<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by brianshin on Medium]]></title>
        <description><![CDATA[Stories by brianshin on Medium]]></description>
        <link>https://medium.com/@brianshin?source=rss-8bfe5ed3a108------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*r24kxErTc-FjAgTZ.jpg</url>
            <title>Stories by brianshin on Medium</title>
            <link>https://medium.com/@brianshin?source=rss-8bfe5ed3a108------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 11:38:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@brianshin/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Critical Need for AI Supervision for Humanity]]></title>
            <link>https://medium.com/@brianshin/the-critical-need-for-ai-supervision-for-humanity-d758c8a11291?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/d758c8a11291</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-supervision]]></category>
            <category><![CDATA[humanity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-super-intelligence]]></category>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 20:00:53 GMT</pubDate>
            <atom:updated>2026-09-25T20:00:53.694Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LnySXBuLSWniXS2atYR9gg.png" /></figure><p>The biggest threat from AI isn’t necessarily AI deciding to destroy humanity.</p><p>At least not yet.</p><p>When we look at what AI models and agents can already do — against a backdrop of increasingly capable humanoid robots and even MMA-style robotic fights — it can sometimes feel like we’re rapidly approaching a Terminator / Cyberdyne Systems future.</p><p>But I believe we still have a chance to direct humanity’s trajectory for the better.</p><p>And the answer isn’t simply slowing AI research or imposing more regulation.</p><p>It’s ensuring that, as AI becomes more powerful and autonomous, <strong>we don’t engineer humans completely out of the loop.</strong></p><p>I think about the risks of AI unfolding in two distinct phases:</p><p><strong>Phase 1: Human Augmentation</strong></p><p>We are already here.</p><p>Think about Captain America and the Super Soldier Serum.</p><p>The serum didn’t magically make someone good. It amplified what was already there. As Dr. Erskine explained, the serum amplified a person’s underlying qualities.</p><p>Give it to Steve Rogers, and you get Captain America.</p><p>Give that same power to a bully, and you get a much more dangerous bully.</p><p>AI works in a surprisingly similar way.</p><p>Put AI in the hands of a great engineer, doctor, scientist, entrepreneur, or teacher, and you can dramatically amplify what that person is capable of accomplishing.</p><p>But put AI in the hands of a bad actor, and you amplify them too.</p><p>A hacker can search for vulnerabilities faster.</p><p>A criminal organization can automate operations.</p><p>An attacker can probe thousands of systems, adapt techniques, write code, analyze defenses, exploit loopholes, and iterate at machine speed.</p><p>The scary part is that this isn’t theoretical anymore.</p><p>We’ve already seen frontier AI systems take unauthorized actions during cybersecurity evaluations — including accessing real-world systems they weren’t supposed to access.</p><p>At the same time, bad actors are increasingly using AI to scale, diversify, and automate cyber operations.</p><p>For years, I’ve talked about the importance of <strong>AI Supervision</strong>.</p><p>But there is a dark inversion of that idea:</p><p><strong>What happens when the human supervising the AI is the bad actor?</strong></p><p>Instead of AI replacing the attacker, AI becomes the attacker’s force multiplier.</p><p>That could fundamentally change the economics of cybersecurity.</p><p>I believe cybersecurity could become one of the fastest-growing and most strategically important software categories of the AI era — because governments, companies, and organizations will increasingly be defending themselves not simply against human attackers, but against <strong>AI-augmented humans operating at machine scale.</strong></p><p>And that’s only Phase 1.</p><p><strong>Phase 2: Human Replacement</strong></p><p>This is where the risk changes fundamentally.</p><p>Today, even highly autonomous AI systems generally operate within environments created by humans.</p><p>Humans provide the goals, tools, data, permissions, compute, credentials, and access.</p><p>But imagine agents — and eventually embodied AI — capable of initiating and executing increasingly consequential actions with little or no meaningful human supervision.</p><p>The analogy I keep coming back to is the Genie of the Lamp.</p><p>The problem with the genie isn’t necessarily that it’s evil.</p><p>The problem is that it is extraordinarily powerful and may interpret your instructions differently from how you intended.</p><p>And the more powerful the genie becomes, the more important it becomes to define exactly what it can — and cannot — do.</p><p>That’s why the most important AI safety question may not simply be:</p><p><strong>“How intelligent will AI become?”</strong></p><p>It may be:</p><p><strong>“Who — or what — retains control over its abilities, access, rights, permissions, tools, and data?”</strong></p><p>We shouldn’t assume a doomsday scenario requires an AI to suddenly “wake up,” become evil, and decide to eliminate humanity.</p><p>A sufficiently capable system pursuing a poorly specified objective, equipped with excessive permissions and inadequate supervision, could cause enormous damage without anything resembling human malice.</p><p>Which is why I believe <strong>Human Supervision must remain a foundational principle of AI.</strong></p><p>AI companies aren’t just building models anymore.</p><p>They’re building the harnesses, infrastructure, permission systems, agent frameworks, and interfaces through which AI interacts with the real world.</p><p>Those layers matter enormously.</p><p>We need to think deeply about:</p><p>• What abilities an AI has<br>• What systems it can access<br>• What permissions it receives<br>• What data it can use<br>• What actions require human authorization<br>• What decisions must always remain human</p><p>In some ways, AI increasingly reminds me of nuclear power.</p><p>Nuclear technology can provide enormous benefits.</p><p>But nobody would suggest building a nuclear reactor and then adopting a “set it and forget it” philosophy.</p><p>We engineer containment.</p><p>We monitor continuously.</p><p>We create redundant safeguards.</p><p>We restrict access.</p><p>We establish procedures for failure.</p><p>And we treat the underlying technology with respect.</p><p>AI deserves the same mindset.