<?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 Egg on Medium]]></title>
        <description><![CDATA[Stories by Egg on Medium]]></description>
        <link>https://medium.com/@tamaokayasuhiro?source=rss-bfdce180adf4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*owF_9yBP0_dOEenef0KTcQ.jpeg</url>
            <title>Stories by Egg on Medium</title>
            <link>https://medium.com/@tamaokayasuhiro?source=rss-bfdce180adf4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 19:30:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tamaokayasuhiro/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 End of “Scale Is All You Need”: Why NeurIPS 2025 Marks the Beginning of the Physical AI Era]]></title>
            <link>https://medium.com/@tamaokayasuhiro/the-end-of-scale-is-all-you-need-why-neurips-2025-marks-the-beginning-of-the-physical-ai-era-421ac0ecca31?source=rss-bfdce180adf4------2</link>
            <guid isPermaLink="false">https://medium.com/p/421ac0ecca31</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[san-diego]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[neurips]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Egg]]></dc:creator>
            <pubDate>Fri, 02 Jan 2026 12:13:35 GMT</pubDate>
            <atom:updated>2026-01-02T12:13:35.556Z</atom:updated>
            <content:encoded><![CDATA[<h3>A Dispatch from San Diego</h3><p>The air in San Diego carries a distinct weight — a heavy, salt-laden calm that stands in stark contrast to the frantic, electric thinness of Silicon Valley or the gray intellectualism of Boston. Standing outside the San Diego Convention Center in early December 2025, watching the Pacific Ocean glitter under a winter sun, one could almost be forgiven for thinking this was a retreat. But inside the glass walls, amidst a record-breaking swarm of over 15,000 researchers, engineers, and investors gathering for NeurIPS 2025, the atmosphere was anything but calm. It was the friction point of a massive tectonic shift.1</p><p>For the past three years, the global artificial intelligence community has been unified by a single, religiously held belief, a dogma that has driven trillion-dollar valuations and geopolitical anxiety alike: <strong>Scale is all you need.</strong> The recipe was simple, brutal, and effective — throw more data, more parameters, and more H100s at the problem, and intelligence will emerge. It was the era of the chatbot, the Large Language Model (LLM), and the absolute supremacy of the digital bit.</p><p>But standing on that expo floor, navigating the labyrinth of posters that stretched like a city map, it became undeniably clear that the wind has changed. The era of “Scale” as the sole, blunt-force driver of progress is hitting a wall of diminishing returns. We are entering a messier, harder, and infinitely more consequential phase: the era of <strong>Efficiency</strong> and <strong>Physical AI</strong>.</p><p>I attended NeurIPS 2025 not merely to parse the 5,290 accepted papers — anyone can do that from a laptop in Tokyo, London, or Singapore — but to feel the <em>heat</em> of the ecosystem.1 To understand where the puck is going when the “compute-only” strategy becomes economically and physically unsustainable. What I found was a brutal reality check for the global tech hierarchy, a surprising masterclass in ecosystem building from a city often dismissed as a “vacation town,” and a stark, quantified warning for Japan and other non-G2 nations.</p><p>This is not a conference report. This is a manifesto on how the rules of the AI game have just been rewritten, and why the next trillion dollars of value won’t be found in a better chatbot, but in the violent collision between code and the concrete world.</p><h3>1. Why NeurIPS 2025 Felt Different</h3><p>To understand the magnitude of what occurred in San Diego, one must look <em>through</em> the numbers, not just at them. NeurIPS (Neural Information Processing Systems) has always been the bellwether of the industry, the “canary in the coal mine” for shifts that eventually reshape the global economy.</p><p>In 1987, it was a fringe gathering of 500 physicists and neuroscientists discussing connectionism in a ski resort.1 In 2012, it was the specific room where AlexNet sparked the Deep Learning revolution, proving that neural networks could see.1 By 2022, it was the launchpad for the Generative AI explosion that gave us ChatGPT.</p><p>But 2025 represents the moment the graph went vertical. The conference received a staggering <strong>21,575 submissions</strong>, a <strong>61% increase</strong> year-over-year.1 To put that in perspective, the number of people <em>trying</em> to get into NeurIPS today dwarfs the entire population of the AI field just a decade ago.