<?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 João Felipe on Medium]]></title>
        <description><![CDATA[Stories by João Felipe on Medium]]></description>
        <link>https://medium.com/@JhonnyPhilipe?source=rss-5e39825a7d65------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*4jc8VJ8-DjS0yfAPkUbgEA@2x.jpeg</url>
            <title>Stories by João Felipe on Medium</title>
            <link>https://medium.com/@JhonnyPhilipe?source=rss-5e39825a7d65------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 00:57:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@JhonnyPhilipe/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[Are We Entering the AI Startup Loop?]]></title>
            <link>https://medium.com/@JhonnyPhilipe/are-we-entering-the-ai-startup-loop-6ffb4b1466c7?source=rss-5e39825a7d65------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ffb4b1466c7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[João Felipe]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 15:40:38 GMT</pubDate>
            <atom:updated>2026-08-02T15:42:20.235Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the past few years, the AI industry has entered an unprecedented period of growth.</p><p>According to the Stanford AI Index 2026, more than 2,000 AI companies received venture funding between 2023 and 2025. At the same time, billions of dollars have flowed into AI infrastructure, foundation models, and enterprise applications.</p><p>This rapid expansion raises an interesting question: Why are so many AI companies being created?</p><p>A few years ago, the race was about building the largest language model. Today, the landscape has changed. Open-source models, lower infrastructure costs, and accessible APIs have significantly reduced the barriers to entry.</p><p>As a result, creating an AI company is no longer reserved for organizations with virtually unlimited resources.Instead, the market is shifting toward specialization.</p><p>Rather than trying to build the next general-purpose model, many startups are focusing on solving highly specific business problems. We are seeing AI designed for healthcare diagnostics, cybersecurity, legal research, financial analysis, software engineering, manufacturing, logistics, education, and dozens of other verticals.</p><p>The competitive advantage is gradually moving away from simply having a larger model.</p><p>Instead, success increasingly depends on combining domain expertise, proprietary data, workflow integration, and measurable business outcomes.</p><p>This is also why the industry sometimes feels repetitive.</p><p>Every few weeks, another company announces “the world’s most powerful model for its size.” While these improvements are often technically meaningful, they also reflect an intensely competitive market where performance gains are becoming increasingly incremental.</p><p>The next stage of AI may not be defined by who builds the biggest model.</p><p>It may be defined by who builds the most useful one.</p><p>Of course, this growth also raises important questions.</p><p>Will every niche sustain enough demand to support dozens of AI startups?</p><p>Will consolidation become inevitable?</p><p>Are we witnessing the formation of an AI bubble, or simply the natural evolution of a technology that is becoming as fundamental as cloud computing or the internet itself?</p><p>History suggests that disruptive technologies often experience periods of exuberance before the market matures. Many companies disappear, some merge, and a smaller number become industry leaders. AI may follow the same path.</p><p>The difference is that this time, the pace of innovation is unlike anything we’ve seen before.</p><p>The companies that survive may not be those with the loudest marketing or the largest models.</p><p>They will likely be the ones that solve real problems, in the right niche, at the right time.</p><p>Source: Stanford AI Index 2026 — Economy Chapter</p><p>https://hai.stanford.edu/ai-index</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/627/1*f5RrCKNDzYhbOJvZrcZIww@2x.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ffb4b1466c7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Signs Your Workplace Is Toxic (And Why It Matters)]]></title>
            <link>https://medium.com/@JhonnyPhilipe/5-signs-your-workplace-is-toxic-and-why-it-matters-e8c06ae0f2d5?source=rss-5e39825a7d65------2</link>
            <guid isPermaLink="false">https://medium.com/p/e8c06ae0f2d5</guid>
            <category><![CDATA[jobs]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[work-life-balance]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[mental-health]]></category>
            <dc:creator><![CDATA[João Felipe]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 13:33:39 GMT</pubDate>