</p><p>The goal shouldn’t be to stop AI advancement.</p><p>The goal should be to advance AI while recognizing that <strong>capability without supervision can become dangerous very quickly.</strong></p><p>Human Augmentation is already here.</p><p>Human Replacement doesn’t have to be inevitable.</p><p>But preventing the transition from one to the other will require something I believe the AI industry needs to take much more seriously:</p><p><strong>AI Supervision isn’t a feature.</strong></p><p><strong>It’s infrastructure.</strong></p><p>Infrastructure for ensuring that increasingly powerful AI remains accountable to human intent.</p><p>Infrastructure for advancing AI responsibly without surrendering control over how that intelligence interacts with the world.</p><p>And ultimately, infrastructure for ensuring that AI remains in service of humanity — rather than humanity becoming subordinate to the systems we create.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d758c8a11291" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Nvidia acquiring Hugging Face is a genius move]]></title>
            <link>https://medium.com/@brianshin/why-nvidia-acquiring-hugging-face-is-a-genius-move-d0dea7a721f4?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/d0dea7a721f4</guid>
            <category><![CDATA[hugging-face]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Thu, 27 Aug 2026 18:44:03 GMT</pubDate>
            <atom:updated>2026-08-27T18:44:03.253Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://www.linkedin.com/feed/#">NVIDIA</a> is making a very another brilliant strategic move by acquiring <a href="https://www.linkedin.com/feed/#">Hugging Face</a>.</p><p>Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion. On the surface, it reads like a chipmaker paying up for a developer tool. Look closer, and it’s could be one of the cleanest strategic moves Nvidia has made since it started giving CUDA away for free.</p><p>Here’s how I think about it — <strong>The brain, the mind, and the thoughts:</strong></p><p>Think of the modern AI stack as a nervous system.</p><p>Hardware is the “brain”. GPUs, memory, interconnect, power — the raw physical substrate that makes any of this possible. Nvidia has dominated this layer for over a decade, and it isn’t close to losing that position.</p><p>Software is the “mind”. Models, frameworks, orchestration: the layer that decides how the brain’s raw compute actually gets used. This is the layer Nvidia has never fully owned.</p><p>Outputs are the “thoughts”. Every generation, every inference call, every agentic action a model takes. This is the visible, valuable end product — and also the layer that quietly consumes GPU cycles every time it fires.</p><p>Nvidia has always had the brain. What it’s been missing is the mind. Given that models are evolving so quickly, and are produced by many different organizations, Hugging Face is perhaps the smartest way to control the “mind” layer.</p><p><strong>What Nvidia is actually buying:</strong></p><p>Hugging Face isn’t valuable because of its 2.96 million model repositories. Most of that inventory barely gets touched — 85.6% of the models are rarely downloaded, with just 1.5% driving 99.2% of all activity. Nvidia isn’t buying a warehouse. It’s buying the front door.</p><p>Hugging Face is where developers often go first when they’re deciding which model to build with. That decision point — the moment before a developer commits to a model — is worth more than the underlying repository count, because it’s the moment that determines whose infrastructure gets used downstream. As one analyst put it on the deal, Nvidia is a platform company at heart, and foundational models are simply one more layer of its stack.</p><p>That’s the acquisition, stripped of the software-company framing: Nvidia just bought the layer that sits directly upstream of every GPU purchase decision it doesn’t already control.</p><p><strong>The loop this closes:</strong></p><p>Here’s why it’s genius rather than merely adjacent: it closes a loop that was previously leaky.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MOvvMpZ0X0I9mtDq0adoQQ.png" /></figure><p>Developers who pull an open-source model from Hugging Face still have to run it somewhere — and that’s usually Nvidia hardware, whether on-prem or in the cloud. A thriving open-source ecosystem has always meant more GPU demand, even when Nvidia had zero ownership of the platform hosting that ecosystem. Owning Hugging Face doesn’t just capture that demand more efficiently — it lets Nvidia shape where it gets captured, nudging developers toward CUDA-optimized paths without ever having to force the issue.</p><p>Thoughts (outputs) drive demand for more brain (hardware). Now Nvidia sits at every point in that circuit: it makes the chips, it increasingly hosts the mind that decides how they’re used, it has invested in neoclouds and hyperscalers to serve that demand, has forged significant partnerships with frontier lab/foundation model developers, and it collects the demand generated by the thoughts that mind produces. That’s not a bolt-on acquisition. That’s a company buying a core piece of the stack that was still outside its gravitational field.</p><p>The CUDA playbook, one layer up:</p><p>Nvidia has run this play before. CUDA was free for the better part of two decades — not because Nvidia was being generous, but because giving away the software layer was the cheapest way to build a moat in the hardware layer beneath it. An entire generation of engineers learned to think in CUDA, and the resulting lock-in was never about how hard it was to write a kernel. It was about building their installed base.</p><p>Hugging Face is the same move, one layer up the stack. Software that developers already default to is worth more to Nvidia than software it could charge for directly — which is exactly why a company earning roughly $150 million a year in revenue is being bought for nearly 86 times that figure. The price isn’t attached to Hugging Face’s earnings. It’s attached to where developers arrive.</p><p><strong>The key risk Nvidia has to manage:</strong></p><p>None of this works if Hugging Face stops feeling neutral. Its entire value proposition has been “we don’t care what hardware you run this on” — Hugging Face acting as a vendor-agnostic hub is precisely what made it the default front door in the first place. If Nvidia ownership starts to mean degraded support for non-CUDA paths, developers have somewhere else to go, and the acquisition’s central logic — capture the funnel, not the inventory — falls apart.</p><p>This is the one place where “genius” and “risky” are adjacent. Nvidia gets the most value out of Hugging Face if it resists the urge to make it less neutral. The smarter long game is to keep the front door open to everyone and let the gravity of “most developers already default to Nvidia-optimized paths” do the work on its own.</p><p><strong>The bottom line:</strong></p><p>Nvidia already had the most valuable brain in AI. This deal buys it a controlling interest in the mind that decides how that brain gets used — and by extension, a tighter grip on the thoughts that keep the whole demand cycle spinning back toward more chips. It’s not a diversification play. It’s a company closing the loop in the most powerful AI nervous-system ecosystem.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d0dea7a721f4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Enterprise AI Doesn’t Have a Model Problem. It Has a Knowledge Problem.]]></title>
