YouTube Digest

English

Based on "Anthropic Just Bought a Dev Tools Startup for $300M. Here's What Its Founder Told Me." from Every Watch the original video

The Internet's Next Evolution: How AI Will Learn to Code Its Own Future

The internet, a marvel of interconnected systems, was fundamentally designed for human interaction. Its architecture, from websites to web services, caters to clicks, scrolls, and visual interfaces. But as artificial intelligence, particularly large language models (LLMs), becomes increasingly sophisticated, a new challenge emerges: how do we empower AI to truly use the internet, not just observe it?

This is the frontier where companies like Stainless, an API and SDK powerhouse behind some of the biggest names in tech, are operating. Founded and led by Alex Ratray, Stainless has long specialized in making computers talk to each other seamlessly. Now, they're at the forefront of enabling AI to do the same, grappling with the complexities of what's known as Model Context Protocol (MCP).

"APIs are the dendrites of the internet," explains Ratray, drawing a vivid analogy to the neural connections in our brains. Just as thoughts don't happen without neurons communicating, the internet ceases to function without servers and applications exchanging information. For decades, this communication has been between humans and computers via user interfaces (UIs), or between computers via APIs. Now, a "new computer has entered the chat": LLMs.

The Promise and Pitfalls of Model Context Protocol (MCP)

Model Context Protocol (MCP) represents the internet's attempt to give LLMs their own native interface. Imagine every website or web service transforming into a set of "tools" an AI can use directly to get work done. For instance, a Gmail MCP might offer tools like "send mail" or "read inbox," allowing an LLM to compose and manage emails without a human intermediary. The promise is profound: AI agents capable of automating complex, multi-step tasks across various applications.

Ratray illustrates this with a common, yet surprisingly convoluted, business scenario. "Let's say Dan walks into my store and buys a pair of stripey socks," he begins. "The next day, something's wrong, and I tell my team, 'Hey, can we refund Dan for those stripey socks and send him a discount code with a thank you note?'"

For a human operator, this seemingly simple task involves a laborious dance across multiple applications: logging into an internal admin, searching a Stripe dashboard for transactions, finding Dan, identifying the specific product, initiating a refund, creating a discount code, and then switching to another SaaS app to send an email. Each step involves multiple clicks, scrolls, and loading screens – a process that could easily rack up 15 different interactions across five different apps.

The vision of Agentic AI, powered by MCP, is to type that same request into an LLM and have it execute the entire sequence autonomously. "That's the big vision of MCP," Ratray emphasizes. "To take that prompt and just have the AI go off and do that."

However, the reality of current MCP implementations falls short. Ratray admits, "it's not working so great." The core problem lies in the sheer volume of information and the way LLMs process it. To truly empower an AI agent to do anything a human operator could, you'd need to expose every single function of every single API and dashboard – hundreds of endpoints for a service like Stripe alone, let alone multiple integrated SaaS tools.

This leads to several critical challenges:

Current solutions often involve drastically limiting the number of tools exposed or using "dynamic mode," where the LLM has only three tools: list endpoints, get endpoint details, and execute endpoint. While this helps with context, it increases the number of turns the model needs to perform a single action, making it slower and more expensive.

Stainless in Action: AI for Business Insights

Despite these challenges, Ratray and his team at Stainless are actively leveraging MCP for internal operations. "I use MCP to actually, funnily enough, not so much on the coding side, but I use it on the business side," he shares.

Stainless employs MCP servers for Notion, HubSpot, Gong, and even a read-only copy of their own PostgreSQL database. Ratray can prompt an LLM, "Hey, what are the interesting customers that signed up for Stainless last week?" The AI then cross-references data from the database, HubSpot, Notion notes, and even Gong call transcripts to provide a comprehensive answer.

Beyond generating business reports, Stainless is experimenting with AI for knowledge management and even bug fixing. Ratray has a "master company Git knowledge repository" where he instructs Claude Code to store interesting customer quotes and their citations, SQL queries, and other valuable insights. This allows the AI to "cache" information, avoiding repetitive searches and quickly retrieving relevant data for future queries or board preparations.

