The Great Copycat Carousel: Everyone Copies Everyone

Amazon pulls an Nvidia move and is now also selling chips

Amazon has made a remarkable discovery: The company already operates one of the three largest data center chip businesses in the world, it just never counted itself as a customer. During the quarterly earnings call, CEO Andy Jassy announced that Amazon's semiconductor division has already reached an annual revenue of $20 billion. If Amazon were to sell its in-house chips (Graviton processors, Trainium AI chips, Nitro security chips) to its own cloud division, AWS, as Intel or Nvidia do, revenue would be at $50 billion. The business is growing by over 100 percent annually. Demand is exploding: Trainium2 is sold out, Trainium3 is almost fully reserved, and even Trainium4—still 18 months away from wide availability—is already pre-ordered. OpenAI and Anthropic have jointly secured seven gigawatts of Trainium capacity, and Meta is using Graviton cores for its AI agents. → Techpresso

Synthszr Take: Amazon is accidentally practicing what economists call “vertical integration,” but has kept it hidden on its balance sheet for years. The company produces chips for its own use and is only now realizing that it's competing with Intel and AMD. This is reminiscent of cities that operate their own power grids and suddenly realize they've become energy suppliers. The $225 billion in Trainium commitments show: Whoever controls the means of production for AI dictates the terms. Jassy's calculation reveals the true power shift in the tech sector: It's not the models that are scarce, but the chips they run on.

Google pulls an Amazon move and is now also selling chips

Alphabet is opening its hardware vault and, for the first time, selling its in-house Tensor Processing Units (TPUs) to select customers for their own data centers. The two new TPU generations for training and inference are intended to compete directly with Nvidia's dominance. Anthropic and Meta have already signed deals. Meanwhile, OpenAI has effectively abandoned its ambitious Stargate project with 20 planned data centers, opting instead for more flexible compute leasing models. Control disputes between the project partners and OpenAI's precarious financial situation (analysts expect liquidity problems by mid-2027) are forcing a change in strategy. → TLDR AI

Synthszr Take: Google is making a virtue out of necessity, turning its TPU development into a franchise system. Instead of just offering cloud access as before, it's now selling the hardware directly to customers who want to run their own infrastructure. This is reminiscent of Intel's foundry strategy: If you can't use all the chips yourself, you become a supplier for others. The irony: While Google decentralizes its infrastructure, OpenAI is failing at this very attempt—Stargate. The difference lies in the starting position: Google already has the hardware, OpenAI wanted to build it. In a world where compute is becoming the bottleneck, the winner is the one who controls the machines, not the one who trains the smartest models.

Musk pulls a China move and is now also copying OpenAI models

Elon Musk has testified under oath in a federal court in Oakland that his AI lab, xAI, has used OpenAI models “in part” to train its own systems. When asked by OpenAI's lawyer, William Savitt, about “distillation”—a technique where smaller AI models are trained to mimic the behavior of larger models—Musk initially answered evasively: “Generally, all AI companies do that.” Only upon further questioning did he confirm with “in part” that xAI also uses this practice. The testimony came during the ongoing legal dispute between Musk and OpenAI, which centers on Musk's failed takeover attempts and his subsequent competitive battle against the ChatGPT developer. The telling detail: OpenAI itself is actively fighting against the distillation of its models by competitors, particularly by Chinese labs like DeepSeek, and has, by its own account, “taken steps to harden our models against distillation.” → Wired

Synthszr Take: Musk's confession reveals the double standard of the AI industry: While OpenAI portrays itself as a victim of Chinese model piracy and the Trump administration takes protective measures against “foreign distillation,” American labs are happily copying each other. The scene is reminiscent of medieval craft guilds that guarded their secrets from outsiders, while journeymen moved between workshops and took knowledge with them. Anthropic has already drawn consequences, blocking both OpenAI and xAI from accessing its Claude models. What Musk defends as “standard practice” turns out to be a fragile gentleman's agreement in an industry that oscillates between cooperation and competition. The real joke: While everyone rails against China's alleged technology theft, American AI CEOs are admitting in court that they do the same.