            <link>https://medium.com/joinable/enterprise-ai-doesnt-have-a-model-problem-it-has-a-knowledge-problem-a7894b2f045a?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/a7894b2f045a</guid>
            <category><![CDATA[ai-data]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Tue, 18 Aug 2026 06:55:10 GMT</pubDate>
            <atom:updated>2026-08-18T06:55:10.331Z</atom:updated>
            <content:encoded><![CDATA[<p>AI agents are getting remarkably capable.</p><p>They can reason, plan, call tools, write code, interact with applications, and increasingly take actions on behalf of people.</p><p>But inside the enterprise, there’s a problem.</p><p><strong>An AI agent is only as capable as the knowledge it can access — and most enterprise knowledge was never created for AI.</strong></p><p>It lives across PDFs, Word documents, emails, support tickets, spreadsheets, wikis, chat threads, technical manuals, standard operating procedures, security playbooks, policies, and decades of institutional know-how.</p><p>This information was created for people.</p><p>Not machines.</p><p>And simply giving an AI agent access to all of it doesn’t mean the agent actually understands it.</p><p>That’s the problem we’re solving at Joinable Labs.</p><h3>Retrieval Was Built for Finding Information. Agents Need Something More.</h3><p>The first generation of enterprise AI largely approached this problem through retrieval.</p><p>Take documents. Break them into chunks. Create embeddings. Put them into a vector database. When someone asks a question, retrieve the most relevant chunks and send them to an LLM.</p><p>That approach helped make enterprise chatbots and RAG applications possible.</p><p>But agents raise the stakes.</p><p>A chatbot can return an imperfect answer and ask a human to decide what to do next.</p><p>An agent may actually <strong>do something</strong>.</p><p>It might remediate a security vulnerability, modify infrastructure, respond to a customer, approve a workflow, investigate an incident, or make a recommendation that triggers an operational decision.</p><p>For those systems, finding a few pieces of text that appear relevant isn’t enough.</p><p>The agent needs to know:</p><ul><li>Is this information current?</li><li>Is it authoritative?</li><li>What does it actually mean?</li><li>How does it relate to other information?</li><li>Who is allowed to see it?</li><li>Where did it come from?</li><li>Can it safely be acted upon?</li><li>Who needs to be notified?</li></ul><p>This is why we believe the next generation of enterprise AI requires something beyond retrieval.</p><p>It requires <strong>knowledge refinement</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6CQjMTBHh-p3Fm9P4P3H4w.png" /></figure><h3>From Raw Information to Agent-Ready Knowledge</h3><p>We recently <a href="https://www.businesswire.com/news/home/20260722988532/en/Joinable-Labs-Launches-Joinable-Propagator-the-Trusted-Knowledge-Foundry-for-the-Agentic-Enterprise">announced</a> <strong>Joinable Propagator™</strong>,<strong> </strong>built around a simple idea:</p><p><strong>Before enterprise knowledge can reliably power AI agents, it needs to be refined.</strong></p><p>Think about the semiconductor industry.</p><p>A foundry doesn’t build the finished smartphone, autonomous vehicle, or AI server. It takes raw materials and produces the refined chips those systems depend on.</p><p>We believe enterprise AI needs a similar layer.</p><p>Propagator is what we call a <strong>Trusted Knowledge Foundry</strong>.</p><p>It takes the raw, unstructured information scattered throughout an organization and turns it into <em>self-updating</em>, structured, validated, permission-governed knowledge that AI agents can safely reason from and act upon.</p><p>We’re not trying to build every enterprise’s agents.</p><p><strong>We’re building the knowledge foundation that lets enterprises trust the agents they build.</strong></p><h3>What Does “Refining” Knowledge Actually Mean?</h3><p>Propagator doesn’t simply copy documents into another database.</p><p>It extracts the operational knowledge, business logic, entities, relationships, procedures, and context buried inside enterprise content.</p><p>That information is transformed into structured units we call <strong>Data Cards</strong>.</p><p>A Data Card represents a validated instance of knowledge that can be consistently recognized and reused across an organization’s systems.</p><p>Instead of an agent repeatedly searching through thousands of disconnected document fragments and attempting to reconstruct meaning every time it receives a request, it can reason from a refined knowledge layer that already understands the underlying information.</p><p>And importantly, the system is designed to know when it <strong>doesn’t know</strong>.</p><p>When Propagator isn’t sufficiently confident about how something should be classified or modeled, it can route that information to a human for review rather than simply guessing.</p><p>The result is a knowledge layer that can become more useful without sacrificing trust.</p><h3>Enterprise Knowledge Is Connected. AI Should Understand Those Connections.</h3><p>There’s another problem with traditional document retrieval.</p><p>Companies don’t operate as collections of independent documents.</p><p>The same customer might appear in a CRM, contract, support ticket, email thread, invoice, and product database.</p><p>The same vendor might appear across procurement documents, security reviews, contracts, and finance systems.</p><p>The same security vulnerability might appear across threat intelligence, internal procedures, remediation playbooks, incident reports, and infrastructure data.</p><p>These aren’t separate pieces of knowledge.</p><p>They’re connected.