</p><p><strong>Metric201820242025 (San Diego)Growth TrendSubmissions</strong>4,85613,400<strong>21,575+61% YoYAccepted Papers</strong>1,0113,459<strong>5,290</strong>Record High<strong>Acceptance Rate</strong>20.8%25.8%<strong>24.5%</strong>Stable Quality 1</p><p>Yet, despite this explosion, the acceptance rate held steady at roughly 24.5%.1 The quality bar hasn’t lowered; the volume of high-quality science has simply exploded. But the <em>texture</em> of that science has changed.</p><p>In previous years (2022–2024), the energy at NeurIPS was manic — a gold rush mentality where every other poster promised Artificial General Intelligence (AGI) within six months. The air buzzed with the euphoria of “magic.” This year, the mood was different. It was serious. It was industrial. The euphoria of discovery has been replaced by the engineering reality of deployment.</p><p>You could see it in the eyes of the researchers from Google DeepMind, Meta FAIR, and OpenAI. They were no longer just showing off capabilities; they were wrestling with the hard limits of physics, economics, and safety. The conversations in the hallways weren’t about who had the biggest model anymore. They were about:</p><ul><li>“How do we make this run on a device that doesn’t require a nuclear reactor?”</li><li>“How do we stop the model from hallucinating when it controls a robot arm holding a scalpel?”</li><li>“How do we break the dependency on English-only internet data?”</li></ul><p>This shift from <strong>“Look what it can do”</strong> to <strong>“Look how efficiently and safely it can interact with reality”</strong> is the hallmark of a maturing technology. We have moved from the “Scientific Discovery” phase to the “Engineering Refinement” phase. And in that transition, the competitive advantage shifts from those with the deepest pockets for GPUs to those with the deepest understanding of the physical world.</p><h3>2. The End of “Scale Is All You Need”</h3><p>If there was a single intellectual theme dominating the keynotes and oral presentations, it was the quiet, almost reluctant admission that the “Scaling Laws” — the observation that performance improves predictably with more compute and data — are hitting a wall. Or rather, the cost of climbing that wall is becoming prohibitive.</p><p>For the last five years, the industry operated on a brute-force logic: “Scale is All You Need.” If your model wasn’t smart enough, you just made it bigger. You trained it on more text. You bought more NVIDIA H100s. This logic favored the incumbents — the “G2” superpowers (US and China) and the hyperscalers who could afford billion-dollar clusters.</p><h3>The Efficiency Imperative</h3><p>NeurIPS 2025 marked the official counter-reformation. The focus has shifted violently toward <strong>Efficiency</strong>.1</p><p>Why? Because we are running out of internet. We are running out of power. And we are running out of patience for latency. The research presented in San Diego highlighted a new frontier: <strong>Test-Time Compute</strong> (often referred to as “System 2” thinking).2 Instead of just making the model bigger during <em>training</em> (which is astronomically expensive and creates a static artifact), researchers are figuring out how to make models “think” longer and harder during <em>inference</em>.</p><p>This is a profound democratization of the playing field. It means the battleground is moving from “who has the biggest cluster” to “who has the smartest architecture.” It suggests that you don’t necessarily need a $100 billion data center to innovate; you need clever mathematics and engineering — areas where resource-constrained nations like Japan and the UK have historically excelled.</p><h3>The “Data Wall” and the Synthetic Future</h3><p>Another whisper running through the poster sessions was the “Data Wall.” We have effectively trained models on the entire public internet. To go further, we need data that doesn’t exist online. We need <em>physical</em> data. We need <em>embodied</em> data.</p><p>This is why the keynote on “Scale is Not All You Need” resonated so deeply.3 The speaker highlighted that synthetic data (AI generating data for AI) causes “model collapse” if not managed carefully. The snake eats its own tail, and the signal degrades into noise. The only way to get fresh, high-quality signal is to have AI interact with the real world, make mistakes, and learn from physics.</p><p>This realization is driving the second massive trend of NeurIPS 2025: the pivot from Bits to Atoms.