            <atom:updated>2026-07-31T13:45:05.318Z</atom:updated>
            <content:encoded><![CDATA[<p>For years, we’ve seen a corporate culture where performance is often treated as the highest priority. Delivering results is important, but when performance becomes more important than people, organizations start paying a hidden price: burnout, disengagement, high turnover, and declining innovation.</p><p>The challenge isn’t high standards. High-performing organizations can also be healthy places to work.</p><p>The real problem begins when pressure replaces leadership.</p><p>Here are five of the clearest signs that you may be working in a toxic environment.</p><ol><li>Fear is the primary management tool</li></ol><p>Employees avoid asking questions, admitting mistakes, or challenging decisions because they’re afraid of being blamed, embarrassed, or punished.</p><p>This creates what researchers call low psychological safety — one of the strongest predictors of poor team performance and low innovation.</p><p>Questions to ask yourself:</p><ul><li>Are people afraid to speak up in meetings?</li><li>Do mistakes lead to learning or punishment?</li><li>Is bad news hidden instead of discussed?</li></ul><p>⸻</p><p>2. Burnout is normalized</p><p>Working late every day, answering messages at midnight, and sacrificing weekends are treated as commitment instead of warning signs.</p><p>Occasional busy periods happen in every company.</p><p>When they become the expectation rather than the exception, the organization is creating a sustainable burnout cycle instead of sustainable performance.</p><p>The World Health Organization officially recognizes burnout as an occupational phenomenon resulting from chronic workplace stress that has not been successfully managed.</p><p>⸻</p><p>3. Recognition is rare, but criticism is constant</p><p>Feedback should help people improve.</p><p>In toxic environments, employees mainly hear from leadership when something goes wrong.</p><p>Over time, motivation decreases because people begin to feel that no matter how much they contribute, it is never enough.</p><p>Recognition doesn’t require bonuses — it requires leaders who notice good work.</p><p>⸻</p><p>4. Politics matter more than results</p><p>Promotions, visibility, and opportunities become driven by relationships, internal politics, or personal preferences rather than competence and impact.</p><p>When employees stop believing that merit matters, trust disappears.</p><p>Eventually, high performers either disengage or leave.</p><p>⸻</p><p>5. Leaders create stress instead of clarity</p><p>Unclear priorities, changing directions without explanation, unrealistic deadlines, and inconsistent communication create unnecessary anxiety.</p><p>Good leaders don’t eliminate pressure.</p><p>They reduce uncertainty.</p><p>People can handle difficult challenges far better than constant confusion.</p><p>⸻</p><p>Final Thoughts</p><p>A demanding workplace is not automatically a toxic workplace.</p><p>Great organizations often expect excellence while also creating psychological safety, recognizing contributions, encouraging healthy debate, and respecting employees’ well-being.</p><p>The healthiest companies understand a simple principle:</p><p>Sustainable performance is built on healthy people — not the other way around.</p><p>Recognizing these warning signs doesn’t necessarily mean you should leave immediately. Sometimes they can be addressed through better leadership, communication, or organizational change.</p><p>But ignoring them often comes at a much higher personal and professional cost.</p><p>⸻</p><p>References</p><ul><li>Amy Edmondson, The Fearless Organization (2018) — Psychological safety and team performance.</li><li>* World Health Organization — Burnout as an occupational phenomenon (ICD-11).</li><li>* Gallup, State of the Global Workplace (2025) — Employee engagement, manager impact, and workplace well-being.</li><li>* Harvard Business Review — Multiple studies on burnout, psychological safety, leadership, and organizational culture.</li><li>* American Psychological Association — Work in America surveys on workplace stress and employee mental health</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*NwK5k9X77CxDUGyoYR2KaA@2x.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e8c06ae0f2d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Technology Rollouts Fail — And It’s Never the Technology]]></title>
            <link>https://medium.com/@JhonnyPhilipe/why-technology-rollouts-fail-and-its-never-the-technology-31c0e3576d14?source=rss-5e39825a7d65------2</link>
            <guid isPermaLink="false">https://medium.com/p/31c0e3576d14</guid>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[telecommunication]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[cisco]]></category>