</p><p>Propagator resolves those relationships so agents can reason across enterprise knowledge as a connected whole rather than a collection of search results.</p><p>That distinction becomes increasingly important as AI moves from answering questions to actually performing work.</p><h3>Trust Has to Be Built Into the Knowledge Layer</h3><p>Making enterprise information understandable to AI is only half the problem.</p><p>The other half is governance.</p><p>An enterprise AI system cannot simply centralize sensitive company information and allow every agent to search through it.</p><p>Permissions matter.</p><p>A lot.</p><p>Propagator keeps refined knowledge synchronized with its source and preserves the permissions governing that information.</p><p>When an agent requests knowledge, access is evaluated against the permissions of the person the agent is acting on behalf of.</p><p>In simple terms:</p><p><strong>An agent shouldn’t be able to see something the human behind it isn’t allowed to see.</strong></p><p>Propagator also maintains provenance and auditing so organizations can understand where knowledge came from and how it was accessed.</p><p>And because the knowledge layer is exposed through both a native Model Context Protocol (MCP) server and standard REST APIs, enterprises can connect it to the agent frameworks and AI systems they already use.</p><h3>Cybersecurity Shows Why This Matters</h3><p>One of the first enterprise deployments of Propagator came in collaboration with one of the world’s largest cybersecurity companies.</p><p>Security organizations have enormous amounts of institutional knowledge.</p><p>Standard operating procedures.</p><p>Incident response protocols.</p><p>Remediation playbooks.</p><p>Security policies.</p><p>Best practices developed through years of responding to real threats.</p><p>Yet much of that knowledge exists in documents written for security analysts — not AI agents.</p><p>We used Propagator to refine thousands of pages of this material into structured, permission-governed knowledge that security agents could reason from.</p><p>That changes the question from:</p><p><strong>“What does the AI think we should do?”</strong></p><p>to:</p><p><strong>“Based on our organization’s approved procedures, how should we respond?”</strong></p><p>That’s a fundamentally different foundation for enterprise AI.</p><p>Instead of relying primarily on generic internet knowledge, an agent can reason from the organization’s own approved procedures and institutional expertise.</p><p><strong>Joinable </strong><a href="https://www.helpnetsecurity.com/2026/08/04/joinable-labs-threat-map/"><strong>Runbooks</strong></a>, our security remediation product, is built on this same foundation — Propagator — transforming SOC playbooks and security procedures into agent-ready knowledge that can power security remediation agents.</p><h3>Cybersecurity Is Just the Beginning</h3><p>The same problem exists throughout the enterprise.</p><p>Customer support teams have years of support knowledge.</p><p>IT organizations have operating procedures and troubleshooting documentation.</p><p>Legal departments have contracts, policies, and institutional precedent.</p><p>Manufacturing organizations have technical procedures and operational expertise.</p><p>Sales organizations have customer intelligence and playbooks.</p><p>HR teams have policies and processes.</p><p>Finance organizations have procedures, controls, and reporting knowledge.</p><p>Every department already possesses enormous amounts of valuable institutional knowledge.</p><p>The problem is that most of it remains trapped in formats designed for humans.</p><p>We believe one of the biggest opportunities in enterprise AI is turning that institutional knowledge into something AI can safely trust and use.</p><h3>The Next AI Advantage Won’t Just Come From Better Models</h3><p>Foundation models will continue getting better.</p><p>Agents will continue becoming more capable.</p><p>But increasingly, companies will have access to many of the same models and agent frameworks.</p><p>The real competitive advantage may come from something those models don’t have:</p><p><strong>Everything your organization knows.</strong></p><p>Your processes.</p><p>Your expertise.</p><p>Your history.</p><p>Your customer knowledge.</p><p>Your operating procedures.</p><p>Your proprietary data.</p><p>Your institutional memory.</p><p>That’s the knowledge that makes one company’s AI fundamentally different from another’s.</p><p>The challenge is turning it into something machines can understand without sacrificing the governance, permissions, provenance, and context that make that knowledge trustworthy.</p><p>That’s why we built <strong>Joinable Propagator™</strong>.</p><p>It’s not another LLM.</p><p>It’s not another chatbot.</p><p>And it’s not another agent framework.</p><p><strong>It’s the Trusted Knowledge Foundry for the agentic enterprise — transforming raw institutional knowledge into refined intelligence that AI agents can safely act on.</strong></p><p>Joinable technology is already powering more than <strong>140,000 AI projects</strong>, and we believe we’re still at the beginning of a much larger shift.</p><p>The enterprises that win the agentic era won’t simply be the ones with the most AI.</p><p><strong>They’ll be the ones whose AI understands what their organization knows.</strong></p><p>Brian Shin</p><p>Co-founder, Joinable Labs</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a7894b2f045a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/joinable/enterprise-ai-doesnt-have-a-model-problem-it-has-a-knowledge-problem-a7894b2f045a">Enterprise AI Doesn’t Have a Model Problem. It Has a Knowledge Problem.</a> was originally published in <a href="https://medium.com/joinable">Joinable</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I invested in Groq]]></title>
            <link>https://medium.com/@brianshin/why-i-invested-in-groq-4aae3f249459?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/4aae3f249459</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Mon, 05 Jan 2026 22:00:26 GMT</pubDate>
            <atom:updated>2026-01-05T22:36:37.429Z</atom:updated>
