Feeling Seen": Beyond Features

In a world where generative AI makes it "very easy to come up with first prototypes" using coding LLMs, Das argues that "the true differentiator is in taste and craft – in how intuitive and well-designed the product is." This goes beyond simply fulfilling user requests.

Yutori adopts an 80/20 approach, prioritizing features based on user feedback but also relying heavily on intuition to identify unspoken needs. "Very often there are ways to build product that users may not be asking for," he says, "but if you built it... then users feel seen and they feel like oh this is someone who is listening to us." He offers the example of two-factor authentication auto-fill on mobile phones. "It is hard to imagine like a user asking for that feature. But it saves a few seconds multiple times a day for people all across the world." These "tiny paper cuts," he explains, are where thoughtful design truly shines, making users "feel seen."

To cultivate this "taste" and "craft," Yutori rigorously "dogfoods" its own product. Weekly sessions are dedicated to testing new features internally, with only a fraction making it to external users. This constant internal refinement builds the "muscle" for discerning what "awesome or magical feels like." It's "a lot of reps to build that muscle," he notes.

Transparency: The Foundation of Trust

Das also champions the concept of "proof of work" in AI. Referencing his involvement in the highly cited Grad-CAM project – which helped visualize what deep learning models "look at" to make predictions – he emphasizes the importance for models to convey "not just the final prediction or the final answer, but also the proof of work: what are the steps that went into coming up with this final prediction or the final answer."

Yutori integrates this philosophy directly into its product. When an agent (or "scout") monitors the web and generates a report, users can "inspect the work that went in in behind the scenes" – seeing which websites were visited and what information the agent actually looked at to pull out a piece of information. This transparency, he stresses, is "very, very important for trust building for users to be able to trust that yes this is a reliable product."

For Das, trust is earned through meticulous attention to detail, both visible and invisible. "If we put attention to detail into parts of the product that users can see, then the user is more likely to trust the parts of the product that they cannot see." This philosophy underpins Yutori’s commitment to "build something meaningful, to build something right, to bring a vision of the future to life and building like delightful and reliable product experiences. It doesn't just appear out of nowhere." It takes time, hard work, and an unwavering dedication to getting it right every single time.

In a world saturated with AI demos that promise much and deliver little, Abhishek Das and Yutori are charting a different course. Their "first-try imperative" isn't just a product development rule; it's a foundational principle for an AI-powered future built on trust, reliability, and genuinely intuitive experiences. As AI agents increasingly weave into the fabric of our digital lives, Yutori aims to ensure they are not just smart, but consistently dependable – transforming mundane digital chores into seamless, invisible assistance.

Based on "The internal AI tool that's transforming how Stripe designs products | Owen Williams" from How I AI Watch the original video

Stripe's AI Secret: How an Internal Tool is Redefining Product Design

In the fast-paced world of tech, where product design can often feel like a bottleneck, companies are constantly searching for ways to accelerate innovation while maintaining a high bar for quality. Stripe, a company renowned for its meticulous design and user experience, is no exception. However, instead of relying solely on off-the-shelf tools, Stripe has cultivated a powerful internal AI-driven prototyping platform called "Protodash" and "Protodash Studio," spearheading a quiet revolution in how its products are conceptualized, built, and refined.

Owen Williams, a design manager at Stripe, found himself at the forefront of this transformation. As he observed designers grappling with generic AI tools, he recognized a critical gap: these tools, while powerful, often produced what he affectionately termed "Tailwind indigo slop"—a bland, unbranded aesthetic that failed to capture Stripe's distinctive visual identity. This "uncanny valley" effect broke immersion in design reviews, making it difficult to assess whether a prototype truly represented a Stripe product. His dream was simple yet ambitious: "I want something that's like vzero but fast," a tool deeply integrated with Stripe's unique design system.

The Pain of Prototyping in a Data-Rich World

The challenge of creating high-fidelity prototypes is particularly acute for data-heavy products like Stripe's dashboards. Imagine the sheer effort involved in prototyping a data dashboard in traditional tools like Figma: accounting for all interactions, filters, multiple states (zero data, a bunch of data, error states), internationalization (how does it look in Dutch?), or different business models (startup vs. enterprise). Owen highlights this pain point vividly: "It is nearly impossible to do that in Figma." Designers would spend countless hours manually duplicating components, faking data with "10 xxx" placeholders, and creating static screenshots for every possible scenario.

This manual, labor-intensive process not only slowed down the design cycle but also limited the scope of exploration. Designers were often forced to present isolated JPEGs or simplified flows, making it difficult for stakeholders to grasp the full user journey or the nuances of dynamic data. The absence of interactive prototypes meant feedback often focused on static visuals rather than functional experience, leading to superficial discussions.