</p><h3>3. From Bits to Atoms: The Rise of Physical AI</h3><p>If 2023 was the year of the Chatbot, 2025 is the year the AI grew a body. But these are not the clunky, pre-programmed robots of the 20th century. We are witnessing the birth of <strong>Physical AI</strong>.1</p><p>This term — <em>Physical AI</em> — was everywhere. It represents the convergence of Foundation Models (like GPT-4) with Robotics, creating systems that understand the world not just semantically, but kinetically.</p><h3>The “Brain” Meets the “Body”</h3><p>Until recently, robots were essentially “dumb.” You had to program every movement. If you moved a coffee cup three inches to the left, the robot would grasp at thin air, executing its pre-written code blindly.</p><p>Now, we are seeing <strong>Vision-Language-Action (VLA)</strong> models.5 These are AI brains that can “see” the world through cameras, “understand” a vague natural language command like “clean up that mess,” and “act” by generating the motor controls to do it — even if the mess is in a configuration it has never seen before.</p><p>At the expo, the manifestation of this shift was striking:</p><ul><li><strong>Wayve</strong>, the UK autonomous driving unicorn, was demonstrating end-to-end learning where the car learns to drive like a human.6 It doesn’t follow a set of if/then rules; it understands the <em>context</em> of the road. It negotiates with other drivers.</li><li><strong>Qualcomm</strong> was showcasing edge-integrated VLA models that process visual and voice data in real-time to generate safe driving trajectories, proving that this intelligence can live on the device, not just in the cloud.7</li><li><strong>NVIDIA</strong> released <strong>GR00T</strong>, an open foundation model for humanoid robots, effectively attempting to create the “operating system” for the physical world.5</li></ul><h3>The Reversal of Moravec’s Paradox</h3><p>We are finally seeing the reversal of Moravec’s Paradox — the observation that high-level reasoning (chess, math) requires little computation, while low-level sensorimotor skills (walking, holding an egg) require enormous computational resources. For decades, AI was great at chess but terrible at being a toddler.</p><p>Now, thanks to the efficiency breakthroughs discussed at NeurIPS, AI is finally learning to be a toddler. And once it masters the physical world, the economic implications are vastly larger than those of a chatbot. A chatbot can write a poem; a Physical AI can build a house, care for the elderly, or repair a power grid.</p><p>This shift is the most important signal for Japan. In the world of “Bits” (pure software, search, social media), the US and China have an insurmountable lead. They own the platforms, the cloud, and the consumer data. But in the world of “Atoms” (robotics, medical devices, materials science), the playing field is different. An LLM cannot learn how to solder a circuit board or perform a delicate endoscopy by reading Reddit. It needs physical data. It needs to be embodied in precision hardware.</p><p>This is where the “Japan Advantage” potentially exists. The country that has the highest density of robots 8 and the deepest repository of tacit manufacturing knowledge is suddenly sitting on the world’s most valuable training data — <em>if</em> it can digitize it.</p><h3>4. Why San Diego Matters (and Silicon Valley Isn’t Enough)</h3><p>One of the explicit goals of my trip was to investigate why NeurIPS chose San Diego, and what this city represents as an ecosystem.1 It is easy to dismiss San Diego as a “vacation town” or a “second-tier” tech hub compared to the hegemony of the Bay Area or the academic prestige of Boston.</p><p>Yet, San Diego has quietly built one of the most robust DeepTech ecosystems in the world. It is the perfect case study for how to win <em>without</em> being Silicon Valley. It offers a blueprint for “Second Cities” and “Second Nations” on how to carve out a winning niche.</p><h3>The “30-Minute Ecosystem”</h3><p>In the Bay Area, software is king. In San Diego, <strong>Bio + Hardware</strong> is king.</p><p>I spent days touring the local ecosystem — places like UCSD, the Salk Institute, and the corporate HQs of Qualcomm and Illumina. What struck me was the density. In San Diego, you can drive from a Nobel Prize-winning research lab at UCSD, to a wireless hardware prototyping facility at Qualcomm, to a mass-production factory in Tijuana, Mexico — all within 30 minutes.1</p><p>This <strong>Research $\rightarrow$ Prototype $\rightarrow$ Production</strong> loop is structurally impossible in Silicon Valley (where manufacturing is outsourced to Asia) or Boston (which lacks the maquiladora proximity). In San Diego, the feedback loop between the idea and the physical object is almost instantaneous.