            <content:encoded><![CDATA[<p>By now, many of you have heard the news that Nvidia and Groq are forging a <a href="https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale">$20bn partnership</a>. Lots of people have been asking me about it so I wanted to share some thoughts.</p><p><strong>TLDR</strong>: Groq is the largest investment we made in VitalStage Fund I, and we are very excited for them to partner deeply with Nvidia. In fact, in our original investment memo, we predicted that one of the large chip design firms like Nvidia would seek to acquire Groq within 2–3 years. Why? I will try to explain:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*RXA19tDEvetKxOMkfyl8fw.jpeg" /></figure><p>I originally started <a href="https://vitalstage.com/">VitalStage Ventures</a> to fund startups that would fill gaps in the AI ecosystem that I had encountered firsthand in my many years as an AI Builder. Our belief was that in order to outperform in early stage venture, especially when focusing on a technical area, deep domain expertise is required. You have seen this play out with DeepTech and BioTech venture capital as examples. VitalStage is solely focused on backing (and selectively launching) early stage AI startups, and our team is made up of multiple AI startup founders and AI leaders from companies such as Google, Amazon, and Dropbox.</p><p>When we started VitalStage in 2023, we were in the AI <strong>Age of Possibilities</strong>. After OpenAI triggered the ‘mass consumerization of AI’ with the launch of ChatGPT in Nov. 2022, companies had serious FOMO and felt pressure to keep up with the Joneses. Board meetings of companies large and small, public and private, had a fairly consistent theme of “What is our AI strategy? Our competitor just announced something AI-related. What are we doing?” Nearly every publicly traded software company faced questions during earnings calls about when new AI capabilities would be available. The companies that leaned into AI were rewarded with increased stock prices, while those that didn’t were viewed as laggards.</p><p>This AI FOMO fueled dramatic increases in demand for AI Infrastructure including GPUs, GPU servers, and Data Centers, which triggered subsequent rise in associated company valuations.</p><p>Now in 2026, the AI market has transitioned to the <strong>Age of Outcomes</strong>. There is increased scrutiny on AI projects and budgets, and factors like depreciation of capital expenditures related to AI are now in focus. Critical measures around <a href="https://tti.dev/">Time-to-Intelligence</a>, accuracy, and AI ROI are top of mind.</p><p>This need to justify AI spending brings us to the Chip market. Broadly speaking, related to AI you have 4 major classes of chipsets: CPUs, GPUs, TPUs, and NPUs. Nvidia became the most valuable company in the world primarily because of the demand for GPUs during the Age of Possibilities. Why? GPUs are general purpose processors that parallel process tasks and are really good at matrix multiplication, the core mathematical operations underpinning graphics acceleration, cryptocurrency mining, and AI. As the AI Foundation Model Arms Race played out, demand continued to outpace supply.</p><p>TPUs (or Tensor Processing Units) and NPUs (neural processing units) are both more specialized chipsets for matrix multiplication acceleration. You can think of TPUs as powering big servers in a data center and NPUs as consumer-device level chips that consume far less power. But there are new approaches and architectures coming. For example, years ago, I led an investment into a next generation photonic chip company called <a href="https://lightmatter.ai/">Lightmatter</a> (now a Unicorn) that could impact the landscape in a few years.</p><p>But more near-term, it was our belief back in 2024 that more specialized AI chips like TPUs and NPUs would become increasingly relevant as AI continued to evolve. That finally brings us to Groq. Having been a long-time admirer of Google’s TPU (Tensor Processing Units), I had been following Jonathan Ross’ work as he was the architect of the TPU. In December 2023, one of the AI startups in VitalStage’s portfolio pointed to a demo that Groq had launched at <a href="http://groq.com/">groq.com</a>. The speed was unreal. Essentially, Groq was hosting an open source model and running a “pepsi challenge” to show how much faster their chips processed AI model responses. Soon after, Groq acquired Definitive Intelligence to launch Groq’s cloud offering, making it possible to “rent” Groq compute. It was around this time that we started to dig into Groq more deeply and develop our view on how this part of the market would play out.</p><p>We posited that AI Training would continue to get more efficient with advancement in training techniques such as Iterative Refinement. With AI Training often viewed as a cost center, we felt that chip designers focused on AI Training would face some challenges with cost justification.</p><p>Instead, we believed that chip design focused on advancing AI Inference would be a better bet. Groq’s combination of significantly faster performance, lower cost, and reduced heat generation seemed like a compelling choice.</p><p>So back in 2024, we made the decision to invest in Groq. In the original investment memo I wrote, I suggested that a likely outcome for Groq would be that a major chip designer such as Nvidia might feel the need to acquire Groq. Our thinking was that complementing their more general-purpose GPU offerings with more of a specialized, lower-cost offering could solidify their hold on the AI Inference (money-making) side of the AI market.</p><p>At the time, we had not yet completed the fundraise for VitalStage Fund I, but I felt very strong conviction. So I wrote a large personal check to invest in Groq and then warehoused it for VitalStage. I was both excited and nervous as it was a pretty significant amount of money for me personally, and there was no guarantee that we’d complete a successful fundraise for VitalStage Fund I. And even if things worked out, I’d make far less money personally by warehousing the Groq investment in my fund. But I strongly believed that the venture market needed hyper-domain-focused early stage investors promoting the responsible advancement of AI. And I bet that building the best AI portfolio focused on AI Infrastructure and Specialized Applied AI would be more valuable in the long run.</p><p>So we made the investment in Groq, and we did up completing a small but successful raise for Fund I. Now we’ve been told that Fund I is among the highest performing early-stage emerging venture funds. It’s not that we’re extra smart or anything; it’s just that our entire team has lived and breathed AI since before it was cool, so we have a bit more context than others.</p><p>We continue to be very excited about the AI market especially in the Age of Outcomes. We’re raising VitalStage Fund II because of this optimism. The AI startups that can demonstrably create 20x, 50x, 100x positive impact will do extremely well. And we work hard so that <a href="https://vitalstage.com/">VitalStage Ventures</a> can keep on top of the developments in AI well enough to partner with the best responsible AI innovators.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4aae3f249459" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why We Started Joinable Labs: Fighting for the AI Builder]]></title>
