<?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 lwi00 on Medium]]></title>
        <description><![CDATA[Stories by lwi00 on Medium]]></description>
        <link>https://medium.com/@doidylouis?source=rss-7943fec93a50------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*3jYkdX4DqFU7Z14U0nLHcQ.jpeg</url>
            <title>Stories by lwi00 on Medium</title>
            <link>https://medium.com/@doidylouis?source=rss-7943fec93a50------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 19:18:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@doidylouis/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[Is NVIDIA really a hardware company ?]]></title>
            <link>https://medium.com/@doidylouis/is-nvidia-really-a-hardware-company-b4b3df5683b4?source=rss-7943fec93a50------2</link>
            <guid isPermaLink="false">https://medium.com/p/b4b3df5683b4</guid>
            <dc:creator><![CDATA[lwi00]]></dc:creator>
            <pubDate>Wed, 07 Jan 2026 10:47:53 GMT</pubDate>
            <atom:updated>2026-01-07T18:53:42.996Z</atom:updated>
            <content:encoded><![CDATA[<h3>Is NVIDIA really a hardware company ?</h3><blockquote>The importance of drivers and firmware in the GPU market</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/714/0*7uCD0-ftFrbtMhqX.png" /></figure><p>In the annals of modern technological history, few companies have undergone a metamorphosis as profound and lucrative as NVIDIA Corporation. By late 2025, the company had not only surpassed a <strong>$4 trillion market capitalization</strong> but had fundamentally redefined the economic logic of the semiconductor industry. Traditional hardware manufacturing is a business characterized by cyclical demand, ruthless commoditization, and gross margins that struggle to exceed 50% even in the best of times. Yet, NVIDIA consistently reports gross margins in excess of 75%, a financial profile typically reserved for elite SaaS monopolies rather than silicon foundries. This discrepancy signals a fundamental misunderstanding of NVIDIA’s true nature by the broader market: NVIDIA is not a hardware company that writes software to support its chips; it is a proprietary software platform company that utilizes high-performance silicon as the exclusive delivery mechanism for its ecosystem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Zr2PFuvi_GLUYqLb.jpg" /><figcaption>NVIDIA earning decompositions</figcaption></figure><p>The central thesis of this report is that NVIDIA main moat is not the transistor count of its Blackwell or Rubin architectures, but in a meticulously constructed “moat” of drivers, firmware, and API libraries. This ecosystem, centered on the <strong>Compute Unified Device Architecture</strong> (CUDA), creates a profound vendor lock-in that renders the raw hardware performance specifications of competitors like AMD and Intel largely irrelevant for the vast majority of commercial applications.</p><p>This report offers an exhaustive analysis of the technical, legal, and economic mechanisms NVIDIA employs to maintain this dominance. It explores the “<strong>CUDA Gap</strong>,” the controversial role of signed firmware in suppressing open-source driver development, and the existential threat posed by the “TorchTPU” alliance between Google and Meta, which seeks to dismantle the software barriers that protect NVIDIA’s empire.</p><h3>Section 1: The Ontological Crisis of Silicon : Hardware vs. Platform</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/498/0*ee_KiIo9Ke8A2J4P.png" /></figure><h4>The Definition of the Ecosystem</h4><p>To understand the current valuation of NVIDIA, one must first deconstruct the traditional dichotomy between hardware and software. For decades, the semiconductor industry operated on the principle of “speeds and feeds” : the idea that the primary value driver was the raw performance metrics of the hardware, such as clock speed, memory bandwidth, and transistor density (not important definitions here). In this paradigm, software was merely an enabler, a necessary utility to unlock the hardware’s potential.</p><p>However, NVIDIA’s trajectory over the last two decades has inverted this relationship. Investors and analysts increasingly draw parallels between NVIDIA and Apple, <a href="https://www.aminext.blog/en/post/why-nvidia-s-true-moat-isn-t-chips-but-cuda-an-investor-s-guide-to-the-ecosystem-wars#google_vignette">noting that buying NVIDIA</a> stock is less akin to purchasing a component manufacturer and more like investing in a closed ecosystem. When a consumer buys an iPhone, they are buying access to the App Store; <strong>when an enterprise purchases an H100 GPU, they are buying access to CUDA.</strong> The hardware, while technologically marvels in their own right, GPUs act as a physical key required to access the vast, accumulated value of the CUDA software library.</p><p>The “moat” protecting NVIDIA is quantified not by the number of developers alone, but by the accumulated volume of optimized code and the switching costs embedded in the industry’s workflow. An AI application today is rarely written using the base CUDA API alone; it leverages a rich stack of highly optimized, proprietary libraries like cuDNN for deep neural network primitives and TensorRT for inference optimization. These libraries are the result of thousands of engineer-years of optimization, creating a barrier to entry that raw silicon performance cannot overcome. A competitor might build a taller castle (faster chip), but without access to the water source (the software ecosystem and in fine the final usage for end users), the castle remains uninhabitable.</p><h3>Section 2: Call the control tower : The CUDA Ecosystem</h3><h4>2.1 The Two Faces of CUDA: Driver vs. Runtime</h4><p>To technical observers, CUDA is often viewed as a monolith. However, its architecture is bifurcated into two distinct levels, a distinction that is critical for understanding the “lock-in” mechanism utilized by NVIDIA.</p><p><strong>Table 1: The Structure of the CUDA API Architecture</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/701/1*Jmm3QSNsgFyAGfknGJa2YQ.png" /><figcaption>Source Data Analysis: 12</figcaption></figure><p>The CUDA Runtime API eases device code management by providing implicit initialization and context management. This leads to simpler code, which accelerates adoption among researchers and developers who are not systems engineers. However, this convenience comes at the cost of control and portability. The Runtime API is typically statically linked or version-locked to the developer’s application. In contrast, the Driver API is dynamically linked to the installed NVIDIA driver.</p><p>This distinction creates a subtle but powerful lock-in. Applications built on the Runtime API are tied to the specific version of the CUDA Toolkit used at compile time. Because the Runtime API depends on the Driver API, and the Driver API is tied to the GPU driver version, developers and users are forced into a perpetual upgrade cycle. To run a newer application (compiled with a newer Toolkit), the user <em>must</em> update their NVIDIA driver. This mechanism ensures that the installed base of NVIDIA drivers is constantly refreshed, preventing fragmentation but also ensuring that older hardware or third-party drivers are quickly rendered obsolete.</p><h4>2.2 The “CUDA Gap”: Software Outperforming Physics</h4><p>A quantitative analysis of the market reveals a phenomenon that can be termed the “CUDA Gap.” This metric quantifies the extent to which NVIDIA’s software optimization improves its hardware’s realized performance beyond what raw hardware specifications (TFLOPS, or TeraFlops, indicating number of operations per seconds) would predict. In a purely hardware-centric view, a competitor’s chip with higher floating-point throughput should outperform an NVIDIA chip. In reality, the opposite is often true.</p><p><strong>Table 2: The CUDA Gap: Theoretical vs. Realized Throughput</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/705/1*ewGC7HWscPyE6JLm8m_oVQ.png" /><figcaption>Source: AIMultiple Research ”GPU Software for AI: CUDA vs. ROCm in 2026&quot;</figcaption></figure><p>The data in Table 2 illustrates the paradox. In the comparison between the NVIDIA H100 and the AMD MI300X, the AMD chip possesses a theoretical TFLOPS advantage of roughly 32%. Yet, in real-world high-concurrency scenarios (128 concurrent users), the NVIDIA H100 delivers nearly 39% <em>more</em> throughput than the AMD chip. This discrepancy results in a score of roughly 70, indicating that NVIDIA’s software stack unlocks performance that is 70% higher than what the hardware specs alone would suggest relative to the competition.</p><p>Why does this gap exist? The answer lies in the maturity of the software stack.</p><ol><li><strong>Library Optimization:</strong> NVIDIA’s libraries (cuBLAS, cuDNN) have been profiled and tuned for nearly two decades against every micro-architecture iteration. They utilize specific instruction sets (like Tensor Cores) with optimal memory access patterns that are automatically engaged without the developer needing to intervene.</li><li><strong>Compiler Sophistication:</strong> The NVIDIA compiler (nvcc) performs sophisticated automatic kernel fusion. It identifies separate operations in the code (e.g., a multiplication followed by an addition) and fuses them into a single GPU kernel launch, reducing the overhead of memory transfers and kernel scheduling. AMD’s compiler stack, while improving, lags in these aggressive optimizations.</li><li><strong>Framework Integration:</strong> Deep learning frameworks like PyTorch and TensorFlow have CUDA integration built into their core. Custom CUDA kernels are written for everyday operations, ensuring that the “happy path” for execution is always on NVIDIA hardware.</li></ol><p>This gap effectively means that when an enterprise purchases a competitor’s chip based on superior datasheet specifications, they are paying for computational capacity that their software cannot effectively utilize. The cost-per-token of the “expensive” NVIDIA chip becomes lower than the “cheap” competitor chip once the software efficiency is factored in.</p><h3>Section 3: The Firmware Fortress : The GSP and the Falcon Processor</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*UirePObld7TW7W1d.jpg" /></figure><h4>3.1 The Shift from Host Driver to Firmware</h4><p>While CUDA dominates the high-level application space, a quieter but equally significant transformation has occurred at the low level: the migration of driver logic from the host CPU to the GPU’s internal firmware. Historically, the graphics driver (a kernel module running on the host OS) managed the GPU’s state. However, starting with the Turing architecture and accelerating with Ampere and Hopper, NVIDIA has introduced the <a href="https://download.nvidia.com/XFree86/Linux-x86_64/510.39.01/README/gsp.html"><strong>GPU System Processor (GSP)</strong>.</a></p><p>The GSP is a RISC-V based coprocessor embedded directly into the GPU silicon. NVIDIA has been progressively moving logic that previously resided in the open-source-accessible kernel driver into the proprietary GSP firmware blob. This includes critical functions such as temperature management or overclocking.</p><p>NVIDIA markets this transition as a performance optimization, arguing that offloading these tasks from the CPU reduces latency and improves system responsiveness. While technically true, this architectural shift serves a dual purpose as a <strong>potent anti-competitive mechanism</strong>. By effectively “black-boxing” the GPU’s operation, NVIDIA ensures that the hardware cannot be controlled without its proprietary, cryptographically signed firmware.</p><h4>3.2 The “Falcon” Security Processor and Signed Firmware</h4><p>The enforcement mechanism for this control is the <strong>“Falcon” (FAst Logic CONtroller)</strong> security microprocessor. The Falcon chip enforces a strict “signed firmware” requirement. To fully initialize a modern NVIDIA GPU, the driver must upload a firmware image that is cryptographically signed by NVIDIA keys. Without this signature, the Falcon processor locks down the hardware, refusing to enable critical features such as high-performance clock states or access to specific memory regions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cjxFuGUMxKqBAuTu.jpg" /></figure><p>This architecture has had a devastating impact on the open-source community’s ability to create independent, third-party drivers. The “Nouveau” project, a reverse-engineered open-source driver for NVIDIA cards, has been severely hampered by this requirement. Because the open-source developers cannot sign their own firmware, they are forced to rely on NVIDIA to release “redistributable” binary blobs.