Protodash: Crafting Prototypes with Code and AI

Owen Williams, with his engineering background, approached this problem pragmatically. He envisioned a system that could lower the barrier to entry for designers while ensuring prototypes adhered to Stripe's exacting quality standards. The solution was Protodash, an internal prototyping tool that leverages a bundle of Cursor rules, React components, and Stripe's proprietary design system, "Sale."

The first iteration of Protodash was a basic setup: a router, Sale components, and an integration with Stripe's internal MCP (Managed Component Platform) server. The core innovation, however, lay in the "bundled rules" fed to LLMs like Cursor (or Claude Code, as mentioned in the transcript). These rules taught the AI how to interpret design requests, prioritize Stripe's design system, and avoid common pitfalls like generating generic Tailwind styles. For instance, if a user pasted a Figma link, the AI was instructed to check the Sale MCP server before writing any code. This ensured that the generated prototype was built from Stripe's established building blocks, not generic elements.

The result was transformative. "It gets you there like 90% of the way most of the time," Owen explains. Designers, who are "high craft," could then refine the remaining 10%. The impact on design reviews was immediate: "It's so convincing that I'm like, is this the real product or am I looking at like something fake?" This level of realism, achieved rapidly, meant stakeholders could interact with prototypes that felt indistinguishable from the live product.

The Magic of Devboxes and Protodash Studio

Initially, Protodash ran locally, requiring designers to "npm run dev." However, Stripe's sophisticated "devbox" infrastructure elevated the experience further. Devboxes are cloud-based development environments that can be spun up in minutes, pre-configured with all necessary tools and dependencies. Now, designers could simply navigate to a URL, and a devbox instance of Protodash would be ready to go, eliminating the need for local setup or managing a large monorepo. This "no npm required" magic was a game-changer, especially for designers who might not have the same powerful machines as engineers.

Building on this, Owen developed "Protodash Studio," an extra layer that transformed Protodash into a browser-based "vibe coding platform." This studio environment allows users to interact with the LLM directly within their browser, without needing to open a separate coding editor. Users can manage their prototypes, view a "vibe prototyping feed" for inspiration, and even remix others' prototypes. The most compelling feature is the embedded LLM, where a user can simply type a request—like "I want to add a new variant of my prototype where there chat a line chat"—and the AI will build it entirely in the browser.

Owen admits much of Protodash Studio was built through an iterative process of "yelling at Claude Code for 18 months," leveraging his engineering background to guide the AI's architecture. The internal nature of the tool meant it didn't need to be "production grade" in the traditional sense, allowing for rapid experimentation and feature addition.

Transforming the Design Review: Demos, Not Memos

One of the most significant impacts of Protodash has been on the design review process. Owen passionately advocates for "demos not memos," a philosophy championed by other Stripe design leaders. He laments the "drowning in presentations" of the past five years, where reviews involved static JPEGs and Figma screenshots. With Protodash, designers can share a URL, allowing everyone in the review to interact directly with a clickable, dynamic prototype. "Can I just say being in a design review where I can click things is my favorite," Owen enthuses, expressing his goal to "never want to see a slideshow again."

Protodash Studio further enhances this by integrating a robust design review mode. Stakeholders can comment directly on the prototype at specific points, eliminating the need for fragmented Google Docs or lengthy email threads. These comments are then summarized by the AI, and crucially, they can be directly fed back to the AI for fixes. Owen describes the delight of being able to tell the AI, "The filter pattern isn't right here. Please add three more options," and then, after the AI implements the changes, responding to feedback providers with "I fixed Katie D's feedback for you. Here's the receipts thread." This drastically reduces the "busy work" associated with post-review follow-ups.

The PM Empowerment Revolution: A Surprising Twist

Perhaps the most unexpected and impactful outcome of Protodash is its adoption by Product Managers (PMs). Owen initially felt "a little nervous" seeing PMs design, wondering, "what's going to happen?" However, this apprehension quickly turned to excitement. PMs can now articulate their ideas with unprecedented clarity by building high-quality, interactive prototypes that look and feel like Stripe products.

This empowerment has several profound benefits:

The tool changes the nature of conversations, shifting them from arguments about staffing to discussions about the work itself. As Owen notes, it helps PMs "talk the same language" as designers and engineers, especially for complex technical features like webhooks with multiple states.

Building a Bespoke Culture Tool

Owen strongly believes in the value of building internal tools that are precisely matched to a company's culture and cadence. While external tools like Vzero are excellent, they can't be as deeply integrated or customized to Stripe's specific needs, design system, and review processes. Protodash isn't just a prototyping tool; it's a cultural enabler.