            <category><![CDATA[rollout]]></category>
            <dc:creator><![CDATA[João Felipe]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 16:04:31 GMT</pubDate>
            <atom:updated>2026-07-30T16:04:31.750Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Spreadsheet Trap</h3><p>Every rollout starts the same way. Someone builds a beautiful Excel file. Color-coded tabs. Conditional formatting. A masterpiece of organization.</p><p>Three weeks in, it’s already wrong.</p><p>Someone updated a site status in an email thread but forgot to touch the spreadsheet. A utility approval came through but the PM who handled it is on vacation. A shipment got delayed in customs — flagged in a WhatsApp group that half the team isn’t in.</p><p>The spreadsheet becomes a historical document masquerading as a live system. And nobody wants to be the one to say it out loud.</p><h3>Nobody Owns the Blockers</h3><p>Here’s the pattern I see constantly: a blocker appears, gets mentioned in a meeting, gets assigned vaguely to “the team”, and disappears into the void.</p><p>Two weeks later, same blocker. Different meeting. Same vague assignment.</p><p>In a technology rollout, blockers compound. A delayed grid connection doesn’t just push one site — it pushes every site downstream that was waiting on that milestone. One missed utility deadline can cascade into a six-month delay across an entire region.</p><p>The problem isn’t that blockers exist. Blockers always exist. The problem is that nobody has a single source of truth for who owns what, by when, and what happens if they miss it.</p><h3>The Visibility Gap</h3><p>Senior stakeholders want dashboards. Field teams want WhatsApp. Middle management wants PowerPoint decks updated every Friday.</p><p>So the PM becomes a full-time translator. Pulling data from five different places, reformatting it for five different audiences, and somehow finding time to actually manage the project.</p><p>This is where rollouts silently die. Not in a dramatic failure — in a slow erosion of PM bandwidth. The person who should be anticipating the next problem is busy reporting on the last one</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TJ6J3wVTxCpHZVFp.png" /></figure><h3>The Uncomfortable Truth</h3><p>Most technology rollout failures are documented in detail — in email threads, in WhatsApp messages, in spreadsheet version histories — and nobody connects the dots until it’s too late.</p><p>The technology worked fine. The grid connection approval was delayed 11 weeks because nobody followed up. The hardware sat in customs for 3 weeks because the broker needed a document that nobody knew was missing. The site went live 4 months late because the PM was spending 60% of their time on status updates instead of risk management.</p><p>If you’re managing a technology rollout right now, ask yourself one question: <strong>if I got hit by a bus tomorrow, could anyone on my team tell me exactly where every blocker stands, who owns it, and when it’s due?</strong></p><p>If the answer is no, that’s your real problem. Not the technology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=31c0e3576d14" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Efficiency Revolution: 4 Hidden Breakthroughs Reshaping Modern AI]]></title>
            <link>https://medium.com/@JhonnyPhilipe/the-efficiency-revolution-4-hidden-breakthroughs-reshaping-modern-ai-f0fe339577ce?source=rss-5e39825a7d65------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0fe339577ce</guid>
            <category><![CDATA[yoco]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[embedded-systems]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[rags]]></category>
            <dc:creator><![CDATA[João Felipe]]></dc:creator>
            <pubDate>Wed, 29 Jul 2026 13:57:03 GMT</pubDate>
            <atom:updated>2026-07-29T14:18:02.126Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Efficiency Revolution: 4 Hidden Breakthroughs Reshaping Modern AI</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-t_wGfFI7RrCxwj1iJWVhw.png" /></figure><h3>1. Introduction: The Invisible Walls of AI Scaling</h3><p>We have reached a pivotal moment in the evolution of Large Language Models (LLMs). While the industry remains fixated on the “scaling laws” of parameter counts, developers on the front lines are hitting invisible walls that no amount of raw compute can easily overcome. These are not limits of intelligence, but of architecture: the O(L \cdot N \cdot D) memory bottleneck of the Key-Value (KV) cache, the glacial prefilling speeds of long-sequence inputs, and the semantic fragmentation that plagues Retrieval-Augmented Generation (RAG).