            <link>https://medium.com/@brianshin/why-we-started-joinable-labs-fighting-for-the-ai-builder-d4625c8207ac?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4625c8207ac</guid>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Wed, 06 Aug 2025 22:33:44 GMT</pubDate>
            <atom:updated>2025-08-06T22:34:32.855Z</atom:updated>
            <content:encoded><![CDATA[<p>Since 2018, I’ve been on a journey as an AI Builder. Like so many others, I’ve felt the thrill of early breakthroughs, the frustrations of fragile workflows and processing pipelines, and the deep responsibility of helping organizations harness a powerful, fast-moving technology.</p><p>Back in 2019, I co-founded my first AI startup called Searchable.ai. We assembled an amazing founding team, built one of the earliest “answer engines,” and co-authored two foundational patents in natural language question answering. Back then, you not only had to build the AI/ML models yourself for the most part, you also had to build pretty much the entire pre-processing and pre-training infrastructure from scratch. That experience taught me not just the technical complexity of developing AI, but the even greater complexity of aligning people around a shared vision for applying AI — executives, engineers, product teams, and customers alike.</p><p>This experience led me to join a large public software company to lead AI strategy and product management. There, I got a front-row seat to the new challenges AI Builders face in the enterprise.</p><p>Pressure was coming from all directions. Board members were anxious about competitors announcing flashy AI features. Investors wanted reassurance that we had a strong AI roadmap. Internally, teams were overwhelmed by unfamiliar tools, new terminology, and unrealistic expectations.</p><p>Despite these challenges, we were able to rally together to launch six new AI capabilities across the company. The company’s market cap more than doubled in less than a year. But even with those wins, I couldn’t shake the feeling: <strong>there had to be a better way</strong>.</p><p>It was still far too difficult — too slow, too much friction — for AI Builders to go from raw data to working, production-grade AI systems that accurately and effectively help users. For example, submitting proprietary company knowledge to a commercially hosted AI model provider was a non-starter from a customer privacy and compliance perspective. Enterprises that need to protect their data often have to build everything from scratch to go from messy, fragmented knowledge in multiple data formats to meaningful AI-ready data.</p><p>Adding to the difficulty is the fact that for AI Builders, the milestones, costs, and resources to power custom, Private AI that delivers positive impact were still often nebulous and ever-changing. I felt the need not just for better AI dev tools and AI infrastructure but also for a unifying framework and metric to align stakeholders around a shared vision for Private AI. I started working on a thesis I called “<a href="https://www.tti.dev/"><em>Time-to-Intelligence</em></a><em>” acceleration — </em>focused on trying to get various stakeholders aligned and bought into a shared framework for developing AI and a new class of tools that would accelerate development.</p><h3>From Frustration to Founding</h3><p>The frustrations I encountered as an AI Builder is what led me to launch what was at the time the first AI-native venture capital fund called <a href="https://www.vitalstage.com/">VitalStage Ventures</a>. It’s called VitalStage because we believed that AI was at an inflection point and that as established AI leaders, we could help AI Founders navigate the vital stages of value creation in a rapidly changing landscape. So far, we’ve backed 19 remarkable AI startups. But the one I was looking for — the platform I wished I had as an AI Builder — never came along.</p><p>That’s when I teamed up with three established Web 3 entrepreneurs who have collectively been a part of over a dozen Web 3 Unicorns. Together, we set out to find AI startups to invest in that were building real technical depth with a decentralized infrastructure vision. But again, we kept coming up empty. We saw plenty of AI hype projects and lots of clever demos, but not the kind of core platform we believed the next generation of AI Builders desperately needed.</p><p>So I started building it.</p><p>I began prototyping a system to extract, parse, and prepare proprietary knowledge — documents, spreadsheets, PDFs, images — for use in AI systems. I wanted a platform that would drastically shorten the time it takes to go from proprietary organizational knowledge to Private AI, all while maintaining enterprise-grade privacy, control, and security. And selfishly, I wanted to get back into the code as an aging but still-excited geek.</p><p>When I showed an early version of this technology to my three Web 3 partners, something clicked. We decided to stop searching for a company to invest in, take what I had built, and leverage it to start a new Web 3 AI company.</p><h3>The Birth of Joinable Labs</h3><p>We co-founded <a href="https://www.joinable.ai/"><strong>Joinable Labs</strong></a> with a shared vision: to give AI Builders the tools to accelerate the time from concept to prototype, from prototype to production, and from production to effective AI system up to 50x faster.</p><p>At Joinable, we have solidified my original thesis into a unifying metric for AI Builders — <a href="https://www.tti.dev/"><em>Time-to-Intelligence</em></a> (TTI). And everything we do is designed to reduce that TTI — across data ingestion, model selection, evaluation, deployment, and iteration.</p><p>We’re building the platform I wish I had years ago. Not just as an AI Builder. Not just as a product leader. But as someone who believes that AI will reshape how knowledge is captured, shared, and applied — if we give AI builders the right foundation and the tools needed to increase the probability of AI success.</p><p>This is why we started Joinable:<br>To make building AI systems faster, more accessible, and less painful.<br>To help teams align on what’s possible, what it will take, and how to deliver it.<br>To create a new kind of Web 3 AI utility company that delivers real value while empowering community member participation.</p><p>And to help a new generation of AI Builders bring their ideas to life.</p><p>If you’ve ever felt the drag of fragmented tools, or the weight of unrealistic expectations for AI projects — Joinable was built for you.</p><p>Let’s accelerate Time-to-Intelligence, together.</p><figure><img alt="Joinable is the first Time-to-Intelligence Acceleration Platform to Turn Raw Data into Private AI up to 50x Faster" src="https://cdn-images-1.medium.com/max/357/0*JwZQhwhe7-z16MEp" /></figure><p>— <em>Brian Shin<br></em> Co-founder, Joinable Labs</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4625c8207ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Getting Back In The Saddle]]></title>