The platform is designed for contribution, with Owen actively encouraging designers to submit pull requests to evolve its features. This fosters a sense of ownership and allows the tool to adapt organically to the changing needs of the design and product teams. The ability to quickly integrate a Google Doc Product Definition (PD) and have the AI build a prototype from it, or to enable "crazy eights" mode for generative, out-of-the-box design exploration, highlights its flexibility. It even allows for temporary deviations from the core design system by granting the LLM access to other frameworks like Tailwind, enabling generative exploration beyond established components.

During a live demo, Owen illustrated this power by prompting Protodash to build a "Black Friday Cyber Monday dashboard for a pet store showing live sales on a chart at the top of the page, a ticker with the latest sales and top performing products, with realistic real-time data." Within moments, the AI began assembling the page, calling on Stripe's Sale MCP server to select appropriate components and structure the layout. This rapid, on-the-fly generation of complex, data-driven interfaces underscores the tool's transformative potential.

Owen's journey with Protodash at Stripe is a compelling case study in how internal tooling, powered by thoughtful AI integration, can fundamentally reshape product design and development. By addressing specific pain points, fostering collaboration, and empowering diverse teams, Stripe is not just building better products; it's building a better way to build.

Based on "What Drives Political Violence in America" from New York Times Podcasts Watch the original video

Beyond the Fringe: How Mainstream America Embraces Political Violence

In recent times, a disturbing drumbeat of political violence has echoed across the United States. From the murder of a state representative in Minnesota to an arson at a governor's residence, and multiple reported assassination attempts against former President Donald Trump, these acts signal a dangerous escalation. The question that persistently arises is: how did we arrive at this point, and how much worse could it possibly get?

To dissect this unsettling trend, the New York Times spoke with Professor Robert Pape of the University of Chicago, a leading expert on political violence who has advised every White House from 2001 to 2024. His insights paint a sobering picture of an America grappling with a new and profound threat: "violent populism."

The Age of Violent Populism: A Threat from Within

Professor Pape asserts that the most critical fact about political violence in America today is not the rise of foreign adversaries or terrorist groups, but the alarming acceptance of violence by tens of millions of Americans across the political spectrum. This widespread normalization, he argues, poses a greater risk to American democracy than any external threat.

What exactly is "violent populism"? It's a phenomenon where political violence is no longer confined to a radical fringe or a few militia groups, but finds significant social approval among a broad segment of the populace. This acceptance, Pape explains, has a two-fold dangerous effect:

  1. Empowering Volatile Individuals: For individuals already predisposed to anger or impulsivity due to their psychosocial reasons, the perceived social approval for violent acts can push them "over the edge." They see a prospect of validation for actions that would otherwise be condemned.
  2. Silencing the Alarms: When millions embrace political violence, it makes it incredibly difficult for law enforcement agencies like the FBI to receive crucial tips. People are more likely to "look the other way" or discount what they see, eroding the informal intelligence network vital for preventing attacks.

This era, Pape clarifies, is distinct from a civil war, as there are no organized armies clashing. Yet, it's far beyond "nasty politics as usual." It represents a spiraling cycle of violence fueled by broad social acceptance.

Alarming Statistics: A Nation Normalizing Assassination

Pape's research, conducted through 20 nationally representative surveys over the last five years by the University of Chicago Project on Security and Threats (CPOS), reveals shocking figures on the social acceptance of political violence.

For much of the Biden administration, surveys found that 10% of the body politic supported political violence to restore Donald Trump to the presidency. Further focus groups revealed an even more chilling detail: 55% of those who endorsed the "use of force" believed it meant assassination. This implies that at one point, roughly half of that 10%—or 20 million American adults—found assassination acceptable for political ends.

Disturbingly, this percentage has since doubled. The latest surveys indicate that between 14% and 21% of Americans now support some form of political violence. The acceptance isn't limited to one side; when asked about the perceived acceptability of violence against figures like Charlie Kirk or the attempted assassination of Nancy Pelosi, 10% for each event found it acceptable. This translates to 20 million American adults on each side of the political divide.

In essence, roughly one in five Americans now supports some form of political violence, and this number is increasing on both sides of the spectrum.

Echoes of History, With a Dangerous Twist

While the current climate feels unprecedented, Pape points to other periods of "violent populism" in American history. The 1920s saw the resurgence of the second Ku Klux Klan, growing from hundreds of thousands to 4-6 million members who actively supported political violence, not just in the South but across the nation. The 1960s also witnessed a string of political assassinations and collective violence, like the 1968 Chicago convention riots, with opinion polls from the era showing similar levels of public support for such acts.

However, a critical difference sets the current era apart: "It is pretty much the first time it's happened on both sides at the same time." The 1920s violence was predominantly from the right, and the 1960s largely from the left. Today, the escalation is bipartisan, making it uniquely perilous.