"That's [expletive] so cool," the interviewer exclaims, highlighting the transformative potential of such systems. While still experimental for tasks like bug fixing, the promise of AI-powered efficiency is clear.

The "Cyborg" Future: AI That Writes Code

Looking ahead, Ratray has a bold vision for how AI will truly integrate with the internet, a future he provocatively calls "cyborgs." This doesn't mean literal human-robot hybrids, but rather AI agents that are "part GPT neural net LLM, part AI, and part code." The "code" refers to traditional CPU-based software, not just GPU-driven neural networks.

Ratray envisions a future where LLMs don't just pick from a vast menu of pre-defined tools, but instead write the code themselves. He proposes a model where, instead of a "bajillion tools," the LLM has just two:

  1. Execute Code: The LLM is given a text box to write TypeScript (or another language) using an API's SDK. For example, it could write Stripe.transactions.list() or Stripe.refunds.create().
  2. Search Docs: A tool to query documentation and ask questions about the API.

This approach offers several compelling advantages:

"LLMs are really good at writing code," Ratray asserts. The main hurdle today is their difficulty working with specific library versions and avoiding hallucinations when dealing with raw HTTP APIs without the guardrails of a typed SDK. By having the API provider execute the code, these issues can be managed within a controlled environment.

The Security Imperative: Back to the API Layer

The "cyborg" vision also offers a clearer path for security. Ratray argues that security should ultimately reside at the API layer itself, not by attempting to limit what's exposed through MCP. "Right now you see people trying to implement security by sort of limiting what's exposed through MCP, and that kind of makes sense, but at the end of the day, you could do anything that's in the API under the hood, right?" he points out.

The solution, he believes, lies in robust authentication protocols like OAuth with granular, properly scoped permissions. This ensures that even if an AI agent generates code, its actions are constrained by the permissions granted at the API level, preventing unintended or unauthorized operations.

The journey to a truly intelligent, internet-fluent AI is complex, fraught with technical and security challenges. But pioneers like Alex Ratray and companies like Stainless are charting a course. By moving beyond simplistic tool definitions to an architecture where AI can dynamically write and execute code, we may soon unlock a future where AI agents seamlessly navigate and operate the digital world, transforming productivity and automation in ways we are only just beginning to imagine.

Based on "Was anything good released at Google I/O 2026?" from How I AI Watch the original video

Google I/O 2026: The AI Revolution, One Flash (and a Few Fumbles) at a Time

Google I/O, the tech giant's annual flagship event, is always a whirlwind of announcements, new features, and ambitious product visions. For Claraveville, a product leader and self-proclaimed AI obsessive behind "How I AI," the 2026 iteration was no different. This year, Google went "full bore into agents," launching a staggering array of AI-powered tools and models, promising to reshape everything from coding to creative production. Yet, as Claraveville meticulously explored these innovations, a recurring theme emerged: the tantalizing promise of AI often collided with the frustrating reality of limited access, confusing branding, and features that, well, just didn't quite work – at least not yet.

The Foundation: Gemini 3.5 Flash — Speed Meets Intelligence

At the core of Google's I/O announcements was the unveiling of the Gemini 3.5 family of models, prominently featuring Gemini 3.5 Flash. Claraveville immediately highlighted its significance: "their fastest, smartest coding model." What truly sets Flash apart, according to Google's benchmarks, is its dual prowess. It not only "rivals the intelligence of some of our favorite coding models" like Opus and GPT-4 but is also "four times as fast." This combination promises a "super smart model... at the speed of something much more like their 3.1 flash model."

The emphasis on agentic capabilities was unmistakable. Google is betting big on AI agents, and Gemini 3.5 Flash is designed to be their engine. Claraveville noted that while some of these agentic features might feel like Google playing "catch-up" to competitors like Anthropic and OpenAI, they come with two distinct advantages: the sheer speed of the Flash models and a pronounced "creative consumer bent."

Beyond coding, Gemini's multimodal strength continues to be a standout. Claraveville has consistently championed Gemini for its ability to handle diverse data types: "I have always told people if you are working with files, videos, any sort of transformative work where you have to go from one modality... Gemini models are really, really good at handling files. I love it for handling videos." The benchmarks presented at I/O seemed to confirm this, with Gemini 3.5 Flash reportedly exceeding its peers in multimodal and agentic capabilities, all while being incredibly fast.