</p><h3>The “Cali-Baja” Advantage</h3><p>The secret weapon of San Diego is <strong>Tijuana</strong>. The two cities, separated by a militarized border, ironically function as a single, fluid economic unit often called the “Cali-Baja Mega-Region”.9</p><p>While national politics often focus on walls, the tech economy has effectively erased the border. I learned that nearly <strong>97% of San Diego’s exports</strong> go to Mexico, supporting nearly 100,000 jobs.10 This is the <strong>Physical AI</strong> supply chain in action:</p><ul><li><strong>Design</strong> happens in the mesas of Torrey Pines (San Diego).</li><li><strong>Manufacturing</strong> happens in the maquiladoras of Otay Mesa (Tijuana).</li><li><strong>Medical Devices</strong> are the star: Tijuana is now the medical device manufacturing capital of North America.11 Companies like Medtronic and Thermo Fisher don’t just assemble there; they have established engineering centers of excellence.12</li></ul><p>For a startup building Physical AI (e.g., a smart surgical robot), San Diego offers a speed of iteration that is 10x faster than outsourcing to Shenzhen and waiting for shipping containers. You can design in the morning, cross the border via the Cross Border Xpress (CBX) to check the prototype line at lunch, and be back in La Jolla for a surf session by sunset.</p><h3>EvoNexus: The “Pay It Forward” Anomaly</h3><p>Perhaps the most shocking discovery was <strong>EvoNexus</strong>, San Diego’s premier incubator. In a global startup culture dominated by the Y Combinator model (taking 7% equity for $125k), EvoNexus is an anomaly.</p><p>EvoNexus takes <strong>1% equity</strong> (which is granted to a non-profit) and charges <strong>zero fees</strong>.13</p><p>How is this possible? It is funded by the “Pay It Forward” culture established by the region’s successful exits. Real estate developers (The Irvine Company) and corporate sponsors (Qualcomm, Franklin Templeton) pay the bills because they want the ecosystem to thrive, not because they are hunting for a quick 100x return.14</p><p>This model — collaborative, hardware-centric, and patient — is far more relevant to Japan’s corporate culture than the cutthroat mercantilism of Sand Hill Road. It proves that you can build a world-class ecosystem based on <strong>community longevity</strong> rather than just <strong>exit velocity</strong>.</p><h3>5. The Brutal Reality for Japan: 1 vs 48</h3><p>Despite the optimism of the “Physical AI” shift and the inspiring model of San Diego, the numbers from NeurIPS 2025 deliver a cold splash of water to anyone hoping Japan will easily catch up.</p><p>I call it the <strong>1:48 Problem</strong>.</p><p>When you look at the accepted papers — the hard currency of future innovation — the disparity is terrifying.</p><ul><li><strong>USA + China (G2):</strong> ~4,820 papers.1</li><li><strong>Japan:</strong> ~100 papers.1</li></ul><p>That is a ratio of roughly <strong>1 to 48</strong>.</p><h3>The G2 Duopoly is Complete</h3><p>The gap between the US and China has narrowed to a statistical margin of error (just 3.3%).1 China is no longer “catching up.” They have arrived. They are peers.</p><p>A specific signal of this parity was the <strong>Best Paper Award</strong> given to the Alibaba Qwen team.15 This was not a paper about “copying” Western tech; it was a fundamental breakthrough in attention mechanisms for Large Language Models. China is innovating on the architectural level.</p><p>Japan, meanwhile, sits at roughly 2% of the global output. Even with the heroic efforts of RIKEN AIP (46 papers) and companies like Preferred Networks (PFN) and Sony, the sheer volume of intellectual capital being deployed by the G2 is overwhelming.1</p><p>If AI were a war of attrition, Japan would have already lost. You cannot fight a 1:48 force in a head-on collision. The “Scale” game belongs to the G2. They have the data centers, the energy grids, and the raw capital.</p><h3>The “Wait and See” Death Spiral</h3><p>The most damaging cultural trait I observed was not a lack of talent — the Japanese researchers I met were brilliant — but a lack of <em>presence</em> and <em>agency</em>.</p><p>In the US ecosystem, the culture is: <strong>“Don’t wait. Go get it.”</strong> At the conference, I watched PhD students boldly approach Chief Scientists from OpenAI or DeepMind to pitch their ideas. They didn’t wait for an introduction. They didn’t apologize for interrupting. They understood that in a hyper-growth industry, passivity is invisibility.