</p><p>If NVIDIA delays the release of these blobs, or releases them with restricted functionality, the open-source driver remains non-functional or severely crippled. For example, for several generations, the Nouveau driver was unable to “re-clock” the GPU, meaning the card was stuck running at low boot-time frequencies (e.g., 400 MHz instead of 2000 MHz), rendering expensive hardware useless for any serious workload.20 This ensures that for any commercial or high-performance application, the user <em>must</em> use the proprietary NVIDIA driver stack.</p><h4>3.3 The “Stable ABI” Problem and Hardware Obfuscation</h4><p>The migration of logic to the GSP firmware has created what is known in the <a href="https://blogs.gnome.org/uraeus/2022/05/11/why-is-the-open-source-driver-release-from-nvidia-so-important-for-linux/">Linux kernel community as the “unstable ABI” (Application Binary Interface) problem.</a> The interface between the kernel driver (which runs on the CPU) and the GSP firmware (which runs on the GPU) is not a fixed standard; it changes with almost every driver version.</p><p>NVIDIA developers have admitted that they do not maintain a stable ABI for the GSP message protocols because development is moving too rapidly. This means that the firmware is tightly coupled to a specific version of the kernel driver. An open-source driver cannot simply write a “standard” interface to talk to the hardware; it must chase a moving target, reverse-engineering the communication protocol with every new release.</p><p>The release of the “Open GPU Kernel Modules” in May 2022 was hailed by some as a step forward, but critics argue it is a “Trojan Horse”. While the kernel module source code is now open (GPL/MIT licensed), it is merely a thin wrapper that passes commands to the massive, closed-source GSP firmware blob. The actual “intelligence” of the driver has retreated deeper into the silicon, where open-source licenses cannot reach. This strategy allows NVIDIA to comply with the letter of open-source requirements (for Linux kernel integration) while violating their spirit, ensuring that they retain absolute control over how the hardware is utilized and preventing competitors or third parties from optimizing the hardware in ways NVIDIA does not approve.</p><h3>Section 4: The Invisible Chains of cuDNN and TensorRT</h3><h4>4.1 The Sticky Layers</h4><p>While CUDA provides the language, the true lock-in is maintained by the optimized libraries that sit on top of it. A developer utilizing PyTorch or TensorFlow is rarely writing raw CUDA C++ kernels. Instead, they are calling high-level functions that dispatch operations to NVIDIA’s proprietary libraries: <strong>cuDNN</strong> and <strong>TensorRT</strong>.</p><p><strong>cuDNN (CUDA Deep Neural Network library)</strong> is the absolute bedrock of modern AI training. It contains highly tuned implementations of standard routines such as forward and backward convolution, pooling, normalization, and activation layers. Importantly, cuDNN is <strong>closed-source and proprietary</strong>. It is specifically optimized for NVIDIA’s hardware features, such as Tensor Cores.</p><p>This creates a significant barrier to portability. If a researcher’s code relies on the numerical behavior or performance characteristics of cuDNN (which is the <strong><em>default backend for PyTorch</em></strong>), porting that code to a competitor’s hardware requires finding an exact equivalent. AMD offers <strong>MIOpen</strong>, but it is not a 1:1 drop-in replacement. Differences in precision handling, memory layout, or API surface area create <strong>implementation friction</strong>. A model that converges in 3 days on cuDNN might diverge or take 5 days on MIOpen due to subtle differences in floating-point operations.</p><p><strong>TensorRT</strong> compounds this lock-in at the deployment (inference) stage. TensorRT is a compiler and runtime that takes a trained network and optimizes it for a specific target NVIDIA GPU. It performs graph fusion, kernel auto-tuning, and precision calibration (e.g., quantizing to FP8 or INT8). Once a model passes through TensorRT, it is effectively compiled into a proprietary binary format that <em>only</em> runs on NVIDIA hardware. The model itself becomes tied to the silicon vendor, making it impossible to migrate a production inference pipeline to another hardware provider without rebuilding the entire deployment stack.</p><h4>4.2 NCCL: The Glue of the AI Factory</h4><p>In the era of trillion-parameter models, the unit of compute is no longer the single GPU, but the cluster. Training a model like GPT-4 requires thousands of GPUs communicating in unison. The software that orchestrates this symphony is <strong>NCCL (NVIDIA Collective Communication Library)</strong>.</p><p>NCCL provides inter-GPU communication primitives (AllReduce, Broadcast, ReduceScatter) that are optimized for NVIDIA’s proprietary interconnects, <strong>NVLink</strong> and <strong>InfiniBand</strong>. It is topology-aware, meaning it knows exactly how the chips are wired together and routes data to minimize latency.</p><p>The dominance of NCCL is a critical component of the moat. Even if a competitor like AMD produces a chip with faster compute (e.g., the MI300X), they struggle to match the efficiency of NCCL in multi-node scaling. AMD’s equivalent, <strong>RCCL</strong>, functions similarly, but NCCL is the default backend for the industry-standard torch.distributed package. NVIDIA’s strategy of &quot;full-stack co-design&quot;—selling the GPU, the NVLink switch, the Quantum InfiniBand cable, and the NCCL software—creates a scenario where the value is in the <em>system</em>, not the component. To replace NVIDIA, a customer must replace the entire data center architecture, a proposition that is financially and technically daunting.</p><h3>Section 5: Why AMD ROCm Still Fails ?</h3><h4>5.1 The ROCm Ecosystem and Usability Gaps</h4><p>AMD’s primary response to CUDA is <strong>ROCm</strong> (Radeon Open Compute). On paper, ROCm offers an open-source, flexible alternative, using the HIP language to facilitate porting. However, in practice, ROCm suffers from a chronic “maturity gap” that has led to a reputation for poor usability.</p><p>A <a href="https://www.reddit.com/r/linux/comments/1duq5nx/despite_nvidia_having_a_bad_reputation_with/">survey </a>of developer sentiment in 2024 and 2025 reveals a consistent pattern of frustration. User reports frequently use terms like “nightmare,” “gave up,” and “broken” to describe the experience of setting up a ROCm environment.</p><h4>5.2 The Cost of “Free” Software</h4><p>While ROCm is free and open-source, the <em>cost</em> of using it is measured in engineering time. For an enterprise, the salary cost of engineers spending weeks debugging driver conflicts or porting libraries often outweighs the capital savings of buying cheaper AMD hardware. This economic reality reinforces NVIDIA’s dominance: the “expensive” H100 is actually the “cheap” option because it works out of the box, while the “cheap” MI300X is expensive because it requires a team of systems engineers to maintain.</p><p>The performance delta further complicates the equation. Even when ROCm is successfully deployed, benchmarks indicate that it often trails CUDA in realized performance by 10–30% on equivalent tasks due to less mature compiler optimizations. This reinforces the perception that NVIDIA is the premium, reliable choice.</p><h3>Section 6: Google, Meta, and the Hyperscalers’ rebellion</h3><h4>6.1 The JAX vs. PyTorch Friction</h4><p>The only entities with the resources to effectively challenge NVIDIA’s software monopoly are the “Hyperscalers”, specifically Google and Meta (since Amazon uses Trainium only internaly and not for public released models). These companies are NVIDIA’s largest customers, giving them both the motive (<em>reducing trillion dollar CAPEX</em>) and the means (massive engineering talent pools) to break the lock-in.</p><p>Google has long possessed a superior hardware alternative: the <strong>Tensor Processing Unit (TPU)</strong>. TPUs are ASICs designed specifically for matrix math, offering significantly better performance-per-watt and performance-per-dollar than GPUs.However, TPU adoption has been hindered by software friction. TPUs were historically optimized for Google’s own frameworks, <strong>TensorFlow</strong> and <strong>JAX</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*f61CYGpkX9OnJV2u.jpg" /><figcaption>Sexy TPU isn’t it</figcaption></figure><p>The broader AI research community, however, has standardized on <strong>PyTorch</strong> (created by Meta), which is deeply intertwined with CUDA. This created a dilemma: to use the efficient TPUs, a company had to rewrite its PyTorch codebase into JAX. This “rewrite penalty” was high enough that most companies, including internal teams at non-Google tech firms, chose to pay the “NVIDIA tax” rather than incur the engineering cost of migration.</p><h4>6.2 The “TorchTPU” Initiative (December 2025)</h4><p>In a landmark move in December 2025, Google and Meta formalized a partnership codenamed <strong>“TorchTPU”</strong>.2 This strategic alliance aims to make PyTorch run <em>natively</em> on TPUs, effectively bypassing the need for JAX or CUDA.</p><p>The technical goal of TorchTPU is to develop a direct backend for PyTorch that maps operations to the TPU’s XLA (Accelerated Linear Algebra) compiler without the “lazy tensor” overhead that plagued previous attempts like torch_xla.</p><p>This partnership poses an existential threat to NVIDIA. If PyTorch code runs seamlessly on TPUs without modification, the “switching cost” drops to near zero. Meta, which plans to spend $70–72 billion on CAPEX in 2025, is using this initiative to leverage pricing pressure on NVIDIA and diversify its infrastructure.10 The “Ironwood” TPU superclusters, capable of scaling to 9,000+ chips, represent the first genuine threat to NVIDIA’s data center dominance because they attack the software barrier directly, rather than just competing on hardware specs.</p><h3>So ? Is NVIDIA really a hardware company? The evidence overwhelmingly suggests the answer is <strong>no</strong>.</h3><p>While NVIDIA manufactures physical products, its business model, margin structure, and strategic defenses are those of a proprietary software platform.</p><p>However, the “TorchTPU” initiative marks a critical turning point. For the first time, the software barrier is being attacked by software giants (Google and Meta) rather than hardware competitors., and it might be the first time in decades someone is encorching on Nvidia’s turf, and it might be a good thing for the industry.</p><p>Sources :</p><ul><li>24/7 Wall St. (2026, January 2). <em>Nvidia (NVDA) bull, base, bear price prediction and forecast</em>. <a href="https://247wallst.com/forecasts/2026/01/02/nvidia-nvda-bull-base-bear-price-prediction-and-forecast/">https://247wallst.com/forecasts/2026/01/02/nvidia-nvda-bull-base-bear-price-prediction-and-forecast/</a></li><li>AI Multiple. (n.d.). <em>CUDA vs ROCm: Performance and feature comparison</em>. Research. <a href="https://research.aimultiple.com/cuda-vs-rocm/">https://research.aimultiple.com/cuda-vs-rocm/</a></li><li>Aminext. (n.d.). <em>Why Nvidia’s true moat isn’t chips but CUDA: An investor’s guide to the ecosystem wars</em>. <a href="https://www.aminext.blog/en/post/why-nvidia-s-true-moat-isn-t-chips-but-cuda-an-investor-s-guide-to-the-ecosystem-wars">https://www.aminext.blog/en/post/why-nvidia-s-true-moat-isn-t-chips-but-cuda-an-investor-s-guide-to-the-ecosystem-wars</a></li><li>ByteIota. (2025, December). <em>Google TorchTPU + Meta challenge Nvidia’s $5T CUDA lock</em>. <a href="https://byteiota.com/google-torchtpu-meta-challenge-nvidias-5t-cuda-lock/">https://byteiota.com/google-torchtpu-meta-challenge-nvidias-5t-cuda-lock/</a></li><li>CloudSyntrix. (n.d.). <em>NVIDIA’s AI dominance: How full-stack thinking built an unassailable moat</em>. <a href="https://www.cloudsyntrix.com/blogs/nvidias-ai-dominance-how-full-stack-thinking-built-an-unassailable-moat/">https://www.cloudsyntrix.com/blogs/nvidias-ai-dominance-how-full-stack-thinking-built-an-unassailable-moat/</a></li><li>Cygnus. (2025, December 18). <em>Google, Meta team up on “TorchTPU” as Nvidia faces $5 trillion market test</em>. Domain-b. <a href="https://www.domain-b.com/technology/artificial-intelligence/google-meta-team-up-on-torchtpu-as-nvidia-faces-5-trillion-market-test">https://www.domain-b.com/technology/artificial-intelligence/google-meta-team-up-on-torchtpu-as-nvidia-faces-5-trillion-market-test</a></li><li>Digitimes. (2025, December 18). <em>Google reportedly launches TorchTPU project to boost TPU compatibility with PyTorch</em>. <a href="https://www.digitimes.com/news/a20251218PD222/google-tpu-nvidia-cuda-software.html">https://www.digitimes.com/news/a20251218PD222/google-tpu-nvidia-cuda-software.html</a></li><li>Fedora Discussion. (2025, August). <em>Heads-up: Nvidia proprietary driver with open kernel and GSP firmware breaks Runtime D3 on Turing cards</em>. <a href="https://discussion.fedoraproject.org/t/heads-up-nvidia-proprietary-driver-with-open-kernel-and-gsp-firmware-breaks-runtime-d3-on-turing-cards/161643">https://discussion.fedoraproject.org/t/heads-up-nvidia-proprietary-driver-with-open-kernel-and-gsp-firmware-breaks-runtime-d3-on-turing-cards/161643</a></li><li>Flox. (n.d.). <em>The Flox catalog now contains NVIDIA CUDA</em>. <a href="https://flox.dev/blog/the-flox-catalog-now-contains-nvidia-cuda/">https://flox.dev/blog/the-flox-catalog-now-contains-nvidia-cuda/</a></li><li>Gigazine. (2024, March 5). <em>NVIDIA bans CUDA from running on other hardware</em>. <a href="https://gigazine.net/gsc_news/en/20240305-nvidia-cuda/">https://gigazine.net/gsc_news/en/20240305-nvidia-cuda/</a></li><li>Gnome Blogs. (2022, May 11). <em>Why is the open source driver release from NVIDIA so important for Linux?