Developers' New Toolkit: Anti-gravity and Google AI Studio

For developers, Google introduced updates to its agentic IDE, Anti-gravity, now in version 2.0, and the Anti-gravity CLI. Claraveville observed that many of these features felt familiar, clearly positioning Google in a competitive space with tools like CodeX. Anti-gravity 2.0 brings "projects" (folder-constrained workspaces) and "scheduled tasks" – a UI reminiscent of existing agentic coding apps.

However, Anti-gravity also introduced some intriguing new capabilities. The core agent features now include sub-agents, allowing the main agent to delegate specific tasks, with the browser sub-agent being a particularly popular use case. "Hooks" allow developers to tap into various lifecycle events of their agent, offering on-demand customization. Claraveville was particularly keen on the new "slash commands," especially /goal, which enables an agent to pursue a long-running task until a defined objective is met.

But the real curiosity lay with /grill me. This command promises a more "aggressive" clarification process, where the agent will "ask clarification questions back and really get to the heart of what your requirements are." Claraveville mused, "Now the question is, is this actually as hardcore as the slash command communicates or is this just a cute way to differentiate against the question and answer tool? We'll see."

Putting Anti-gravity to the test, Claraveville attempted to convert a blog generator UI into an API for an AI agent. While the outcome was successful – generating a programmatic API endpoint with authentication and visual workflow triggers – the speed didn't quite match the "Flash" hype in this specific instance. "I'm not noticing it is particularly faster... it doesn't feel that much faster to me," she noted, concluding that while it was "very nice, fast, kind of what I would expect from a coding agent," it still felt like "playing catch-up."

Beyond the IDE, Google AI Studio received updates, leveraging Gemini 3.5 Flash for low-code/no-code app development. The most significant announcement here was the ability to build apps "connected to your Google Workspace apps" – reading Sheets, drafting Gmails, organizing Drive, and managing calendars. This is Google's strategic move to "own that experience themselves," carving off internal enterprise and personal assistant use cases. The studio also now supports creating Android apps.

However, Claraveville's attempt to build a simple calendar management app for her kids' sporting events quickly hit a wall. Despite being a paying customer, the promised Workspace integration was nowhere to be found. "It didn't do anything magical," she reported, highlighting a recurring frustration: "they've announced a bunch of stuff and then at the very, very bottom they're like it's available to some subset of people or it's available but later this summer."

The Consumer & Creative Frontier: Gemini AI, Omni, and Flow

Before diving deeper into consumer-facing AI, Claraveville paused to voice a significant critique: Google's product naming strategy. "I cannot keep the product and brand name straight," she lamented, listing the myriad of names already encountered: Gemini 3.5 Flash, Anti-gravity, Google AI Studio, and now "Gemini, Google AI, the consumerish AI product." Her plea to the Google team was clear: "a comprehensive brand analysis and singular brand that I can work with."

The consumer-facing Gemini AI product received a redesign, dubbed "something ridiculous" by Claraveville but intended to "make you feel good" with its "glow" and inspiring prompts. More substantially, Gemini now integrates Nano Banana, an acclaimed image generation model, with a focus on templates to inspire users. Claraveville put it to the test, asking it to "upscale and beautify the image of this podcast host, change the background to a professional podcasting studio." The result was a mix of the promising and the peculiar: "This is not my face. This is horrifying in every way possible." Yet, she noted improvements in text generation and photorealism (even if not face-realism), all generated "pretty fast."

The real excitement in the creative space revolved around Omni, Google's new video generation model, positioned as "Nano Banana for video." Omni promises "longer, more consistent, more photorealistic videos using AI," capable of leveraging "reference materials." Claraveville emphasized Gemini's strong multimodal capabilities, noting that Omni allows users to "create video from anything" and "ground Gemini in real-world knowledge." Key features include conversational video editing (e.g., changing a structure into bubbles), descriptive video editing, and the ability to refine videos while maintaining character consistency – a crucial advancement for production-level video.