Quantifying this escalation, Pape's study of threats to members of Congress from 2001 to 2024, based on Department of Justice prosecutions, reveals a fivefold increase in threats since the first year of the Trump presidency in 2017. This surge affected both Democratic and Republican targets, indicating that Trump's rise acted as a "lightning rod on both sides." The dramatic spike has remained consistently high for an eight-year period, spanning both the Trump and Biden presidencies, suggesting that the underlying drivers are deeper than any single individual.

The Tectonic Shifts: Demographics and Wealth

At its core, this rise in political violence is not simply the impact of a president or social media, but the profound effect of social change. Pape identifies two major, interconnected social changes reshaping the United States:

  1. The Demographic Tipping Point: For the first time in its 250-year history, America is transitioning from a white-majority democracy to a white-minority democracy. In 1960, non-Hispanic whites constituted about 88% of Americans; today, that figure is 57%, projected to be 49% in about 20 years. Immigration is a primary driver of this shift, making it the "number one lightning rod issue in America today."

    • On the Right: The established group fears losing political power and declining economic prospects, leading to a desire to halt or reverse this change.
    • On the Left: The emerging groups welcome and often seek to accelerate this change, as they stand to benefit. This fundamental upheaval directly fuels support for political violence, as shown in Pape's surveys.

  2. The Concentration of Wealth: Beginning in the mid-1980s, an accelerating shift of wealth to the top 1% has occurred, regardless of which party holds power. The bottom 90% of Americans are effectively subsidizing this enormous gain for the wealthiest few. Both Republican and Democratic constituents, comprising the bottom 90%, perceive their economic prospects being "clobbered even more." This unresolved economic disparity, ignored by both parties, further fuels the bipartisan nature of rising political violence.

These shifts create a sense of profound powerlessness and a fear of "political exclusion" or "lockout" for many. As groups feel their grievances cannot be addressed through conventional politics, and their very future seems existentially threatened, the appeal of more drastic measures grows.

The New Face of the Attacker: From Fringe to Mainstream

A persistent misunderstanding, Pape argues, is the profile of those who commit political violence. We often imagine them as the "losers of society"—marginalized, unemployed individuals with little to lose. However, Pape's analysis of the January 6th attackers, compared to right-wing political violence perpetrators from previous decades, reveals a starkly different picture.

The "old profile" typically consisted of unemployed individuals, often members of militia groups—the "fringe." But among the hundreds arrested after January 6th, the profile was dramatically different:

This new profile indicates that the "halves"—educated, middle-class, and even upper-middle-class individuals—are increasingly supporting political violence. Their motivation stems not from having nothing, but from the fear of losing what they have, of seeing their stature and influence decline rapidly. "We like our monsters to be villains that are far away from us," Pape notes, explaining why this shift in attacker profile is so difficult for society to acknowledge.

From Rhetoric to Reality: The Role of Leaders and Social Media

While structural issues are the "fire," rhetoric acts as an accelerant. Pape highlights how top-down rhetoric from political leaders can push already predisposed individuals towards violence. He cites examples like Donald Trump famously suggesting "punching people out," Beto O'Rourke advocating to "throw that punch first," and Gavin Newsom speaking of "punching a bully back in the mouth." When authority figures normalize or even glorify aggressive action, others are more likely to follow suit.

Social media, while not the root cause, certainly "adds gasoline to a fire." It provides an immediate manifestation of the millions who support political violence, creating echo chambers where such acts are cheered on and lionized. Donald Trump Jr.'s social media post mocking the attack on Paul Pelosi, which generated "enormous social media support, lots of laughter," is a prime example of how top-down rhetoric can interact with bottom-up amplification, fostering a sense of social approval for would-be attackers. Yet, Pape cautions against overstating social media's role, reminding us that previous eras of violent populism occurred long before the internet.

A Choice for a Better Future

Given these deep-seated drivers, what can be done? Pape outlines a two-stage approach:

1. Longer-Term Structural Solutions (Post-Midterms):

2. Immediate Action (Before the Midterms):

The overarching message, Pape emphasizes, is that despite all structural determinants and pressures, "the future of political violence is a choice." While tens of millions support violence, a significant counterweight exists: 75% of the U.S. population "abhors political violence." This silent majority, Pape insists, must exert its agency. They need to actively voice their opposition to violence, sending letters and making calls to their representatives, demanding joint condemnation from leaders.

"Leaders listen to their constituents," Pape concludes. If their inboxes are filled only with demands to "hit the other side harder," that's what they'll do. But if the 75% who oppose violence step up and make their voices heard, a different future is possible. It is a choice that "we the people" have to make, for the sake of the country and its future.