</p><p>The next frontier of AI is not defined by size, but by <strong>architectural efficiency</strong>. We are witnessing a fundamental shift in how models process, normalize, and retrieve information. By rethinking the fundamental building blocks of the Transformer, researchers are moving beyond the brute-force era and into a phase of hardware-software co-design that prioritizes throughput and context preservation without sacrificing expressive power.</p><h3>2. YOCO: Breaking the Memory Wall with “You Only Cache Once”</h3><p>We are witnessing the obsolescence of the standard Transformer decoder as the default unit of compute. The primary culprit is the KV cache, which in traditional architectures scales linearly with both the number of layers and the sequence length. For a 65B model, a 512K context window demands roughly 86GB of GPU memory — exceeding the capacity of a single H100. This is the “Memory Wall,” and it renders long-context AI prohibitively expensive for most enterprises.</p><p>The <strong>YOCO (You Only Cache Once)</strong> architecture, introduced by Sun et al. (2024), effectively dismantles this wall. By transitioning from a standard decoder to a <strong>decoder-decoder stack</strong>, YOCO separates the encoding of global KV caches from the generation process.</p><ul><li><strong>The Mechanism:</strong> YOCO utilizes a <strong>self-decoder</strong> (using efficient attention modules like sliding-window or gated retention) to encode global KV caches exactly once. These caches are then reused by a <strong>cross-decoder</strong> via cross-attention across all subsequent layers.</li><li><strong>Physical Compute Wins:</strong> Because the cross-decoder reuses the outputs of the self-decoder, the prefill stage can “early exit” before the cross-decoder even begins. This provides at least a 2x structural speedup. In practice, for a 512K context, YOCO slashes prefilling latency from 180 seconds to <strong>under 6 seconds</strong>.</li><li><strong>Strategic Impact:</strong> YOCO reduces KV cache memory by an staggering <strong>80x</strong> for 65B models. By moving the bottleneck from a memory-bound O(L \cdot N \cdot D) problem to a compute-bound O(N) problem, it enables trillion-token context lengths on existing hardware.</li></ul><p>“YOCO improves inference memory, prefill latency, and throughput by orders of magnitude across context lengths and model sizes… The results position YOCO as a strong candidate model architecture for future large language models with native long-sequence support.” — <strong>Sun et al., 2024</strong></p><h3>3. RMSNorm: Why “Less is More” in Neural Normalization</h3><p>In the pursuit of training stability for deep networks, normalization is the unsung hero. While <strong>LayerNorm</strong> has been the industry standard since the original Transformer, modern heavyweights like Llama 3, Mistral, and Gemma have pivoted to <strong>RMSNorm (Root Mean Square Layer Normalization)</strong>.</p><p>The strategic insight here is “simplification for speed.” LayerNorm is computationally heavy because it performs two distinct operations: re-centering (subtracting the mean) and re-scaling (dividing by standard deviation). RMSNorm drops the re-centering step entirely, focusing solely on re-scaling.</p><ul><li><strong>The Mental Model:</strong> Researchers discovered that the re-scaling step provides the vast majority of the stabilization benefit. Crucially, the <strong>residual connections</strong> within the Transformer architecture are more than capable of handling mean shifts on their own, making re-centering mathematically redundant.</li><li><strong>The “Pre-Norm” Evolution:</strong> Modern LLMs apply RMSNorm in a <strong>“Pre-norm”</strong> configuration — normalizing the input <em>before</em> it hits the attention or feed-forward (FFN) blocks. This shift is critical for the stability of deep networks, allowing for cleaner gradient flow.</li><li><strong>Efficiency Gains:</strong> By stripping away the mean subtraction and the associated “beta” parameter, RMSNorm reduces training time by <strong>7% to 64%</strong> depending on the task. This is a classic case of removing redundant math to gain raw hardware throughput.</li></ul><h3>4. Late Chunking: Solving the “Lost Context” Problem in RAG</h3><p>Retrieval-Augmented Generation (RAG) is only as effective as the semantic integrity of its retrieval phase. The industry has long suffered from the “Lost Context Problem” caused by <strong>Naive Chunking</strong>. When a document is split into fragments before embedding, pronouns and relational cues lose their referents. In a Wikipedia entry for Berlin, a chunk starting with “Its population is…” becomes an orphaned string, its semantic link to “Berlin” severed.