</em> <a href="https://blogs.gnome.org/uraeus/2022/05/11/why-is-the-open-source-driver-release-from-nvidia-so-important-for-linux/">https://blogs.gnome.org/uraeus/2022/05/11/why-is-the-open-source-driver-release-from-nvidia-so-important-for-linux/</a></li><li>Google Cloud. (n.d.). <em>Guide to JAX for PyTorch developers</em>. <a href="https://cloud.google.com/blog/products/ai-machine-learning/guide-to-jax-for-pytorch-developers">https://cloud.google.com/blog/products/ai-machine-learning/guide-to-jax-for-pytorch-developers</a></li><li>Google Cloud. (n.d.). <em>PyTorch/XLA: Performance debugging on Cloud TPU VM: Part I</em>. <a href="https://cloud.google.com/blog/topics/developers-practitioners/pytorchxla-performance-debugging-tpu-vm-part-1">https://cloud.google.com/blog/topics/developers-practitioners/pytorchxla-performance-debugging-tpu-vm-part-1</a></li><li>How AI Works. (2025). <em>TPU vs GPU vs ASIC: AI hardware market 2025</em>. <a href="https://howaiworks.ai/blog/tpu-gpu-asic-ai-hardware-market-2025">https://howaiworks.ai/blog/tpu-gpu-asic-ai-hardware-market-2025</a></li><li>Hyperframe Research. (2025, December 24). <em>Can Google’s TorchTPU eventually bridge NVIDIA’s CUDA moat?</em> <a href="https://hyperframeresearch.com/2025/12/24/can-googles-torchtpu-eventually-bridge-nvidias-cuda-moat/">https://hyperframeresearch.com/2025/12/24/can-googles-torchtpu-eventually-bridge-nvidias-cuda-moat/</a></li><li>Kanjinghat, N. (n.d.). <em>Deep learning in practice: A technical comparison of PyTorch and JAX</em>. Medium. <a href="https://medium.com/@nijesh-kanjinghat/deep-learning-in-practice-a-technical-comparison-of-pytorch-and-jax-6458a115dcde">https://medium.com/@nijesh-kanjinghat/deep-learning-in-practice-a-technical-comparison-of-pytorch-and-jax-6458a115dcde</a></li><li>Larabel, M. (2024, August). <em>AMD’s ZLUDA CUDA implementation taken down</em>. Phoronix. <a href="https://www.phoronix.com/news/AMD-ZLUDA-CUDA-Taken-Down">https://www.phoronix.com/news/AMD-ZLUDA-CUDA-Taken-Down</a></li><li>Larabel, M. (2025, May 17). <em>NVIDIA upstreams newer GSP firmware for open-source Nouveau driver</em>. Phoronix. <a href="https://www.phoronix.com/news/NVIDIA-GSP-RM-570-Firmware">https://www.phoronix.com/news/NVIDIA-GSP-RM-570-Firmware</a></li><li>Lee, J. (n.d.). <em>TPU: Why Google doesn’t wait in line for NVIDIA GPUs (2/2)</em>. Medium. <a href="https://medium.com/@jiminlee-ai/tpu-why-google-doesnt-wait-in-line-for-nvidia-gpus-2-2-2267e4ed686f">https://medium.com/@jiminlee-ai/tpu-why-google-doesnt-wait-in-line-for-nvidia-gpus-2-2-2267e4ed686f</a></li><li>Mann, T. (2024, August 9). <em>AMD’s legal team appears to have clawed back control of much of the ZLUDA project’s code base</em>. The Register. <a href="https://www.theregister.com/2024/08/09/amd_zluda_take_down/">https://www.theregister.com/2024/08/09/amd_zluda_take_down/</a></li><li>Massed Compute. (n.d.). <em>Can cuDNN be used with AMD or Intel GPUs?</em> <a href="https://www.google.com/search?q=https://massedcompute.com/faq-answers/%3Fquestion%3DCan%2520cuDNN%2520be%2520used%2520with%2520AMD%2520or%2520Intel%2520GPUs">https://massedcompute.com/faq-answers/?question=Can%20cuDNN%20be%20used%20with%20AMD%20or%20Intel%20GPUs</a>?</li><li>Nasdaq. (n.d.). <em>Nvidia’s broadening moat: Securing the AI ecosystem</em>. <a href="https://www.nasdaq.com/articles/nvidias-broadening-moat-securing-ai-ecosystem">https://www.nasdaq.com/articles/nvidias-broadening-moat-securing-ai-ecosystem</a></li><li>NVIDIA. (2022, May). <em>NVIDIA releases open-source GPU kernel modules</em>. NVIDIA Developer Blog. <a href="https://developer.nvidia.com/blog/nvidia-releases-open-source-gpu-kernel-modules/">https://developer.nvidia.com/blog/nvidia-releases-open-source-gpu-kernel-modules/</a></li><li>NVIDIA. (n.d.). <em>CUDA runtime API: Driver vs runtime API</em>. <a href="https://docs.nvidia.com/cuda/cuda-runtime-api/driver-vs-runtime-api.html">https://docs.nvidia.com/cuda/cuda-runtime-api/driver-vs-runtime-api.html</a></li><li>NVIDIA. (n.d.). <em>GSP firmware README</em>. <a href="https://download.nvidia.com/XFree86/Linux-x86_64/510.39.01/README/gsp.html">https://download.nvidia.com/XFree86/Linux-x86_64/510.39.01/README/gsp.html</a></li><li>NVIDIA. (n.d.). <em>Software license agreement for NCCL</em>. <a href="https://docs.nvidia.com/deeplearning/nccl/sla/index.html">https://docs.nvidia.com/deeplearning/nccl/sla/index.html</a></li><li>One More Thing in AI. (2025, December 21). <em>Google &amp; Meta tie-up to combat NVIDIA monopoly</em>. <a href="https://www.onemorethinginai.com/p/google-meta-tieup-combat-nvidia-monopoly">https://www.onemorethinginai.com/p/google-meta-tieup-combat-nvidia-monopoly</a></li><li>Open Source For U. (2023, August). <em>Google and Meta bet on open source PyTorch to break Nvidia’s CUDA lock-in</em>. <a href="https://www.opensourceforu.com/2023/08/biden-administration-kicks-off-18-5m-ai-cyber-challenge-to-boost-open-source-software-safety/">https://www.opensourceforu.com/2023/08/biden-administration-kicks-off-18-5m-ai-cyber-challenge-to-boost-open-source-software-safety/</a></li><li>Phoronix. (2014, August 16). <em>NVIDIA tightens the belt on hardware: Signed firmware required</em>. <a href="https://www.phoronix.com/news/MTc5ODA">https://www.phoronix.com/news/MTc5ODA</a></li><li>PyTorch. (2023, May 2). <em>PyTorch 2.0 on Cloud TPUs</em>. <a href="https://pytorch.org/blog/pytorch-2-0-xla/">https://pytorch.org/blog/pytorch-2-0-xla/</a></li><li>Reddit. (2024, March 4). <em>[N] Nvidia bans translation layers like ZLUDA</em>. <a href="https://www.reddit.com/r/MachineLearning/comments/1b70ubo/n_nvidia_bans_translation_layers_like_zluda/">https://www.reddit.com/r/MachineLearning/comments/1b70ubo/n_nvidia_bans_translation_layers_like_zluda/</a></li><li>Reddit. (2024, August). <em>Share your GPU and Flux dev render times</em>. <a href="https://www.reddit.com/r/StableDiffusion/comments/1epnb8i/share_your_gpu_and_flux_dev_render_times_help/">https://www.reddit.com/r/StableDiffusion/comments/1epnb8i/share_your_gpu_and_flux_dev_render_times_help/</a></li><li>ResearchGate. (n.d.). <em>CUDA, TensorRT and cuDNN: Understanding NVIDIA’s ML acceleration stack</em>. <a href="https://www.researchgate.net/publication/394591605_CUDA_TENSORRT_AND_CUDNN_UNDERSTANDING_NVIDIAS_ML_ACCELERATION_STACK">https://www.researchgate.net/publication/394591605_CUDA_TENSORRT_AND_CUDNN_UNDERSTANDING_NVIDIAS_ML_ACCELERATION_STACK</a></li><li>ResearchGate. (n.d.). <em>Economic analysis report of Nvidia</em>. <a href="https://www.researchgate.net/publication/399222098_Economic_Analysis_Report_of_Nvidia">https://www.researchgate.net/publication/399222098_Economic_Analysis_Report_of_Nvidia</a></li><li>Techi. (n.d.). <em>Nvidia’s success exposes vulnerabilities: Antitrust probe &amp; geopolitical risks</em>. <a href="https://www.techi.com/nvidia-success-exposes-vulnerabilities-antitrust-probe-geopolitical-risks/">https://www.techi.com/nvidia-success-exposes-vulnerabilities-antitrust-probe-geopolitical-risks/</a></li><li>TechInformed. (2025, December). <em>Google moves to make TPUs feel native to PyTorch as it targets Nvidia’s CUDA advantage</em>. <a href="https://techinformed.com/google-moves-to-make-tpus-feel-native-to-pytorch-as-it-targets-nvidias-cuda-advantage/">https://techinformed.com/google-moves-to-make-tpus-feel-native-to-pytorch-as-it-targets-nvidias-cuda-advantage/</a></li><li>TechNewsWorld. (n.d.). <em>AMD’s AI surge challenges Nvidia’s dominance</em>. <a href="https://www.technewsworld.com/story/amds-ai-surge-challenges-nvidias-dominance-179781.html">https://www.technewsworld.com/story/amds-ai-surge-challenges-nvidias-dominance-179781.html</a></li><li>TechPowerUp. (2024, March 6). <em>NVIDIA cracks down on CUDA translation layers, changes licensing terms</em>. <a href="https://www.techpowerup.com/319984/nvidia-cracks-down-on-cuda-translation-layers-changes-licensing-terms">https://www.techpowerup.com/319984/nvidia-cracks-down-on-cuda-translation-layers-changes-licensing-terms</a></li><li>Teki, S. (2025). <em>Nvidia’s AI moat in 2025: A deep dive</em>. <a href="https://www.sundeepteki.org/blog/nvidias-ai-moat-in-2025-a-deep-dive">https://www.sundeepteki.org/blog/nvidias-ai-moat-in-2025-a-deep-dive</a></li><li>TensorWave. (n.d.). <em>Transitioning to high performance: A comprehensive guide to switching from CUDA to ROCm</em>. <a href="https://tensorwave.com/blog/transitioning-to-high-performance-a-comprehensive-guide-to-switching-from-cuda-to-rocm">https://tensorwave.com/blog/transitioning-to-high-performance-a-comprehensive-guide-to-switching-from-cuda-to-rocm</a></li><li>The Scimus. (n.d.). <em>Transitioning from CUDA to ROCm</em>. <a href="https://thescimus.com/blog/transitioning-from-cuda-to-rocm/">https://thescimus.com/blog/transitioning-from-cuda-to-rocm/</a></li><li>Thunder Compute. (n.d.). <em>ROCm vs CUDA: GPU computing</em>. <a href="https://www.thundercompute.com/blog/rocm-vs-cuda-gpu-computing">https://www.thundercompute.com/blog/rocm-vs-cuda-gpu-computing</a></li><li>TLDR AI. (2025, December 18). <em>Google works to erode Nvidia’s software advantage with Meta’s help</em>. <a href="https://tldr.tech/ai/2025-12-18">https://tldr.tech/ai/2025-12-18</a></li><li>vLLM Team. (2025, October 16). <em>vLLM on TPU: Unlocking performance with JAX</em>. <a href="https://blog.vllm.ai/2025/10/16/vllm-tpu.html">https://blog.vllm.ai/2025/10/16/vllm-tpu.html</a></li><li>Winvesta. (2025). <em>Nvidia gaming vs data center revenue analysis</em>. <a href="https://www.winvesta.in/blog/investors/nvidia-gaming-vs-data-center-revenue-analysis-what-us-stock-investors-need-to-know">https://www.winvesta.in/blog/investors/nvidia-gaming-vs-data-center-revenue-analysis-what-us-stock-investors-need-to-know</a></li><li>Zuhair, M. (2024, August 9). <em>ZLUDA, the open-source library for running NVIDIA CUDA on AMD GPUs, has now been taken down amid legal concerns</em>. Wccftech. <a href="https://wccftech.com/zluda-open-source-library-nvidia-cuda-on-amd-gpus-taken-down-amid-legal-concerns/">https://wccftech.com/zluda-open-source-library-nvidia-cuda-on-amd-gpus-taken-down-amid-legal-concerns/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4b3df5683b4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing sillyCUDA: My Quest to Learn CUDA as a Mac Enjoyer]]></title>
            <link>https://medium.com/@doidylouis/introducing-sillycuda-my-quest-to-learn-cuda-as-a-mac-enjoyer-38a28dc725d4?source=rss-7943fec93a50------2</link>
            <guid isPermaLink="false">https://medium.com/p/38a28dc725d4</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[c]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[gpu]]></category>
            <dc:creator><![CDATA[lwi00]]></dc:creator>
            <pubDate>Mon, 29 Dec 2025 16:40:29 GMT</pubDate>