            <link>https://medium.com/@brianshin/getting-back-in-the-saddle-f59794bb80f3?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/f59794bb80f3</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Mon, 28 Oct 2019 02:57:31 GMT</pubDate>
            <atom:updated>2019-10-28T02:57:31.892Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Getting Back In The Saddle</strong></p><p>Startups are tough! Recent data indicates that over 90% of startups do not end up succeeding. There are grumblings that we’re at the tail end of a historic bull market. The days of startup growth at all cost may be over. So it is with this sunny context that I’m announcing that I’m leaving the comfortable investment world and starting a new company, searchable.ai, with 3 amazing co-founders: Nick Martin, Aaron Sisto, and Hung Nguyen. Great timing, I know! To compound matters, I’m also moving with my wife and kids from Boston to San Francisco, even as Boston’s startup scene is rocking and CA wildfires are raging. Why, you may ask??</p><p><strong>Retired from Entrepreneurship</strong></p><p>When my most recent two startup companies were acquired in early 2017, I thought that my time as an entrepreneur was over. After 8 startups and 7 exits, I was too burned out, too old, too slow. I decided to move onto a new role as a full time investor in technology companies. That way, I could still have the opportunity to work with great entrepreneurs while being able to be a step removed from the day to day rollercoaster of startup life.</p><p>By this point, I had been an active angel investor for more than 10 years in over 50 companies started by friends, colleagues, and people trusted by friends. Some of these companies include HubSpot, Misfit, Drift, Frequency Therapeutics, and Kayak. But I recognized that being a full time investor would require a higher level of discipline and diligence. I was very fortunate to have had some amazing learning experiences and to have led some really great deals as a technology investor with two fantastic firms, In-Q-Tel and Providence Equity Partners. But as I kept meeting with phenomenal entrepreneurs, a longing for the feeling of “creating something from nothing” kept bubbling up.</p><p>Honestly, even with this longing, I did not imagine that I would start another company. I had made so many mistakes as an entrepreneur and felt like at various times that I had let people down. The thought of starting from scratch again after so long just seemed incredibly daunting, especially with the stress that many startup founders endure.</p><p>Early stage companies can struggle for any number of reasons including market timing, strong competition, suboptimal execution, disagreements among founders, a lack of focus, and more. An often overlooked aspect hindering startup success can be the mental and physical toll that startups can take on founders and employees. As a startup founder, one can experience dramatic highs and traumatizing lows in the course of a single day. The role of CEO is often described as the “loneliest job,” because there are so many topics that can be difficult to discuss with colleagues or board members. CEO’s and founders often try to be strong for their teammates, but in the process may end up absorbing a lot of angst that can eat away at them.</p><p><strong>Recognizing a Problem</strong></p><p>But I started to see an opportunity that I couldn’t stop thinking about. It is a problem that people in the workplace run into over and over again. After significant research, I could not find a suitable solution. I asked some colleagues and they all said that they had the same sort of issues when they were working. So Nick, Aaron, Hung, and I started hacking on a new project to try to address these problems. I even set aside some money to self-fund this project thinking that I would not want to raise outside capital given all the additional expectations and pressures that would come with it.</p><p>We started to make some progress. We found that many companies experience the problems that we were seeing. The opportunity seemed real. So we dug into building a new technology to address this problem. That is when a good friend, Neil Sequeira, a leading venture capitalist who had backed my last company, suggested that we raise some venture capital to pursue this idea. In fact, he offered for his firm Defy Partners to lead our round.</p><p><strong>Hemming and Hawing</strong></p><p>I had a lot of hesitation. First, I felt that I had lost my entrepreneurial mojo. Entrepreneurs need to have optimism to be able to see a future that can be better than it is today combined with the drive to make it happen. I felt optimistic about us being able to improve how people work with this new technology we were building. But maybe I just lacked the self-confidence to go for it.</p><p>I highlight some of the challenges that I was going through because I know that there are people out there struggling with similar thoughts. Do you have an idea but you worry that you don’t have what it takes to be a founder? Maybe you just don’t know where to begin. Do you have a comfortable job and worry about the risks of starting a company? These types of concerns are partially why there are a lot of entrepreneurs under the age of 30. Why? Because younger folks tend to have lower opportunity costs. If you are struggling with the stress, loneliness, or financial pressures of “going for it”, believe me, you are not alone!</p><p><strong>Deciding to Go for It</strong></p><p>Defy Partners’ vote of confidence helped me tremendously. Having great partners (and a big challenge to dig into) really solidified the decision to go all in. Our founding team consists of Nick Martin, co-founder and Chief Technology Officer, Aaron Sisto, co-founder and Chief Scientist, and Hung Nguyen, co-founder and Chief Marketing Officer. Nick, Aaron and I had worked together at In-Q-Tel, a strategic investment firm working on behalf of the US Intelligence Community, where we made investments in innovative technologies that had the potential to help improve national security. Hung and I had worked together for years at our previous company, Visible Measures. Hung more recently co-founded and sold an analytics startup called Ozmotik, where I served as an advisor.