</p><p><strong>Late Chunking</strong>, pioneered by Jina AI, provides a “free” upgrade to retrieval accuracy. The name “Late” refers to the fact that boundary cues and pooling are applied <strong>after</strong> the model has generated token-level embeddings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/687/1*3zUqB4tJCTdv01taqCuHJg.png" /></figure><p>By leveraging a long-context embedding model (like jina-embeddings-v2) to process the entire document first, every token embedding is &quot;conditioned&quot; by the surrounding text. This results in a <strong>3.63% relative improvement</strong> in retrieval tasks. Unlike Anthropic’s approach, which requires calling an external model like <strong>Claude-3-Haiku</strong> to summarize every chunk—a high-latency, high-cost endeavour—Late Chunking achieves superior semantic integrity within the existing embedding workflow.</p><h3>5. SwiGLU: Moving from Static Math to Dynamic Gating</h3><p>The activation function is the “gatekeeper” of a model’s expressive power. Traditional models relied on ReLU or GELU, which use static thresholds to filter signals. Modern FFN blocks have moved to <strong>SwiGLU</strong>, a hybrid of the Swish (or SiLU) function and Gated Linear Units (GLU).</p><p>Instead of a simple “on-off” switch, SwiGLU uses <strong>dynamic gating</strong>. It splits the input into two branches: a “gate” and a “value.” The input is transformed by two parameterized linear layers that learn <em>how</em> to route information based on the data itself.</p><ul><li><strong>The Technical Edge:</strong> The SiLU/Swish component is key because it ensures <strong>smooth, non-zero derivatives almost everywhere</strong>. This prevents the “dead neuron” problem that plagues ReLU, where neurons can become permanently inactive during training.</li><li><strong>Expressive Power</strong>: By replacing the entire hidden layer mechanism with this data-dependent routing, models like Llama 3 gain the expressive power needed to capture complex, non-linear dependencies. This architectural choice is now the standard for modern FFN blocks, providing a significant boost to training stability and performance in massive context windows.</li></ul><h3>6. Conclusion: The Path to Trillion-Token Context</h3><p>The common thread linking YOCO, RMSNorm, Late Chunking, and SwiGLU is the transition from brute-force compute to <strong>hardware-aware architectural efficiency</strong>. We are entering an era where the software architecture is designed to respect the physical limits of <strong>SRAM and H100 memory bandwidth</strong>.</p><p>By removing the KV cache ceiling (YOCO), simplifying normalisation (RMSNorm), preserving semantic integrity (Late Chunking), and enabling dynamic routing (SwiGLU), we have cleared the path toward models that can ingest entire libraries, codebases, and video streams in real time.</p><p>As we look toward the next generation of AI, the question for every strategist and developer changes: <strong>If memory is no longer the primary bottleneck for scaling, what becomes the next limit for AI?</strong> The future belongs to those who look beyond parameter counts and master the models&#39; internal efficiencies.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0fe339577ce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Não deduza, converse]]></title>
            <link>https://medium.com/@JhonnyPhilipe/n%C3%A3o-deduza-converse-e96528a623f?source=rss-5e39825a7d65------2</link>
            <guid isPermaLink="false">https://medium.com/p/e96528a623f</guid>
            <dc:creator><![CDATA[João Felipe]]></dc:creator>
            <pubDate>Fri, 13 Mar 2020 11:55:50 GMT</pubDate>
            <atom:updated>2020-03-13T12:15:41.835Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/677/1*miQ63EyzB0aueNpHGkqT8A.jpeg" /></figure><p>Quantas vezes você já não fez pré-julgamentos sobre alguém, algumas né? A roupa de fulano é assim, o jeito que ele fala, as atitudes dela… são alguns pré-conceitos que fazemos da personalidade de alguém antes mesmo de conversar com sicrano ou beltrano.</p><p>Muitos negócios, amizades e acordos deixaram de ser feitos por conta da dedução. Até mesmo um casamento deixou de acontecer por conta de um dos lados não ter deixado a dedução de fora, e colocado em prática o diálogo.</p><p>Existem perfis e perfis de pessoas, umas mais tímidas outras mais extrovertidas. Algumas pessoas ficam confortáveis em usar a dedução a seu favor, outras preferem partir para o diálogo e de fato, conhecer o caráter da pessoa.</p><p>Antes de fazer julgamentos de alguém, converse, conheça antes, busque um diálogo. Talvez seja uma maneira de construir relações mais sólidas e duradouras.</p><p>Sair da abstração para o concreto, requer coragem de abrir mão da sua zona de conforto e se colocar em posição de busca de novos horizontes e conhecimento.</p><p>Questione sempre, seja inquieto e busque respostas.</p><p>J.F Aguiar</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e96528a623f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>