</p><p>In contrast, the “Wait and See” (Machi no Bigaku) approach often favored by Japanese institutions is a death sentence in this environment.1 The speed of AI evolution is measured in weeks, not years. By the time you “wait to see if the technology matures,” the G2 players have already standardized the platform, hired the talent, and locked in the IP. The “1:48” ratio is not just a measure of output; it is a measure of <em>attempts</em>.</p><h3>6. Japan’s Real Chance: Changing the Ring</h3><p>So, is it game over? Is Japan destined to be a digital colony of the G2?</p><p>No.</p><p>The <strong>1:48</strong> ratio matters only if you try to fight the G2 on their own turf (Generative LLMs, Consumer Internet, Search). If Japan tries to build the next OpenAI, it will likely fail. The capital requirements and data scale are simply too high, and the lead is too long.</p><p>But remember the vibe shift at NeurIPS 2025: <strong>Scale is ending. Physical AI is beginning.</strong></p><p>Japan’s strategy must be to <strong>change the ring</strong>. Don’t fight in the “Bits” arena. Fight in the “Atoms” arena.</p><h3>The Winning Formula: Japan’s “Hidden” Assets</h3><p>The “San Diego Strategy” — winning by dominating a niche intersection of Bio/Hardware — is the blueprint for Japan. Japan possesses three critical assets that the G2 struggle to replicate, assets that are suddenly increasing in value in the Physical AI era:</p><ol><li><strong>High-Context Physical Data:</strong> Japan has the world’s highest density of industrial robots (631 per 10,000 employees).1 This means Japan has the best <em>training data</em> for factory automation, elderly care robotics, and precision control. The US has data on <em>text</em> (Reddit/Twitter); Japan has data on <em>movement</em> (Factory Floors/Hospitals). In an era of Embodied AI, movement data is the new oil.</li><li><strong>Trust &amp; Safety in Regulated Markets:</strong> Just as San Diego dominates medical devices because of its rigor, Japan’s reputation for quality control is a massive asset in “Safety-Critical AI.” You can hallucinate a haiku, but you cannot hallucinate a brake line or a surgical incision. The world trusts Japanese hardware to <em>not kill people</em>. That trust is a moat.</li><li><strong>Materials Science (The “Chef’s Secret”):</strong> AI is increasingly being used to discover new materials (batteries, semiconductors). Japan still holds a 50% global share in many critical semiconductor materials.1 Applying AI to this “black box” chemical knowledge — what researchers call Materials Informatics (MI) — is a massive opportunity.</li></ol><h3>The “Third Option” Strategy</h3><p>San Diego didn’t try to become Silicon Valley. It became the “Place where Hardware meets Biology.” Japan must become the “Place where AI meets Reality.”</p><p>This means:</p><ul><li><strong>Doubling down on Robotics/Physical AI:</strong> Preferred Networks (PFN) is already doing this.1 We need 10 more PFNs.</li><li><strong>Importing the “San Diego Loop”:</strong> We need to create zones where R&amp;D (Tokyo/Kyoto) is tightly coupled with rapid prototyping and mass production (Ota City/Osaka/Fukuoka), mimicking the Cali-Baja speed.</li><li><strong>Abandoning the “All-Japan” Obsession:</strong> San Diego relies on Mexico. Japan must rely on global talent. We cannot win with a closed ecosystem.</li></ul><h3>7. What We Should Build Next (Call to Thought)</h3><p>Walking out of the San Diego Convention Center on the final day, the Pacific Ocean was glittering, immense and indifferent. The contrast between the dark, screen-lit expo halls and the vast, bright natural world was striking.</p><p>It felt like a metaphor for the state of AI. We have spent the last decade trapped in the dark room of “Bits,” optimizing matrix multiplications and predicting the next token. We have built incredible digital minds that live in a void.</p><p>Now, we are opening the door. We are letting those minds out into the sunlight. We are giving them eyes, hands, and bodies. We are asking them to drive our cars, cure our diseases, and build our cities.</p><p>For the VC, the Founder, and the Policymaker reading this, my question to you is simple:</p><p><strong>Are you still betting on the scaling curve of yesterday? Or are you building the physical infrastructure of tomorrow?</strong></p><p>The “Scale” war is over. The G2 won it.</p><p>The “Physical” war has just begun. And the winner of that war hasn’t been decided yet.</p><p>It’s time to stop looking at the scoreboard (1:48) and start changing the game.