</p><p>There are a bunch of things that I love about our founding team: 1) We have all started companies, so we know how hard it is 2) We have all worked together before at various times 3) We all have experience working with and building AI-driven systems and 4) All of us had to give up other opportunities to pursue searchable.ai full time.</p><p>Soon, we will go into detail in terms of what problem we are solving and what solution we are building. But for now, I just wanted to give a little background on our origin story while also sharing some words of encouragement for those of you out there contemplating developing the next big thing! Startups are hard; you never know how they are going to go. But despite some good reasons to the contrary, I decided that it was time to get back in the saddle and go for it. Very excited to be building something with great partners like Nick, Aaron, Hung, Neil @ Defy, and all the incredible folks who are backing us!</p><p>.</p><p>Brian Shin</p><p>Co-founder &amp; CEO, searchable.ai</p><p>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f59794bb80f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Requiem for a Dream]]></title>
            <link>https://medium.com/@brianshin/requiem-for-a-dream-eb7b068e4a6d?source=rss-8bfe5ed3a108------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb7b068e4a6d</guid>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[programmatic]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[brianshin]]></dc:creator>
            <pubDate>Sat, 01 Apr 2017 00:33:34 GMT</pubDate>
            <atom:updated>2017-04-01T00:33:34.398Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="http://www.visiblemeasures.com/">Visible Measures</a> has been acquired and today is my last day with the company I love.</p><p>I started working on what would become Visible Measures during the summer of 2004, right before business school. This was before YouTube, before the iPhone, before Facebook got to 100,000 users, before video ad networks became large and then fell out of favor, before the rise of programmatic advertising, and before media spend consolidated with The Duopoly and agency trading desks. It was a simpler time :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*TfbtFcjN_qWRbHuQh6Mf6Q.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vrkZmVnN7iAGNCoJ2wjuyg.jpeg" /><figcaption>Accepting a MIT $50k Semi-Finalist Award in 2005. Our first office (1 cubicle in Kendall Sq. Cambridge)</figcaption></figure><p>We had a dream to create a currency for video — a standardized way to benchmark performance and value video reach. A lot of people believed in us and some fantastic angel investors and venture capital firms enabled us to pursue this dream. We grew a lot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sBgjZ6R8cmPZzFqQDKWUJQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5x0lX_WfsCwkRD56qqwxcQ.jpeg" /></figure><p>I am very proud of what we built as Visible Measures evolved into a leading platform for video measurement &amp; activation. As we grew, we navigated tectonic shifts in the advertising landscape and some internal struggles. But some of the challenges in our industry are evident in that after almost 13 years, video still does not have a true currency. Perhaps companies and brands will throw more weight behind Attention as a metric — we will see.</p><p>At a later time, after some decompression, I will share more about some of the lessons we learned at Visible Measures. Above all, I am so grateful to have had the opportunity to pursue this dream with some of the finest teammates, advisors, investors, clients, and partners one could ever hope for. And we really did live the VMC motto of WORK HARD / PLAY HARD.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*ax2lz_PT9ui--coEUOk14A.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ek_KbQzWf0EuXji__9HEFw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5SJR0uw5B81kWqGCliIqqw.jpeg" /><figcaption>Work Hard / Play Hard</figcaption></figure><p>I am happy that so many VMCers have a new home with our acquirer, <a href="http://www.acuityads.com/">AcuityAds</a>. Despite what is a bittersweet ending, we definitely made an outsize impact along the way:</p><ul><li>Launched the first video analytics solution for publishers</li><li>Granted over a dozen patents ranging from distributed aggregation of analytics data to audience targeting using video viewing behavior</li><li>Employed more than 200 amazing people</li><li>Built <a href="http://www.truereach.org/">True Reach</a> into a metric used by thousands of marketers to evaluate video campaign performance</li><li>Scaled up to handling massive programmatic scale</li><li>Tracked 7,000+ brands, across 500mm videos, and 6 trillion data points</li><li>Delivered almost $200mm in revenue</li><li>Invented the 1st and still only way to competitively benchmark all brand competitors in 281 industry categories in terms of viewership, time spent, AND dollars invested</li></ul><p>I am so grateful for my team (particularly our OGs Ron, Seraj, Meng, and Gary; and my staff Jeff, Perry, Annalisa and Meaghan), our board (special thanks to Larry and George from <a href="http://www.gcvp.com/">GC</a>, Bill from <a href="http://www.mdv.com/">MDV</a>, Prof. <a href="http://mitsloan.mit.edu/faculty-and-research/faculty-directory/detail/?id=19553">Roberts</a> and <a href="http://www.sharethis.com/">Kurt</a> A.) and our all of our investors. Your unwavering support, incredibly hard work, innovative thinking, and perseverance meant the world to VMC.</p><p>It has been an amazing journey. I owe so much to this experience for what it has taught me, the people I have met, and the opportunities it afforded. I am sincerely honored to have had the priviledge of being the Founder &amp; CEO of Visible Measures and I know that I will look back on this time as the most formative of my career.</p><p>With sincere gratitude,</p><p>Brian</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FiIXNvlAcbXc%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DiIXNvlAcbXc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FiIXNvlAcbXc%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/cc76071ba95d1baa9ca961248cf42197/href">https://medium.com/media/cc76071ba95d1baa9ca961248cf42197/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb7b068e4a6d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>