            <atom:updated>2025-12-29T16:40:29.276Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/863/1*kreBICkMOaK_2oqvJe7XEQ.png" /></figure><p><strong>A 42 student’s journey into the depths of GPU architecture by building a CUDA-to-Metal translation layer.</strong></p><p>If you are a student, a researcher, or an AI enthusiast, you are likely familiar with this frustration: You love your MacBook. The Apple Silicon chips (M1, M2, M3, M4) are marvels of engineering, efficient, fast, and cool.</p><p>But then, you open a textbook on “Massively Parallel Processors,” or you try to run a specific legacy machine learning repository, and you hit a wall. <strong>CUDA.</strong></p><p>NVIDIA’s CUDA is the lingua franca of GPGPU (General-Purpose computing on Graphics Processing Units). If you don’t have an NVIDIA GPU, you are often left out of the party. While projects like ZLUDA exist for AMD, the bridge between CUDA and Apple’s Metal API is still a construction zone.</p><p><strong>Meet </strong><strong>sillyCUDA.</strong></p><h3>What is sillyCUDA?</h3><p><strong>sillyCUDA</strong> is an open-source student project with a somewhat irrational goal: to build a compatibility runtime that allows CUDA code to execute on Apple Silicon chips by translating calls to Metal.</p><p>I am currently a student at <strong>42</strong>, and I am undertaking this project not just to build a tool, but to force myself to master GPU architecture, software design, and the low-level intricacies of hardware acceleration.</p><h3>The Problem: The Green Wall</h3><p>The problem is simple:</p><ol><li><strong>CUDA is proprietary</strong> and tied to NVIDIA hardware. It is not allowed to read, fork or rebuild native CUDA calls, only uses the public API CUDA doc.</li><li><strong>Apple Silicon uses Metal</strong>, a completely different API and shading language (MSL).</li><li><strong>The gap is wide:</strong> Porting scientific code or legacy kernels from CUDA to Metal manually is time-consuming and difficult.</li></ol><h3>The Solution: A Translation Layer</h3><p>The goal of sillyCUDA is to intercept standard CUDA calls and map them to their Metal equivalents in real-time.</p><ul><li><strong>You write:</strong> cudaMalloc, cudaMemcpy, cudaLaunchKernel.</li><li><strong>sillyCUDA does:</strong> Allocates MTLBuffer, handles data transfer via Metal, and translates the kernel logic.</li><li><strong>The Mac executes:</strong> Native Metal Compute commands.</li></ul><p>It sounds hard but that’s the fun of it.</p><h3>Under the Hood: The Architecture</h3><p>Building a translation layer is a beast. To tackle it, I’ve broken the architecture down into three main components.</p><h4>1. The Runtime (cuda_runtime_api)</h4><p>This is the face of the operation. It mimics the CUDA C API headers. When a program calls cudaMalloc, my library intercepts it. Instead of talking to an NVIDIA driver, it initializes a Metal Device and allocates an MTLBuffer.</p><h4>2. The Translator (The Hard Part)</h4><p>This is where the magic (and the headaches) happens. We need to take CUDA kernels (written in a C++ dialect) and translate them into Metal Shading Language (MSL).</p><ul><li><strong>Parser:</strong> Reads the CUDA code.</li><li><strong>AST:</strong> Builds an Abstract Syntax Tree.</li><li><strong>Codegen:</strong> Spits out valid MSL code that can be compiled by the Metal backend.</li></ul><h4>3. The Metal Backend</h4><p>This manages the heavy lifting on the Apple Silicon side: managing the command queues, pipelines, and compute states to ensure the GPU gets fed the right data at the right time.</p><p><strong>Current Tech Stack:</strong></p><ul><li><strong>Languages:</strong> C++ (Runtime), Objective-C++ (Metal Interface), CMake.</li><li><strong>Target:</strong> macOS 13+ (M1/M2/M3/M4/etc).</li></ul><h3>The Roadmap: From “Hello World” to Matrix Multiplication</h3><p>I am giving myself 10 <strong>months</strong> to build a functional Proof of Concept (PoC). This is a learning marathon, which will be documented through this Medium blog.</p><p><strong>Phase 1: The Basics (Months 1–2)</strong> Before building the translator, I need to understand what I’m translating. This involves deep diving into:</p><ul><li>GPU Architecture (Threads, Blocks, Grids).</li><li>Mastering the CUDA C Programming Guide.</li><li>Writing the “hello worlds” of GPU: Vector Addition and Matrix Multiplication.</li></ul><p><strong>Phase 2: Mastering Metal (Months 3–4)</strong> You can’t translate a language you don’t speak. I will be spending these months reimplementing my CUDA learnings using native Metal Compute to understand the architectural differences.</p><p><strong>Phase 3: The Build (Months 5–7)</strong> This is where sillyCUDA comes to life. I will study existing projects like <strong>ZLUDA</strong> (CUDA on AMD) and <strong>HIPIFY</strong> to understand how they handle call interception. The goal for this phase is a &quot;Vertical Slice&quot;: <strong>getting a single Vector Addition CUDA program to compile and run on my Mac without touching the source code.</strong></p><h3>Scope and Limitations (Managing Expectations)</h3><p>Let’s be real: I am one student, not a team of NVIDIA engineers. sillyCUDA Version 1.0 will have strict limitations:</p><ul><li><strong>✅ Will do:</strong> Basic memory management (malloc, memcpy), Kernel launches, and simple math operations.</li><li><strong>❌ Won’t do (yet):</strong> Advanced shared memory, Texture memory, Streams/Events, or highly optimized cuBLAS libraries.</li></ul><p>The definition of success for this project isn’t to replace NVIDIA. It is to:</p><ol><li>Make a simple CUDA program run on a Mac.</li><li>Achieve “reasonable” performance (50% of native Metal).</li><li>Learn <em>everything</em> about how these architectures work.</li></ol><h3>Why Open Source?</h3><p>I believe the best way to learn is in public. I am releasing sillyCUDA under the <strong>Apache 2.0 license</strong>.</p><p>Whether you are a fellow student wanting to learn how compilers work, a Mac user wishing for CUDA support, or a senior engineer who wants to point out how “silly” my implementation is: you are welcome to join the ride.</p><h3>Quick Start (Join me!)</h3><p>I have just initialized the repo. Currently, it’s a skeleton, but over the coming weeks, you will see the runtime start to take shape.</p><p>You can follow the progress, star the repo, or contribute here:</p><p><a href="https://github.com/lwi00/sillyCUDA">https://github.com/lwi00/sillyCUDA</a></p><p>I will be posting updates here on Medium as I hit major milestones (like the first time I successfully trick my Mac into thinking it’s an NVIDIA GPU).</p><p><strong>Let’s make silicon silly.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/368/1*yttepnc-Oe2h6Rebi8O2mw@2x.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=38a28dc725d4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI music is about to stop]]></title>
            <link>https://medium.com/@doidylouis/the-ai-music-is-about-to-stop-1dad57f82da2?source=rss-7943fec93a50------2</link>
            <guid isPermaLink="false">https://medium.com/p/1dad57f82da2</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[digital-marketing]]></category>
            <category><![CDATA[digital-economy]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[lwi00]]></dc:creator>
            <pubDate>Fri, 24 Oct 2025 15:26:41 GMT</pubDate>
            <atom:updated>2025-10-24T15:26:41.396Z</atom:updated>
            <content:encoded><![CDATA[<h4>And you know what comes after</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*V7KwGcBDE3Q6Cm-4" /></figure><p><em>A little note about my current series the economy of attention, this time more about how the tables are turning, and how the recent announcements from the top AI labs while living one of the worst both politicial and financial crisis of the western bloc could, in fact, free us from slop way more easily than expected</em></p><p>After a euphoric rise, top AI labs like OpenAI and Perplexity now find themselves squeezed by unprecedented financial and political pressures in the Western world. Recent announcements from these companies, from OpenAI’s <strong>ChatGPT Atlas</strong> browser to Perplexity’s <strong>Comet</strong> AI web browser reveal a strategic scramble. They’re running low on both the <strong>fuel</strong> (quality data) and the <strong>oxidant</strong> (easy money) that drove the early AI boom. In response, these labs are resorting to desperate measures: launching what I call <em>“data-black-hole”</em> that voraciously ingest user data both from businesses and users, offerings what is approaching closed app ecosystems. Combined with the poor ROI of today’s AI services and the erosion of the current financial market, these moves could herald an unexpected freedom from the AI “slop” that many feared would drown our attention.</p><h3>The Well of Data Is Drying Up</h3><p>AI’s voracious appetite for training data is running up against a hard limit. Tech companies have already <strong>scraped the Internet nearly bare</strong> to feed their large language models, and now even insiders admit the gold rush is over. OpenAI co-founder Ilya Sutskever <a href="https://www.theverge.com/2024/12/13/24320811/what-ilya-sutskever-sees-openai-model-data-training">warned</a> late last year that <em>“we’ve achieved peak data and there’ll be no more” </em>likening the remaining supply of human-created internet text to a <strong>finite fossil fuel</strong> that’s nearly tapped out. In other words, <strong>there’s only one internet</strong>, and AI models have chewed through most of its readily available content.</p><p>Making matters worse, a <em>“</em><a href="https://observer.com/2024/07/ai-training-data-crisis/#:~:text=The%20study%20found%20that%20in,as%20C4,%20RefinedWeb%20and%20Dolma"><em>rapid crescendo of data restrictions</em></a><em>”</em> has hit in the past couple of years. Major websites, media platforms, and content creators are <strong>locking down their data</strong>. A recent MIT-led study found that between April 2023 and April 2024, <strong>25% of the high-quality content sources</strong> that feed popular AI training datasets became <a href="https://observer.com/2024/07/ai-training-data-crisis/#:~:text=The%20study%20found%20that%20in,as%20C4,%20RefinedWeb%20and%20Dolma"><strong>restricted or off-limits</strong></a> to web crawlers. OpenAI’s own GPTBot, for example, is now blocked from about <strong>26% of top-tier sites’ content</strong>, a dramatic shrinking of the open data firehose that AI labs once drank from freely.</p><p>To compensate, AI companies are now scrambling to <strong>buy or barter for data</strong> that used to be free. OpenAI has reportedly offered major news publishers <strong>$1–5 million deals</strong> for access to their text archives, essentially paying for premium datasets to train the next models. It has already struck content licensing agreements with outlets like <em>The Associated Press, the Financial Times, and Vox Media</em>, sometimes even <strong>trading access to ChatGPT</strong> in return for their data. In one imaginative gambit, OpenAI considered leveraging its Whisper speech-to-text tech to <strong>transcribe YouTube videos</strong> for training material. Others like Meta have gone so far as to explore <strong>acquiring entire publishing companies</strong> just to own their data troves. All these moves underscore a stark reality: the era of abundant, free training data is ending, and AI labs are <strong>running out of both open and paid data</strong> to feed the machine.</p><h3>Burning Cash and Diminishing Returns</h3><p>Hand-in-hand with the data drought comes a financial one. The economics of the AI boom are proving <strong>alarmingly unsustainable</strong>. For all the hype and astronomical valuations, <strong>almost no one is making money</strong>. OpenAI is a prime example: by mid-2025, it was <em>losing about three times more money than it was earning</em>. According to <a href="https://www.theregister.com/2025/10/15/openais_chatgpt_popular_few_pay/#:~:text=OpenAI%20is%20losing%20about%20three,to%20help%20stem%20the%20losses">internal figures</a> reported in the press, OpenAI pulled in roughly <strong>$4.3 billion in revenue in the first half of 2025</strong> yet still posted a <strong>net loss of $13.5 billion</strong> for that six-month period. Despite ChatGPT’s viral popularity (an estimated 800 million users), <strong>95% of those users aren’t paying a dime, </strong>only about 5% shell out for a subscription, leaving the vast majority of usage to be subsidized. Even the $20–$200/month paid plans have been so heavily utilized that CEO Sam Altman admitted OpenAI is <em>“losing money on its $200-per-month ChatGPT Pro plan because people are using it so much.”</em> The result is essentially a <strong>free-tier frenzy</strong> that burns through expensive computing resources without equivalent revenue.