</p><h3>Extended Analysis: The “Pay It Forward” Mechanics of EvoNexus</h3><p><em>Detailed Context for Ecosystem Builders</em></p><p>To understand why EvoNexus works — and why it serves as such a critical model for non-Silicon Valley ecosystems — we must look at its mechanics. EvoNexus was founded in 2010 by Rory Moore (co-founder of Peregrine Semiconductor) and Vice Admiral Walter Davis.16 It emerged not from a desire to profit from startups, but from a desire to save the San Diego tech ecosystem during a downturn.</p><p>The model is structurally unique:</p><ul><li><strong>Equity:</strong> 1% common stock (fully diluted) granted to the non-profit entity.17</li><li><strong>Rent:</strong> Up to 2 years of rent-free residency in Class A office space (provided by The Irvine Company).17</li><li><strong>Mentorship:</strong> Access to a network of mentors who are there to help, not to angel invest.</li></ul><p>This creates a different incentive structure. In a traditional VC-backed accelerator, the goal is to pump up the valuation for the next round. At EvoNexus, the goal is <strong>survivability and sustainability</strong>. The result? Since 2010, EvoNexus companies have raised over $5.1 billion in funding and outcomes 18, with an 85% success rate in raising venture funding.18</p><p>This “Community-Supported” model is a powerful alternative for Japan. Instead of relying solely on VC risk capital (which is scarce in Japan compared to the US), an ecosystem could be built on the backs of corporate giants (Toyota, Sony, Mitsubishi) providing “in-kind” support (office space, lab equipment, manufacturing lines) in exchange for a healthier innovation ecosystem, rather than immediate financial returns. This aligns perfectly with the stakeholder capitalism model prevalent in Japan.</p><h3>Deep Dive: The “Cali-Baja” Medical Device Cluster</h3><p><em>A Case Study in Cross-Border Synergy</em></p><p>The “Cali-Baja” region is not just a marketing term; it is the <strong>world’s largest medical device cluster</strong>.19 This dominance was not accidental. It was engineered through the North American Free Trade Agreement (NAFTA) and cemented by the USMCA.10</p><ul><li><strong>The Logic:</strong> High-value R&amp;D and engineering stays in San Diego (UCSD, Scripps, Salk). High-precision, labor-intensive manufacturing moves to Tijuana.</li><li><strong>The Scale:</strong> 97% of San Diego’s goods exports go to Mexico.10</li><li><strong>The Players:</strong> Thermo Fisher Scientific, Medtronic, and ResMed all operate massive facilities in Tijuana.20</li></ul><p>Crucially, Tijuana has evolved. It is no longer just cheap assembly. It is moving up the value chain. Organizations like <strong>ITJ</strong> (founded by Maritza Diaz) are helping US companies set up software engineering centers of excellence in Tijuana, tapping into the local talent pool of engineers who are culturally aligned with the US but cost significantly less than Silicon Valley talent.21</p><p>For Japan, the lesson is clear: Where is your Tijuana? Is it Vietnam? Is it the Philippines? Or is it a revitalization of regional Japan through high-tech manufacturing zones? The “closed loop” of domestic manufacturing is powerful, but the “cost-arbitrage loop” of cross-border manufacturing (like Cali-Baja) provides the economic margin needed to compete with China.</p><h3>References (Integrated Sources)</h3><ul><li>1: NeurIPS 2025 Trip Report (General Statistics, 1:48 Ratio, San Diego Context, Acceptance Rates).</li><li>1: Historical context of NeurIPS (1987–2025).</li><li>1: Statistical growth of NeurIPS submissions.</li><li>13: EvoNexus equity model details.</li><li>14: EvoNexus “Pay It Forward” culture and corporate sponsorship.</li><li>4: World Economic Forum report on Physical AI trends.</li><li>5: NVIDIA GR00T and robotics foundation models.</li><li>15: Alibaba Qwen Best Paper Award details.</li><li>6: Wayve’s booth and embodied AI presence.</li><li>2: Radical Ventures on the shift to “System 2” reasoning and efficiency.</li><li>9: Cali-Baja Mega-Region economic integration.</li><li>8: Global robot density statistics.</li><li>13: EvoNexus 1% equity structure.</li><li>10: World Trade Center San Diego report on cross-border trade stats.</li><li>17: EvoNexus application requirements and benefits.</li><li>3: NeurIPS Keynote “Scale is Not All You Need”.</li><li>18: EvoNexus cumulative funding stats ($5.1B).</li><li>16: History of EvoNexus founding.</li><li>20: Specifics on medical device companies in Tijuana (ResMed, Medtronic).</li><li>12: ITJ and the evolution of Tijuana’s engineering talent.</li><li>19: Cali-Baja as the world’s largest medical device cluster.</li><li>11: Cali-Baja nearshoring report.</li><li>7: Qualcomm’s edge AI and VLA demonstrations.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=421ac0ecca31" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>