Claraveville demonstrated Omni's power by animating a drawing her child made: "Animate this superhero breaking a kid out of class to go have fun." The result was a delightful success, a 10-second video that her "kid is really going to like," significantly longer than previous models like Sora.

For more prescriptive video editing, Google introduced Flow, a tool embedding Omni with a focus on "cinematic quality production." Flow allows users to blend multimodal references, edit conversationally, and even define characters (e.g., "Captain Escapo School") for consistent use across videos. Most intriguingly, Flow promised the ability to create an AI avatar of oneself. Claraveville eagerly attempted this, scanning a QR code, reading numbers aloud, and turning her head for the camera: "Please don't steal my identity!" she quipped. The outcome? Another I/O disappointment. "No. No sir, I couldn't... it didn't even create the avatar." This led to a stark conclusion: "The promise is really good for some of these things, but the reality is if you're not able to use them or they're broken on the day... people are going to lose patience for some of this."

Design & Marketing Tools: Pamelli and Stitch

The product naming confusion continued with Pamelli, Google's brand and marketing content generator, and Stitch, their design tool. Pamelli introduced an agent that can create a "core brand identity from a website," including brand books, campaign assets, and photoshoots – a feature reminiscent of Claude Design. Claraveville highlighted Google's underlying design.md standard, which allows product designs to be encoded in markdown for AI agents. Her own website, chatprd.ai, was used to demonstrate Pamelli's ability to generate a brand book, which later yielded "definitely my brand colors. This is my tagline. And now I can create... a new website." Her final assessment of the generated website: "it's fine."

Stitch, described as an "in-browser Figma," showcased features like streaming into a design canvas, starting with existing designs, and inline AI edits. Claraveville noted, "This is where I think I feel the speed of the Flash models is actually in design and this sort of experience versus the coding experience." She demonstrated Stitch by asking it to design an app in the style of the Google I/O site, showcasing the streaming design process. Her observation on AI design tools was candid: "Every AI tool ships their own version of slop. Google's slop looks like Google."

The Verdict: A Glimpse of the Future, Still Under Construction

Google I/O 2026 presented a comprehensive vision of an AI-powered future, driven by the speed of Gemini 3.5 Flash and the ambition of agentic AI. From developer tools like Anti-gravity and Google AI Studio to creative powerhouses like Omni and Flow, and specialized design/marketing tools like Pamelli and Stitch, Google's portfolio is undeniably expansive.

Claraveville's journey through these announcements was a microcosm of the current state of AI: breathtaking potential mixed with frustrating limitations. While the video generation capabilities of Omni and the speed felt by designers using Stitch were genuinely exciting, the struggles with accessing promised features, the broken avatar creation, and the pervasive product naming confusion painted a picture of a revolution still very much under construction.

Ultimately, Claraveville's key takeaways were clear: "The video piece is the part that I'm most excited about. probably second is just the speed of coding that comes from these Flash models." Yet, she concluded with a tempered optimism: "there are some sharp edges, some things that definitely need to be improved and some things that actually need to be released." Google I/O 2026 offered a tantalizing glimpse into a more intelligent, agentic, and creative future – one that, with a bit more polish and accessibility, might just live up to its ambitious promise.

Based on "Trump’s Taxpayer-Funded Revenge Plan" from New York Times Podcasts Watch the original video

Taxpayer-Funded Retribution: Unpacking Trump's $1.7 Billion Slush Fund

In a move that has sent shockwaves through Washington and ignited bipartisan fury, President Trump's administration has unveiled an unprecedented $1.776 billion fund of taxpayer money. Ostensibly designed to compensate "victims of weaponization and warfare" by the federal government, critics across the political spectrum are decrying it as a thinly veiled political slush fund, potentially funneling taxpayer dollars to the President's loyalists, including those convicted for their roles in the January 6th Capitol insurrection.

The announcement came after President Trump abruptly dropped a contentious $10 million lawsuit he and his family had filed against the Internal Revenue Service (IRS). This sudden withdrawal, coupled with the creation of such a massive and vaguely defined compensation fund, has raised profound questions about conflicts of interest, the politicization of federal agencies, and the very fabric of American governance.