</p><p>The rush of users and features isn’t translating into sustainable business — <strong>poor ROI is rampant</strong>. Advertising was thought to be one solution, but that’s proving tricky as well. OpenAI only recently <a href="https://www.reuters.com/technology/openai-unveils-ai-browser-atlas-2025-10-21/#:~:text=competition%20for%20ad%20dollars">began considering</a> selling ads in ChatGPT’s interface (something Altman once swore he’d avoid), seeing it as a way to monetize the huge user base. Rival Perplexity tried an ad-supported model for its AI search engine but quietly <strong>paused new advertising</strong> this year after finding it wasn’t paying off.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CuWkDW8lyVfmQcZTCaGNEw.jpeg" /></figure><p>In short, <strong>the AI gravy train is running on fumes</strong>: training data is harder to get, and returns on the massive investments have yet to materialize. The result is that even deep-pocketed labs are starting to <strong>run short on money</strong> to continue business-as-usual.</p><h3>AI Browsers: Data-Black-Hole or Last-Ditch Pivot?</h3><p>How do you respond when both your training data and your profit margins are dwindling? For OpenAI and Perplexity, the answer seems to be: <strong>grab whatever data you can, directly from the users.</strong> This brings us to the latest trend after the infamous slop feed: AI-powered <strong>web browsers</strong> that double as data vacuums. In recent weeks, a lot of companies unveiled their own browsers (or browser-like apps) that deeply integrate their AI assistants into the web experience. OpenAI has Atlas, Perplexity has Comet, and Google quietly announced Gemini will soon be integrated into Chrome futur updates. Ostensibly, these products promise futuristic convenience. They can summarize any webpage, act as a shopping or research agent, and even automate tasks like booking travel or filling out forms, all from a sidebar or chat interface. But make no mistake: these AI browsers are also <strong>designed to ingests massive amounts of user behavior data</strong>, plugging a crucial hole in the labs’ data pipeline.</p><blockquote><em>Comet, Perplexity’s new AI web browser, integrates an assistant that can answer questions about whatever website you’re viewing. By baking AI into the browsing experience, companies hope to both attract users and gather rich data on their interests and behavior.</em></blockquote><p>When OpenAI launched ChatGPT Atlas (initially on macOS) in October, it explicitly framed the browser as a way to “expand into more aspects of users’ online lives”, effectively <strong>capturing their browsing behavior data for the AI, </strong>and in the same way ingesting all their past behavior via data portability if you make it your main web browser. With Atlas, every time you open a site, read an article, or shop online, ChatGPT is right there in the window. If you enable the optional <em>“</em><a href="https://openai.com/index/introducing-chatgpt-atlas/"><em>browser memories</em></a><em>”</em> feature, it will remember everything from the pages you visit to the content of those pages. In agent mode, Atlas can even navigate websites on your behalf and perform actions (like purchasing items via Instacart or scraping recipe ingredients). All of this gives OpenAI an unprecedented firehose of <strong>first-party data</strong> on what users do on the web, and those can be extremely valuable for <strong><em>fine-tuning AI models or targeting ads</em></strong><em> </em>for instance<em> (wink wink)</em>. Simply put, the browser isn’t just a product; it’s a <strong>data-collection funnel</strong> and a potential revenue machine (via ads or affiliate fees) for a company desperate to monetize its AI after failing to sell it to businesses</p><p>Perplexity’s Comet follows a similar playbook. Launched in July as <em>“a web browser built for today’s internet,”</em> Comet uses Perplexity’s AI answer engine as the default search and comes with a built-in assistant that <em>“</em><a href="https://www.theverge.com/news/703037/perplexity-ai-web-browser-comet-launch"><em>can answer questions about what’s on your screen.</em></a><em>”</em> It can even <strong>buy products or book hotels on your behalf</strong> via agent-like abilities. Initially, Perplexity tried to charge a hefty $200/month subscription for Comet’s early access, but it later dropped the paywall, making the AI browser free for all users. Why give away such a premium tool for free? Because <strong>the real gold is in user data and scale</strong>. By getting more people to use Comet as their everyday browser, Perplexity can collect streams of information on what users search, which sites they visit, and how they interact.<br><a href="https://www.theverge.com/news/703037/perplexity-ai-web-browser-comet-launch">Privacy advocates</a> have already raised concerns, calling Comet a browser <em>“built to collect more data about you,”</em> since Perplexity’s privacy policy allows linking of your searches and content interactions to your personal account. (Both OpenAI and Perplexity do offer privacy controls for instance, OpenAI says Atlas won’t train on your browsing content by default unless you opt in. But one has to wonder how many users will dig through the settings to ensure their data isn’t vacuumed up.)</p><p>Another point of why those products became free is also because they are flawed in a lot of ways, vulnerable to prompt injections and lack ability to automatize web tasks. And, as always, the best beta tester is always the final user when you have the luxury of not depends of the product.</p><p>In essence, these AI browsers represent a <strong>last-ditch pivot to secure data and users directly</strong>, cutting out the middleman. Rather than rely on scraping Google’s index or open web data (which is drying up), the labs want to own the portal through which users access the web. If successful, they gain a twofold win: a captive audience for their AI (which could later be monetized via subscriptions or ads), and a <strong>constant feed of real-world interaction data</strong> to feed into their AI improvement cycle. It’s a bold move absorbing everything a user does online into the AI’s orbit. But given the circumstances, OpenAI and Perplexity see it as necessary to keep the AI revolution humming.</p><h3>Funneling Users into Walled Gardens (B2B and Beyond)</h3><p>Another prong of the AI labs’ strategy is to <strong>tighten their grip on user funnels</strong> through partnerships and enterprise offerings. In the early boom, AI companies were happy to let users come to them organically. Now, with growth slowing and competition rising, they’re actively <strong>embedding AI into the devices and platforms</strong> people use, effectively meeting users where they already are, and in the process, locking in valuable data streams. Nowhere is this more evident than in Perplexity’s recent deal with <strong>Motorola</strong>.</p><p>In April 2025, Perplexity <a href="https://www.perplexity.ai/fr/hub/blog/announcing-our-global-partnership-with-motorola">announced</a> a global partnership to have its AI assistant <strong>pre-installed on all new Motorola smartphones</strong>, including the flagship Razr series. This move will drop the Perplexity app onto <strong>millions of Android phones worldwide by default</strong>, giving users instant access to its answer engine without even a trip to the app store. For Motorola, it’s a selling point (AI-ready phones); for Perplexity, it’s a huge B2B distribution win. In practical terms, it also means Perplexity will now directly gather data from a user’s everyday smartphone interactions (searches, commands, etc.), further filling its coffers with proprietary usage data. The funnel of user attention that once might have gone to Google or other apps is now <strong>captured in-app</strong> by Perplexity’s ecosystem.</p><p>OpenAI is pursuing a parallel path, albeit more focused on the enterprise and developer side. In mid-2023 it rolled out <strong>ChatGPT Enterprise</strong>, a business-grade version of its chatbot meant to integrate into company workflows (with promises not to use a client’s data for training). By 2024, over 80% of Fortune 500 companies had at least experimented with ChatGPT, and OpenAI was quick to tout that adoption as a sign of future revenue. The idea is clear: if individual consumers won’t reliably pay for AI, maybe corporations will, especially if AI is embedded in the tools employees already use. Microsoft (OpenAI’s key backer) has done exactly that by weaving ChatGPT-powered copilots into Office, Teams, and Windows. Every such integration <strong>funnels users into the AI’s embrace</strong> without them having to seek it out.</p><p>This B2B and integration push also reflects the reality that <strong>the easy consumer hype is fading</strong>. To actually earn money, AI labs need paying customers locked into contracts (enterprises) or hardware deals. It’s a return to classic tech business principles: distribution deals, OEM pre-installs, enterprise sales, not just viral buzz. The downside for users is the potential rise of <strong>walled gardens</strong> in AI. If one company’s assistant is built into your phone and another’s is built into your office software, each will vie to keep you inside <em>their</em> loop. The open playground of swapping AI tools might give way to closed ecosystems reminiscent of the old browser wars or mobile OS silos, except this time it’s AI assistants and data pipelines at stake.</p><h3>Hype Meets the Harsh Reality</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Hu5OUuu2NZKdtHjZ4JQBug.jpeg" /></figure><p>All these machinations occur against a backdrop of <strong>cooling hype and tightening finances</strong> in the broader market. After a frenzied 2023–2024 where “AI” was the magic word inflating stock prices and startup valuations, the sobering truth is setting in. Global venture funding has slumped drastically from its peak; the era of near-zero interest rates that fueled moonshot bets is over. High interest rates and inflation have created a cash crunch for speculative tech ventures: global VC investment in 2023 was down roughly <strong>35% from the previous year</strong>, hitting its <a href="https://kpmg.com/xx/en/media/press-releases/2023/07/global-vc-investment-remains-muted-falling-to-77-4-billion-across-7783-deals-in-q2-23.html#:~:text=Global%20VC%20investment%20remains%20muted,,high%20inflation,%20domestic%20and">lowest levels </a>since 2019. Simply put, <strong>money isn’t cheap anymore</strong>, and investors are now looking harder at actual returns.</p><p>Even AI’s biggest champions are acknowledging the bubble-like conditions. <em>“Are we in a phase where investors as a whole are overexcited about AI? My answer is yes,”</em> OpenAI’s Sam Altman <a href="https://www.reuters.com/business/finance/opinions-split-over-ai-bubble-after-billions-invested-2025-10-16/#:~:text=SAM%20ALTMAN,%20OPENAI%20CEO%20%E2%80%9CAre,media%20The%20Verge%20in%20August">confessed</a> in August.</p><blockquote>He warned that <em>“someone is going to lose a phenomenal amount of money”</em> in this boom, even as others perhaps profit.</blockquote><p>The <em>fear of missing out</em> that poured billions into any company with “AI” in its pitch is morphing into fear of a <strong>dot-com style bust</strong>. The Bank of England recently cautioned that an <strong>AI market correction could trigger a sharp tumble</strong> in tech-heavy stocks, albeit likely without destroying the whole economy. The IMF’s chief economist echoed that an AI boom <em>“may be followed by a dot-com-style bust”</em>, though it might not be systemic.</p><p>We’re essentially at a crossroads where <strong>technological advancement is colliding with economic reality</strong>. On one hand, AI capabilities continue to leap forward, new models, multimodal abilities, impressive demos abound. On the other, the financial foundation under many AI ventures is cracking: <strong>costs remain sky-high</strong>, revenue is lagging, and macroeconomic winds are no longer at tech’s back. The top AI labs now face a moment of truth. They can’t just rely on breathless innovation to carry them; they have to prove a sustainable model or risk flame-out. We may soon see a <strong>shakeout</strong> where only a few winners (likely those with big cash reserves or diversified parent companies) survive, and many of the rest either pivot away, get acquired for pennies, or shut down. As Jeff Bezos <a href="https://www.reuters.com/business/finance/opinions-split-over-ai-bubble-after-billions-invested-2025-10-16/#:~:text=JEFF%20BEZOS,%20FOUNDER%20AND%20EXECUTIVE,new%20tab%20on%20October%203">mused recently</a>, in hype cycles <em>“every experiment gets funded”</em> and it’s hard to tell good from bad but <em>“when the dust settles… society benefits from the inventions”</em> after the excess has been cleared. The dust hasn’t settled on AI yet, but one gets the sense that the <strong>music is slowing</strong> and a lot of players may soon be scrambling for a chair.</p><h3>Provocative Conclusion: Will an AI Bust Save Us from the “Slop”?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/498/0*J6ej7XE_MkRuhuCR.png" /></figure><p>It’s a delicious irony: after years of breathless warnings that hyper-advanced AI would soon drown humanity in an unstoppable flood of auto-generated <em>“slop”</em> content and seize our attention economies, the more immediate threat to the AI revolution might simply be <strong>running out of money and data</strong>. The very forces that once propelled AI’s rise, abundant free data and ultra-cheap capital, are evaporating. And as a result, the all-powerful AI labs are resorting to increasingly convoluted, almost grasping maneuvers to stay afloat: vacuuming up user data via new browsers, making <strong>Faustian bargains</strong> for training material, and begging enterprises to open their wallets. This isn’t the narrative of tech omnipotence we were sold a couple years ago; it’s starting to look more like a desperate last dance.</p><p>And maybe that’s not a bad thing. Perhaps the <em>Great AI Correction</em> will do what regulators and ethicists could not, nip <strong>the firehose of AI-generated mediocrity in the bud</strong>. If the bubble bursts, the frantic race to plaster every surface of our lives with generative AI might pause, giving society a chance to breathe and set some ground rules. The end of easy money means AI companies must prioritize what actually delivers value (and revenue), rather than simply chasing maximal <strong>engagement and attention at all costs</strong>. In a paradoxical way, the market could enforce a form of quality control and restraint on an industry that was barreling ahead irresponsibly.</p><p>None of this is to celebrate potential job losses or investor pain, a collapse always hurts. But there’s a <em>“cleansing”</em> aspect to bubbles bursting. The dot-com crash, painful as it was, cleared the way for the more disciplined internet innovations that followed. A similar shakeout in AI could <strong>separate the genuine breakthroughs from the BS</strong>. It might free us from being inundated by half-baked AI products and content whose main purpose is to chase venture capital or collect data. As one era of the AI gold rush ends, we could emerge on the other side with a saner perspective on what these technologies are actually good for (and what we truly <em>want</em> them for).</p><p>So yes, <em>the AI music </em><strong><em>might</em></strong><em> stop</em>. And when it does, some big names will undoubtedly stumble. But as the saying goes, that’s only when the tide goes out do you discover who’s been swimming naked. The coming shakeout could indeed <strong>liberate us from the slop, </strong>the endless stream of AI-generated noise and the attention-hijacking schemes, perhaps more easily than any of us expected. In the silence that follows, we might just rediscover the value of human creativity and meaningful innovation beyond the hype. And the next time the VC music starts up again, we’ll maybe know to listen a bit more critically.</p><p><strong>Sources:</strong></p><ul><li>OpenAI’s Ilya Sutskever on “peak data” and finite internet content <a href="https://www.theverge.com/2024/12/13/24320811/what-ilya-sutskever-sees-openai-model-data-training#:~:text=%E2%80%9CWe%E2%80%99ve%20achieved%20peak%20data%20and,%E2%80%9D">theverge.com</a></li><li>Study on web content restrictions cutting off AI training data <a href="https://observer.com/2024/07/ai-training-data-crisis/#:~:text=The%20study%20found%20that%20in,as%20C4%2C%20RefinedWeb%20and%20Dolma">observer.co</a>m</li><li>Reports of AI data deals (OpenAI paying for publishers, etc.) <a href="https://observer.com/2024/07/ai-training-data-crisis/#:~:text=As%20Big%20Tech%20scrambles%20to,products%20like%20ChatGPT%20in%20return">observer.com</a></li><li>OpenAI’s financials: operating losses vs revenue, 95% users on free tier <a href="https://www.theregister.com/2025/10/15/openais_chatgpt_popular_few_pay/#:~:text=According%20to%20a%20report%20published,during%20that%20six%20month%20period">theregister.com</a></li><li>Analysis of generative AI companies’ unprofitability and cloud costs <a href="https://www.wheresyoured.at/why-everybody-is-losing-money-on-ai/#:~:text=My%20gut%20instinct%20is%20that,that%27s%20my%20god%20damn%20point">wheresyoured.at</a></li><li>Reuters on OpenAI launching <em>ChatGPT Atlas</em> to collect browser behavior data <a href="https://www.reuters.com/technology/openai-unveils-ai-browser-atlas-2025-10-21/#:~:text=The%20launch%20marks%20OpenAI%E2%80%99s%20latest,competition%20between%20OpenAI%20and%20Google">reuters.com</a></li><li>The Verge on Perplexity launching <em>Comet</em> AI browser with assistant features <a href="https://www.theverge.com/news/703037/perplexity-ai-web-browser-comet-launch#:~:text=Perplexity%2C%20the%20startup%20behind%20the,%E2%80%9D">theverge.com</a></li><li>Privacy discussion of Comet browser data collection practices <a href="https://www.reddit.com/r/perplexity_ai/comments/1lxgq01/comet_is_here_a_web_browser_built_to_collect_more/#:~:text=,for%20account%20creation%20and%20management">reddit.com</a></li><li>Perplexity’s partnership with Motorola to pre-install AI on phones <a href="https://www.perplexity.ai/hub/blog/announcing-our-global-partnership-with-motorola#:~:text=We%27re%20excited%20to%20announce%20our,our%20answer%20engine%20and%20assistant">perplexity.ai</a></li><li>Investor and economic context: Altman on AI bubble, IMF/BoE warnings <a href="https://www.reuters.com/business/finance/opinions-split-over-ai-bubble-after-billions-invested-2025-10-16/#:~:text=SAM%20ALTMAN%2C%20OPENAI%20CEO%20%E2%80%9CAre,media%20The%20Verge%20in%20August">reuters.com</a></li><li>The Register on OpenAI’s user metrics and ad plans amid massive spending <a href="https://www.theregister.com/2025/10/15/openais_chatgpt_popular_few_pay/#:~:text=About%2070%20percent%20of%20OpenAI%27s,Free%2C%20%2420%2Fmonth%2C%20%24200%2Fmonth">theregister.com</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1dad57f82da2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who is feeding the “slop machine”?]]></title>
            <link>https://aws.plainenglish.io/who-is-feeding-the-slop-machine-bdd486771dbb?source=rss-7943fec93a50------2</link>
            <guid isPermaLink="false">https://medium.com/p/bdd486771dbb</guid>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ads]]></category>
            <category><![CDATA[digital-marketing]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[lwi00]]></dc:creator>
            <pubDate>Fri, 10 Oct 2025 16:16:46 GMT</pubDate>
            <atom:updated>2025-10-14T06:18:37.457Z</atom:updated>
            <content:encoded><![CDATA[<p>My 2nd article of my series, “The economy of attention”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wBAOmXSPjaPKk34-" /></figure><p>The short answer: <strong>money and data.</strong> Generative video isn’t just a new creative toy; it’s the centerpiece of an escalating battle for <strong>attention and commerce</strong>. Sama, Zucc, Elon, all are betting that <strong>short‑form AI video will drive the next wave of revenue. </strong>Subscriptions, ad targeting, personalized automated shopping, all of this is worth the launch of half‑baked products, court controversy and burn mountains of compute to get there.</p><h3>When IP Doesn’t Matter: The Sora 2 Rollout</h3><p>OpenAI’s <strong>Sora 2</strong> illustrates just how important inertia is in this race. The company debuted the model and its companion iOS app on September 30 2025 with minimal guardrails; users promptly filled the TikTok‑style feed with AI videos of SpongeBob selling drugs, South Park episodes and deepfaked celebrities. After an uproar from rights holders, OpenAI <strong>reversed its “opt‑out” policy</strong> and promised an opt‑in system with more <a href="https://www.creativebloq.com/ai/ai-art/did-openai-really-not-see-the-sora-2-copyright-controversy-coming">granular controls</a>. By then, the damage was done: Sora’s brand was synonymous with “slop,” and <strong>OpenAI admitted it still had </strong><a href="https://futurism.com/artificial-intelligence/sora-2-financial-problem#:~:text=As%20Altman%E2%80%99s%20blog%20post%20indicates,,advertising%20or%20users%20%E2%80%94%C2%A0remains%20unclear"><strong>no plan</strong></a><strong> to make money off video generation</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/404/1*TE23y1EkdTGKIBK1rVUhtA.gif" /><figcaption>skibidi SAMA</figcaption></figure><p>Why launch such a messy product? One answer is adoption. Sora head Bill Peebles said the app <strong>surpassed one million downloads in fewer than five days</strong>, outpacing ChatGPT’s growth despite being invite‑only. Those downloads matter: they create a <strong>captive audience</strong> that can be nudged toward paid tiers. BetaNews revealed that InVideo, a platform offering Sora access, sells a “Plus” plan for <strong>US $28 per month</strong> (ten credits and 50 minutes of video) and “Max” and “Generative” tiers at <strong>$50 and $100 per month</strong>. A Skywork analysis suggests that Sora 2 Pro users (available to <strong>$200/month</strong> ChatGPT Pro subscribers) pay roughly <strong>$4 per five‑second 1080p clip</strong>. Generating video at scale is expensive, <strong>Hugging Face researchers found that the energy demands of text‑to‑video models </strong><a href="https://futurism.com/artificial-intelligence/sora-2-financial-problem#:~:text=It%E2%80%99s%20a%20pressing%20issue%20for,required%20increases%20quadratically,%20not%20linearly"><strong>quadruple</strong></a><strong> when video length doubles, </strong>so OpenAI needs paying customers to offset compute costs.</p><p>Sora’s revenue strategy is still opaque, but OpenAI’s CEO Sam Altman hinted that it will include <strong>paid add‑ons for additional generations</strong> and a <strong>revenue‑share with rights holders</strong>. Investors are pressuring the company to deliver: The Information reported that OpenAI burned <strong>$7.8 billion</strong> in the first half of 2025 despite ChatGPT earning more than $1 billion per month. A successful social video platform could generate the cash needed to fund bigger AI dreams.</p><h3>Meta’s Vibes: Slop on Purpose, Ads in the Wings</h3><p>But while OpenAI stumbled into slop, <strong>Meta embraced it</strong>. In late September 2025, as described in my previous article, the company unveiled <strong>Vibes</strong>, a feed of AI‑generated shorts built into the Meta AI app. They are knitting the AI feed into its broader <strong>personalization engine</strong>. Starting December 16 2025, the company will <a href="https://www.theverge.com/news/789168/meta-ai-chats-personalized-advertising"><strong>use your conversations</strong></a><strong> with its AI assistant to personalize the posts, reels and ads you see across Facebook, Instagram and WhatsApp</strong>. There’s no opt‑out; if your AI chat mentions hiking, expect ads for boots, if you are complaining about your couple, expect dating ads. This data‑harvesting extends Meta’s <em>Advantage+</em> suite, where generative AI already <strong>turns images into animated ads</strong> and automatically applies brand logos and fonts. Meta claims these tools deliver up to a <a href="https://searchengineland.com/meta-generative-ai-tools-automated-video-branding-creative-ads-457221#:~:text=Why%20we%20care,more%20engaging,%20personalized%20content%20faster"><strong>22 % boost</strong></a><strong> in return on ad spend</strong>. Vibes is simply the next step: a cheap source of video inventory that can be personalized and monetized at scale, while collecting with way more precision what actually arouses you.