The Roots of a Peculiar Lawsuit

To understand the genesis of this controversial fund, one must rewind to President Trump's first run for office in 2016. Breaking with decades of presidential tradition, Trump famously refused to release his tax returns, a decision that became a recurring point of contention throughout his presidency.

The Unreleased Tax Returns and the IRS Leak

The mystery surrounding Trump's finances deepened in 2020 when The New York Times published a series of bombshell stories based on leaked tax returns, revealing that for many years, Trump had paid little to no federal income taxes. This was followed by ProPublica, another news organization, releasing detailed tax strategies of some of the world's wealthiest individuals, including Jeff Bezos and Elon Musk.

It soon became clear that the source of this highly sensitive financial information was an internal IRS leaker. In the fall of 2023, Charles Littlejohn, a former IRS contractor, was caught. He pleaded guilty to providing this information to the Times and ProPublica and was subsequently sentenced to prison in early 2024. Leaking someone's tax information is a federal crime, and civil lawsuits against the IRS are permitted for individuals whose tax data has been improperly released.

It was under this premise that President Trump, his two sons, and the Trump Organization filed a lawsuit in January, demanding at least $10 million from the IRS. Their argument: the IRS should have prevented Littlejohn from accessing and leaking their information.

A President Suing His Own Government

What made this lawsuit uniquely problematic was not just the claim, but the plaintiff himself. As President, Donald Trump controlled the very agencies he was suing. He appointed the leadership of the IRS and the Treasury Department, both named in the suit. Furthermore, the Justice Department (DOJ), tasked with defending the IRS, had been significantly reshaped under his influence. Adding another layer of complexity, the acting Attorney General, Todd Blanche, was also President Trump's personal lawyer before returning to the White House.

This created an unprecedented web of conflicts. The President, as a private citizen, hired private lawyers to sue the government. Simultaneously, as the head of the executive branch, he effectively controlled the government lawyers who were supposed to defend against that very suit. As one commentator aptly put it, "You almost need a whiteboard to like map out all of these conflicts." Donald Trump, in essence, sat atop both parties in the lawsuit, raising profound questions about the legitimacy of any potential outcome.

The Judge's Dilemma and the Sudden Deal

The inherent conflict did not go unnoticed by the judiciary. The presiding judge began to express serious concerns, questioning whether a genuine legal conflict even existed. Under legal principles, both sides in a lawsuit must be genuinely adversarial; one cannot sue oneself or a friend to manufacture a favorable ruling. The judge indicated she was considering dismissing the case outright, viewing it potentially as a "charade" filed under false pretenses. She ordered both the Justice Department and Trump's private lawyers to submit briefs explaining why the lawsuit should be allowed to proceed.

Facing a potentially embarrassing and unfavorable ruling, the administration reportedly worked "feverishly" to avoid this scenario. The outcome was a dramatic pivot: President Trump would drop his lawsuit entirely. In exchange, the Justice Department would create an "unprecedented gigantic pot of taxpayer money" to compensate "victims of weaponization."

The $1.776 Billion "Victim" Fund Takes Shape

The fund, totaling $1,776,000,000 – a figure many observers note is no coincidence, referencing the year of America's founding – will be drawn from the Judgment Fund, a standing account the Justice Department uses to settle legal claims against the United States. This sum, however, is far larger than typical withdrawals.

The money will be deposited into a separate account controlled by five individuals appointed by the Attorney General. These appointees will then consider claims from people who assert they have been victimized by the federal government.

Who Qualifies as a "Victim"?

While the Justice Department's official language refers to "victims of weaponization and lawfare," the practical interpretation of this definition has sparked intense alarm. Among the primary groups expressing excitement about the fund are individuals convicted for their roles in the January 6th, 2021, Capitol insurrection – many of whom President Trump had previously pardoned. Others include Trump supporters and political allies who claim they were improperly investigated by the Justice Department or FBI under previous administrations simply for their political leanings.

The possibility that "somebody who stormed the Capitol on January 6th could get paid by the government would be shocking, to put it mildly," noted one observer. Indeed, the scenario of convicted criminals receiving taxpayer money from a government they attacked has fueled much of the public outrage. It remains unclear whether payments would be limited to covering legal fees or if individuals could "profit from the experience overall."