</p><h3>The Next Frontier: Personalization and Agentic Commerce</h3><p>The economic allure of AI video isn’t just more eyeballs and cute Italian brainot : as always, it’s <strong>sales</strong>. Generative systems make it easy to tailor content and ads to micro‑segments. The Interactive Advertising Bureau (IAB) found that <a href="https://www.marketingdive.com/news/over-half-of-ad-buyers-are-using-generative-ai-for-video-creation-iab/753048/#:~:text=,in%202026"><strong>30 % of digital video ads in 2025 </strong></a><strong>are built using generative AI</strong>, and that share is projected to reach <strong>39 % by 2026</strong>. Among marketers using gen‑AI, <strong>42 % create multiple versions of the same video for different audiences</strong>, while <strong>38 % adjust style and 36 % refine contextual relevance. <br></strong>Smarter targeting drives ROI: Meta says its AI video tools can boost return on ad spend by <strong>up to 22 %</strong>, and CTV buyers believe that programmatic buying will deliver better ROI and scale.</p><p><strong>Are you seeing the schema, anon ?</strong></p><p>Beyond ads, the same personalization technologies are transforming <strong>shopping</strong>. <a href="https://www.bcg.com/publications/2025/agentic-commerce-redefining-retail-how-to-respond#:~:text=Consumer%20adoption%20of%20agentic%20shopping,(See%20Exhibit%202">Boston Consulting Group notes</a> that <strong>traffic to US retail sites from generative‑AI browsers and chat services surged 4,700 % year‑over‑year in July 2025</strong>, with these visitors spending <strong>32 % more time</strong>, viewing <strong>10 % more pages</strong> and bouncing <strong>27 % less often</strong>. Platforms are embedding <strong>agentic commerce</strong>: Perplexity’s <em>Buy with Pro</em> lets users make one‑click purchases; ChatGPT’s <strong>Instant Checkout</strong> allows purchases directly within the chat; and Google’s AI Mode will soon let users track prices and pay via Google Pay. As AI assistants become the <strong>default entry point for shopping</strong>, brands risk being disintermediated. BCG warns that by 2029, <strong>AI search ads in the US will reach $26 billion</strong>, about <strong>14 % of total search ad spend</strong>. And the more you slope, the more they will sell.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*_GFqeu1e9qrE44_y2BEh0A.jpeg" /></figure><h3>Slop as a Loss Leader</h3><p>Viewed in this light, AI slop feeds are <strong>loss leaders: </strong>cheap amusements designed to capture data, lock in users and condition them for on‑demand commerce. Sora’s million‑plus downloads are a data gold mine; every prompt and scroll trains a model that will eventually push <strong>personalized promotions, brand collaborations and instant purchases</strong>. Meta’s Vibes extends Advantage+ across a new video surface, using AI chat interactions to refine your advertising profile. Musk’s Grok Imagine ties video creation to paid subscriptions and will likely integrate shoppable links into X’s feed. The economic logic is clear: <strong>own the attention funnel</strong>, then <strong>sell ads and products through it</strong>.</p><p>There are risks. Training and serving these models is expensive, and users may balk at paying for AI slop. Regulators are circling around privacy and copyright issues, and right‑sholders may demand a share of the proceeds. Yet the upside is too big to ignore. If OpenAI, Meta and generally all the generative AI companies can convert even a fraction of their slop‑watchers and chat-users into shoppers, they unlock a new revenue stream that could fund dozens of new models.</p><h3>Conclusion: Following the Money</h3><p>In part one, we questioned why our feeds are overrun by AI‑generated junk. Now it’s clear: <strong>the slop is a feature, not a bug.</strong> It’s a low‑cost, high‑engagement way to train models, harvest data and pave the way for personalized ads and agent‑run shopping. The coming years will see these feeds evolve from free playgrounds into <strong>automated storefronts</strong>, where your scrolls and prompts seamlessly trigger purchases.</p><p>That raises a provocative question for the next article: <strong>If the economics of slop hinge on turning our attention into transactions, what happens when AI is both the buyer and the seller?</strong></p><p><strong>Sources :</strong></p><ul><li>Estes, Adam C. <em>Vox</em> : “Little videos are cooking our brains”</li><li>Roth, Emma. <em>The Verge</em> : “OpenAI’s Sora has already hit more than 1 million downloads”</li><li>Malik, Aisha. <em>TechCrunch</em> : “Meta launches ‘Vibes,’ a short-form video feed of AI slop”</li><li>Roth, Emma. <em>The Verge</em> : “Meta will soon use your AI chats to personalize your feeds”</li><li>Karlovitch, Sara. <em>Marketing Dive</em> : “Over half of ad buyers are using generative AI for video creation: IAB”</li><li>Hutchinson, Andrew. <em>Social Media Today</em> : “New insight suggests that xAI is seeing solid revenue increases”</li><li>The Economist. <em>The Economist</em> : “AI is turning the ad business upside down”</li><li>Malik, Nesrine. <em>The Guardian</em> : “With ‘AI slop’ distorting our reality, the world is sleepwalking into disaster”</li><li>Bharadia, Priya. <em>The Guardian</em> : “Cat soap operas and babies trapped in space: the ‘AI slop’ taking over YouTube”</li><li>Boston Consulting Group: “Agentic AI commerce could top 180 billion dollars annually”</li><li>Warmoth, Brian. <em>Digital Commerce 360</em> : “Ecommerce trends: what agentic commerce means for online retail”</li></ul><h3>A message from our Founder</h3><p><strong>Hey, </strong><a href="https://linkedin.com/in/sunilsandhu"><strong>Sunil</strong></a><strong> here.</strong> I wanted to take a moment to thank you for reading until the end and for being a part of this community.</p><p>Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? <strong>We don’t receive any funding, we do this to support the community. ❤️</strong></p><p>If you want to show some love, please take a moment to <strong>follow me on </strong><a href="https://linkedin.com/in/sunilsandhu"><strong>LinkedIn</strong></a><strong>, </strong><a href="https://tiktok.com/@messyfounder"><strong>TikTok</strong></a>, <a href="https://instagram.com/sunilsandhu"><strong>Instagram</strong></a>. You can also subscribe to our <a href="https://newsletter.plainenglish.io/"><strong>weekly newsletter</strong></a>.</p><p>And before you go, don’t forget to <strong>clap</strong> and <strong>follow</strong> the writer️!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bdd486771dbb" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/who-is-feeding-the-slop-machine-bdd486771dbb">Who is feeding the “slop machine”?</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[About the “Slopification” of society]]></title>
            <link>https://medium.com/@doidylouis/about-the-slopification-of-society-b73775a56b1b?source=rss-7943fec93a50------2</link>
            <guid isPermaLink="false">https://medium.com/p/b73775a56b1b</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[content-creation]]></category>
            <dc:creator><![CDATA[lwi00]]></dc:creator>
            <pubDate>Thu, 02 Oct 2025 10:27:59 GMT</pubDate>
            <atom:updated>2025-10-02T10:28:38.802Z</atom:updated>
            <content:encoded><![CDATA[<h4>The first article of my series, “The economy of attention”</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*yuS6waHgN0m_LwVjHWfahg.jpeg" /></figure><h4><strong>Quality content is dead. Long live slop.</strong></h4><p>AGI was supposed to unleash creativity and knowledge; instead, it has unleashed an <strong>avalanche of unbearable slop and brainrot </strong>that is overwhelming our digital lives. We’re now witnessing <em>“the largest information quality crisis in internet history”</em>, as <a href="https://ipullrank.com/ai-search-manual/geo-challenge#:~:text=The%20internet%20is%20drowning,far%20more%20insidious%3A%20AI%20slop">one analysis put it</a>. <br>In plain terms, <strong>AI slop</strong> refers to the mass of low-quality, easily produced content made by generative AI (yes those linkedin posts) all <strong>churned out with minimal human effort or originality</strong>. It’s fast, cheap, and often sloppy in accuracy or coherence, dripping with emojis and windy phrases. And thanks to the <strong>economics of online attention</strong>, this digital junk is flooding every corner of the internet at machine speed.</p><p>How did we get here? Just a few years ago, generative AI was the shiny new revolution in tech. Tech visionaries promised democratized creativity; anyone could make art, write articles, or produce videos with AI assistance. <strong>That hype has quickly curdled into disillusionment.</strong> Today, AI isn’t just helping humans create, in fact, experts have predicted that <strong>up to 90% of online content could be AI-generated by 2025–2026</strong>, and we’re already seeing the signs. A Stanford study of hundreds of millions of documents found a <em>sharp surge in AI-written text</em> after ChatGPT’s release in late 2022. On LinkedIn, over <strong>54% of long posts are likely AI-authored</strong>. News sites, blogs, and forums are quietly filling up with algorithmically generated prose. What began as a promising tool to <em>enhance</em> content creation has instead enabled a <strong>deluge of low-effort content</strong> that’s crowding out the genuine, high-quality material.</p><h3>From Creative Promise to Content Chaos</h3><p>Crucially, the explosion of this slop content is enabled by AI tools that make content creation <strong><em>trivial, </em></strong>and more importantly<strong><em>, easy and stupidly fast</em></strong>. You no longer need cameras, studios, or even writing skills, you can just input a text prompt and let the generative model do the rest. As a result, the <em>cost</em> (in time, effort, money, or even thinking) to create a piece of content has plummeted. <strong>It’s <em>fast, easy, and inexpensive</em> to pump out AI-made videos, images, or articles</strong>. A single person (or bot) can spawn an entire <em>“content farm”</em> of AI-generated posts in minutes.</p><p>Platforms have enthusiastically provided these tools to users. For instance, Google recently rolled out its <strong>Veo 3</strong> text-to-video model that can auto-generate 8-second video clips (with sound) from a prompt, a <a href="https://www.vox.com/technology/419430/ai-tiktok-youtube-shorts-instagram-reels">tool one reviewer</a> called <em>“a slop-monger’s dream”</em>. String a few AI clips together, and you’ve got a ready-made short video for social media. YouTube announced that this technology would be built directly into YouTube Shorts, meaning creators can populate the feed with algorithmically generated videos at the click of a button. Not to be outdone, <strong>TikTok and even OpenAI have been launching similar generative media features</strong> as of late 2025. In other words, the very apps that command billions of eyeballs are actively <em>inviting</em> more AI slop into their ecosystems.</p><p>It’s no wonder then that <strong>AI-generated content is flooding our feeds</strong>. We’ve known this was coming, but when a lot of people thought that people would adapt to this societal change, most of the people were, in fact, <strong>one shotted.</strong></p><p>Consider <strong>YouTube</strong>, traditionally a home for human-created videos: an August 2025 Guardian analysis found that nearly <em>1 in 10 of the fastest-growing YouTube channels globally consisted entirely of mass-produced AI videos</em>. Out of the top 100 rapidly growing channels, <strong>nine were pure AI content </strong>featuring surreal fare like <em>“babies trapped in space, zombie football stars, and cat soap operas,”</em> as the report described. These channels aren’t niche oddities; together they have millions of subscribers (one channel of an AI-animated infant in space had 1.6 million subs, and a cat drama channel “Super Cat League” boasted 3.9 million). Their videos <strong>rack up views in the millions</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/251/0*jE1PgfGvVnknmmSE" /><figcaption>The so-said “Super Cat League”</figcaption></figure><p>On <strong>TikTok and Instagram</strong>, the situation is arguably even more pronounced. The <em>short-form video boom</em> has created an environment where <strong>quantity and immediacy trump quality, </strong>perfect conditions for AI slop to thrive. TikTok’s average user now spends about <strong>95 minutes per day</strong> scrolling an infinite feed of quick videos, far more time than on any other social network, even if most of them are purely AI generated. To keep viewers hooked, TikTok’s algorithm rewards <em>constant output</em> and viral visuals.</p><h3>Meta’s “Vibes” and the Embrace of Infinite Slop</h3><p>Nothing illustrates this trend better than Meta’s latest move. In September 2025, <strong>Meta (Facebook’s parent company) </strong>rolled out a new feature called <strong>“Vibes”</strong> , effectively <em>a TikTok-style feed composed entirely of AI-generated videos</em>. As TechCrunch quipped in the announcement, <em>“Think TikTok or Instagram Reels, but </em><strong><em>every single video</em></strong><em> you come across is essentially AI slop.”</em>. Mark Zuckerberg himself demonstrated Vibes in a flashy Instagram post, showing off surreal AI videos: <em>“a group of fuzzy-looking creatures hopping between fuzzy cubes; a cat kneading dough; an ancient Egyptian woman taking a selfie on a balcony in Ancient Egypt”</em> . In the Vibes feed, not a single clip is real footage; <em>the entire experience is algorithmically generated</em>. Users can prompt the AI to create a video from scratch or <strong>remix</strong> an existing AI video, add AI-generated music or styles, and then share it instantly. Meta even partnered with well-known AI image generators (like Midjourney) and a startup called Black Forest Labs to power the early version of Vibes, underscoring that this is a serious investment, not just a gimmick. Essentially, one of the world’s biggest tech companies has built a dedicated firehose for <em>AI-produced mini movies</em>. A real attention black hole capable of absorbing the attention of millions, if not billions, of users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UZt69n1x1cYq3-MTy4USjw.jpeg" /></figure><p>We’re already at the edge of the attention event horizon. Past this line, everything flattens into the same glow: news, jokes, grief, AI dreams. The feed optimizes; we rationalize. And the more we scroll, the more the system learns our orbit.<br>But black holes don’t appear from nowhere. They’re <strong>built</strong>, layer by layer, from incentives we rarely name: ad markets, payout formulas, watch-time targets, the cost curves of synthetic media. That’s part two of this study. This one ends with an open problem: <strong>what is the business model of our boredom?</strong> Until we can answer that (not morally, but <strong>economically</strong>)the slop keeps compounding. Next, I’ll map the money trails and ask the uncomfortable question: are we the audience, the product, or the subsidy?</p><p><strong>Sources:</strong></p><ul><li>Malik, Nesrine. <em>The Guardian</em> : “With ‘AI slop’ distorting our reality, the world is sleepwalking into disaster”</li><li>Nemeroff, Adam. <em>Live Science : </em>“AI slop is on the rise, what does it mean for how we use the internet?”</li><li>Bharadia, Priya. <em>The Guardian</em> : “Cat soap operas and babies trapped in space: the ‘AI slop’ taking over YouTube”</li><li>Malik, Aisha. <em>TechCrunch</em> :“Meta launches ‘Vibes,’ a short-form video feed of AI slop”</li><li>Perez, Sarah. <em>TechCrunch</em> : “YouTube prepares crackdown on ‘mass-produced’ and ‘repetitive’ videos, as concern over AI slop grows”</li><li>Estes, Adam C. <em>Vox : “</em>Little videos are cooking our brains”</li><li>Focal AI Blog : “Why ‘AI Slop’ Is Flooding TikTok, Facebook, and YouTube Shorts”</li><li>iPullRank : “The Content Collapse and AI Slop, A GEO Challenge”</li><li>Buffer (DataReportal) : <em>TikTok Statistics 2025</em> (average daily usage)</li><li>Wired — <em>“Yes, That Viral LinkedIn Post… Was Probably AI-Generated”</em></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b73775a56b1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ML is dead, long live ML]]></title>
            <link>https://medium.com/@doidylouis/ml-is-dead-long-live-ml-620779c0ee54?source=rss-7943fec93a50------2</link>
            <guid isPermaLink="false">https://medium.com/p/620779c0ee54</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[lwi00]]></dc:creator>
            <pubDate>Sun, 28 Sep 2025 13:31:10 GMT</pubDate>
            <atom:updated>2025-09-28T14:39:52.711Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Are you ready for the sklearn resurgence anon ?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m-grSW2QQR-YNBx1SpJhQA.png" /></figure><h3>From Hype to Reality: Machine Learning’s Resurgence</h3><p>Over the past few years, <em>generative AI</em> (e.g. large language models and image generators) has dominated headlines and overshadowed more traditional machine learning in the data science world. This led some to believe classical ML was becoming obsolete : a perceived <strong>“fall”</strong> of ML amid the generative AI hype. However, as the initial frenzy cools, the industry is experiencing a <strong>“rise”</strong> of machine learning once again. Early excitement around generative AI has been tempered by risk concerns and <em>disappointment</em> in practical results (e.g. a lack of killer apps and weak ROI from early projects, as illustrated by the now famous headline of 95% inefficient projects).</p><p>Gartner’s hype cycle describes generative AI as sliding into a <em>“trough of disillusionment”</em> — interest wanes as early implementations fail to meet high expectations. Many companies are now confronting the <strong>limits</strong> of generative models as general-purpose solutions and re-evaluating their approach, mainly due to a general lack of precision and predictability in generative AI results. In this reality check, classical machine learning is proving to be the more <strong>understood, reliable, and ROI-driven</strong> tool for core business problems. Generative AI is often kept to <em>“low-stakes”</em> applications and corporate-friendly PoC, whereas traditional ML offers a clearer line of sight to value on mission-critical tasks.</p><h3>Massive Compute Investments Lower the Cost of AI</h3><p>However, if the generative AI has driven both the SP500 and NVIDIA stocks to new ATH, it has lowered the computational price to abyssal lows. Like the Dot Com bubble has seen the development of cheap internet access to millions of households, the AI bubble is witnessing one of the biggest <strong>increases in affordable computing power</strong> resulting from huge capital expenditures in recent years. Hyper-scale tech firms have spent tens of billions to build AI data centers and specialized hardware, dramatically expanding available compute.<br>The top four hyperscalers (Google, Amazon, Meta, Microsoft) collectively make up about half of this spending, investing in <strong>vertically-integrated chips and architectures to <em>optimize performance and lower the cost of compute</em></strong>. As a result of this arms race in AI infrastructure, global data center investments are hitting record highs — projected to reach <strong>$1.2 trillion by 2029</strong> (21% CAGR) — fueled largely by AI needs. These massive outlays have, in effect, <em>subsidized </em>the compute costs for the whole industry, making advanced ML and AI workloads more accessible.</p><p>Concrete examples illustrate how <strong>compute cost per task is dropping</strong>. OpenAI reportedly spent on the order of <strong>$80–100 million</strong> to train GPT-4, a state-of-the-art generative model. Yet by late 2024, a Chinese lab (DeepSeek) claimed it trained a comparable model for only <strong>$3 million</strong> using 2,000 GPUs and an over <strong>30×</strong> cost efficiency improvement.</p><p>The success of smaller, efficient models has <strong>“punctured some of the capex euphoria,”</strong> suggesting ever-bigger spending isn’t the only path forward. The <strong>bottom line</strong> is that thanks to economies of scale, new hardware, and algorithmic efficiency gains, the cost of machine learning computation has been trending <strong>downward</strong>. Lower AI compute costs are expected to <strong>enable broader AI adoption</strong> across industries, from automotive to smartphones, by lowering barriers to entry and making advanced models economically viable in more contexts. This drop in cost strengthens the case for applying <em>ML solutions</em>, since projects that might have been cost-prohibitive a few years ago (due to required GPU hours or cloud bills) are increasingly within reach for many organizations.</p><h3>Enterprises Realize the Limits of Generative AI</h3><p>Another factor in the “rise” of classical ML is a growing realism among industries about what <strong>generative AI can and cannot do</strong> in production. The initial wave of generative AI pilots often hit stumbling blocks when applied to enterprise needs like accuracy, compliance, and maintainability. These challenges have led to a degree of <strong>“AI solution fatigue”</strong> in some organizations that expected magic from GenAI. As one governance expert noted, the obsession with generative AI was somewhat <em>myopic</em>, and it’s time to remember that other AI technologies (like traditional ML) can drive major impact, and now for a 1/10th of the cost.</p><p>We now see companies recalibrating their strategies: rather than trying to generalize a generative model to every problem, they are picking the <em>right tool for each job</em>. In many cases, a well-tuned gradient boosting model or a straightforward classification algorithm can outperform a huge language model <em>on a specific structured task</em>, with less cost and risk. Moreover, the trust factor is leading firms back to interpretable ML models for high-stakes uses.</p><p>Finally, the <strong>ROI pressure</strong> is reinforcing this pragmatism. As noted, many executives have started questioning the <em>return on large AI investments if meaningful revenue streams don’t materialize in the near term</em>. It’s telling that <em>enterprise AI spending is now scrutinized for tangible value</em>. Machine learning projects with clear metrics (e.g. reduced churn by X%, improved yield by Y%) are easier to justify than open-ended generative experiments. With economic uncertainty, leadership tends to favor proven solutions — which often means doubling down on classical predictive models and incremental AI improvements rather than moonshots. Of course, generative AI will continue to advance and find its footing (as Gartner’s <em>“slope of enlightenment”</em> phase suggests). But in the interim, the shine has worn off enough that <strong>good old-fashioned machine learning is back on center stage</strong>, solving the problems that matter with reliable techniques.</p><p>In short, the industry is moving past the hype cycle peak and into a phase of <em>balance</em>: pairing the creativity of generative AI with the dependability of classical ML. Combined with the development of ultra light models, the future seems to be oriented towards “all-local” models.</p><p><strong>Sources:</strong></p><ol><li>Mardi Witzel — <em>“Generative AI and the Trough of Disillusionment,”</em> CIGI (Nov 2024) — Discusses cooling hype, ROI concerns, and renewed focus on mature AI</li><li>Rob Petersen — <em>“ROI of AI — 6 Steps to ROAI,”</em> BarnRaisers (Apr 2025) — Cites study: 73% of firms invest $1M+ in GenAI, but only ~33% see positive ROI</li><li>Sebastian Moss — <em>“AI Capex Leads US Economy… Data Center Capex to $1.2Tn by 2029,”</em> DCD (Aug 2025) — Reports one-third of data center spending is on AI hardware; hyperscalers driving costs down</li><li>McKinsey &amp; Co. — <em>“The Cost of Compute: $7 Trillion Race to Scale Data Centers,”</em> (Apr 2025) — Notes $6.7T needed by 2030; mentions OpenAI’s ~$100M GPT-4 cost vs $3M rival claim</li><li>Manish Singh — <em>TechCrunch:</em> <em>“DeepSeek ‘punctures’ AI leaders’ spending plans…,”</em> (Jan 2025) — Describes Chinese model trained for $5.6M, analysts on lower-cost AI future</li><li><strong>SuperAnnotate Blog</strong> — <em>“Direct Preference Optimization (DPO): Complete Overview,”</em> (Sep 2024) — Explains DPO vs. RLHF, noting DPO’s simplicity and lower compute needs</li><li><strong>DataCebo Team</strong> — <em>“Synthetic Data in 2024: Year in Review,”</em> (Dec 2024) — Highlights synthetic data market growth to $3.1B<a href="https://datacebo.com/blog/synthetic-data-in-2024-the-year-in-review/#:~:text=A%20new%20report%20from%20Allied,larger%20landscape%20of%20synthetic%20data">datacebo.com</a> and examples of using GenAI to boost classical ML (e.g. Meta Llama 3.1 with synthetic data)</li><li>Payal Rawat — <em>“Why Classical Machine Learning Still Matters in a GenAI World,”</em> TransOrg Blog (Aug 2024) — Emphasizes classical ML strengths: interpretability, data efficiency, robustness</li><li>Rahul — <em>“Beyond Generative AI: Classical ML is Still Essential,”</em> Medium (Sep 2025) — Argues that enterprises should integrate both approaches; GenAI as catalyst, ML as backbone</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=620779c0ee54" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>