A Veil of Secrecy

Crucially, many details about the fund remain shrouded in secrecy. The exact criteria for evaluating claims, the precise definition of "weaponization," and how payment amounts will be determined are largely unknown. The identities of the five individuals who will control this enormous sum and make these critical decisions have also not been disclosed. This lack of transparency has led to concerns that the entire operation "might be very hidden from public view," allowing for discretionary payouts with little public accountability.

Outrage on Capitol Hill and Beyond

The creation of the fund immediately triggered a firestorm on Capitol Hill. Acting Attorney General Todd Blanche found himself defending the arrangement during a scheduled testimony before Congress. He argued that while "unusual," it was "not unprecedented" and claimed the fund was "not limited to Republicans," "not limited to January 6th," and open to "anybody in this country... if they believe they are a victim of weaponization."

Democrats Cry Foul, Republicans Get "Squeamish"

However, Blanche's defense did little to quell the outrage. Democratic lawmakers lambasted the fund as an "outrageous unprecedented slush fund" designed to pay the president's political allies. Senator Chris Van Hollen accused Blanche of "pure theft of public funds" and called the idea of rewarding convicted criminals "obscene." He further asserted that it was part of President Trump's "broader revenge tour" and bluntly told Blanche, "You are still acting as the president's personal lawyer, not as acting Attorney General."

Even Republicans, traditionally staunch allies of the President, showed signs of discomfort. Senate Majority Leader John Thune, while measured, stated that "there are and will be continue to be a lot of questions around that that the administration is going to have to answer." This "mild discontent" from top Republican lawmakers signaled that even within Trump's own party, there was a distinct lack of enthusiasm or pride in the arrangement.

Internal Dissent

The controversy also reportedly spilled into the administration itself. The top lawyer at the Treasury Department, an agency involved in the fund's creation, suddenly resigned on Monday, just hours after the announcement. Initial reports suggest this resignation was, at least in part, due to the creation of the fund, indicating significant internal dissent over the move.

Trump's Personal Windfall: Beyond the Fund

While the President, his sons, and the Trump Organization are explicitly barred from receiving payouts directly from the $1.776 billion fund, the agreement that led to its creation includes another extraordinary concession that could directly benefit Trump financially. The IRS will drop any audits it has of Trump, his family members, and any related business entities.

This is an "extremely unusual and extraordinary concession," as audits can result in significant tax liabilities. Previous reports have indicated that the President has faced audits in the past where, if the IRS prevailed, he could owe more than a hundred million dollars in additional taxes. While the specifics of current audits remain unknown, this provision effectively "robbed the IRS of the opportunity to demand that the Trumps pay the taxes that the IRS says that they may owe under the law." This could directly deliver "tens of millions of dollars" or more in financial benefit to the President, representing a substantial personal windfall.

The Unanswered Question: Can It Be Stopped?

The question now looms: can anything be done to stop this controversial fund? Attempts by House Democrats to intervene in Trump's original lawsuit against the IRS were ultimately unsuccessful. Now, the focus shifts to Congress's "power of the purse" – its constitutional authority to control federal spending. Theoretically, Congress could pass laws to address how money can be spent on legal settlements or specifically target this fund. However, this power has been "eroded" under the current presidency, and it remains "unclear if that will actually happen."

A New Chapter in Political Retribution

For many ethics experts and watchdogs, this fund marks a dangerous escalation in the politicization of government. It represents a significant advancement in President Trump's stated goal of "retribution against people and agencies that he saw as having been opposed to him." What began as a campaign to punish perceived enemies has now, according to critics, evolved into direct "compensation for President Trump's allies" with taxpayer dollars.

Ethicists are calling this "the most corrupt act in American history," highlighting the almost complete lack of limits or controls on how this money will be used. Five individuals, appointed by the President's former personal lawyer, will have "seemingly total freedom to give money to whoever they deem is worthy of it and to give them as much money as they want."

This "radical and striking situation" means that those deemed "loyal to the president" could be "cashing out to the tunes of thousands or tens of thousands, hundreds of thousands, or even millions of dollars." It signifies a profound shift, not just in punishing enemies, but in directly rewarding allies using the public's money, fundamentally challenging the principles of impartial governance and public trust.