Google pulls an OpenAI move and is integrating ads into Gemini

Google is keeping all its options open: While the company has not yet shown any ads in its Gemini app, Chief Business Officer Philipp Schindler hints that this could change. The initial focus is on “AI Mode,” the AI-powered conversational version of Google Search, where ads are already being tested. Schindler emphasizes that formats that work in AI Mode can be successfully transferred to the Gemini app. As recently as January, Google's VP of Global Ads assured that there were “no plans” for ads in Gemini. However, on the earnings call, Schindler argued that ads have historically been crucial for scaling products to billions of users. “When done well, ads can be really valuable and helpful,” said Schindler. OpenAI is already testing ads in ChatGPT for free and low-cost subscription tiers, while Anthropic is making fun of this decision. → Business Insider

Synthszr Take: Google is following the natural law of digital platforms: first attract users, then monetize. The cautious rhetoric (“no rush,” “when done well”) is reminiscent of Facebook in 2012, when Zuckerberg swore that the user experience was more important than ad revenue. Today, Meta makes 98% of its revenue from ads. The real leverage lies in the fusion of AI answers and commercial recommendations: If Gemini suggests a recipe, why not order the ingredients from a partner right away? Google is already testing this blurring of lines in “AI Mode,” where the distinction between organic search results and advertising becomes hazy. Anthropic's mockery of OpenAI's ad experiments seems like the arrogance of a startup that hasn't yet grasped that even a $10 subscription fee barely covers inference costs. The question is not if, but how subtly the commercialization will be implemented.

China Speed through AI: Uber takes on booking.com

Uber is expanding its offerings to include hotel bookings, a move that only looks like diversification on the surface. US customers can now book over 700,000 hotels worldwide directly in the app, made possible through a partnership with Expedia Group—the company that Uber CEO Dara Khosrowshahi led for twelve years. Later this year, vacation rentals from Vrbo will also be added. Uber One members receive a 20% discount on select hotels and 10% cashback in the form of Uber Credits. According to CTO Praveen Neppalli Naga, the new features—including AI-powered voice booking and enhanced search functions—were developed within a few months, a process that would have previously taken at least a year. This was made possible by agentic AI tools like Cursor, which have fundamentally changed software development at Uber since late last year. → Techpresso

Synthszr Take: Khosrowshahi now calls Uber “an app for everything,” but that misses the point. What we're seeing here isn't a strategic expansion, but the result of evolutionary pressure from AI agents. As users increasingly book through chat interfaces and voice assistants (“Hey, organize a trip to Berlin for me”), apps must become all-in-one providers or disappear into the aggregation layer. The acceleration from twelve to six months of development time through agentic AI is just the enabler—the real driver is the looming irrelevance of single-purpose apps in a world where AI agents handle the orchestration. Uber isn't building a hotel feature because it wants to, but because in two years, it would otherwise just be a backend supplier for GPT-7.

Pinterest: AX is the new UX

Pinterest has developed a specialized two-tower retrieval model that optimizes shopping ads for actual purchases instead of just clicks. The problem: Traditional engagement signals like clicks are abundant but correlate only weakly with real purchase intent. The new architecture combines parallel DCN v2 and MLP Cross Layers with a multi-task approach, uses clever training techniques for the sparse and noisy conversion data, and implements an advertiser-specific loss function. The system has to deal with the fundamental asymmetry between abundant click data and rare purchase events. Pinterest solves this with a hybrid architecture that learns from frequent interaction signals while also prioritizing valuable but rare conversion events. → TLDR Data

Synthszr Take: Pinterest is building the invisible infrastructure for a world where AI agents will handle most online purchases. Humans still browse for fun; agents shop with a clear mission: “Find me a blue summer dress under $100.” The two-tower architecture acts as a translator between two worlds—one optimized for human dopamine hits (clicks, likes, saves) and one that only counts the final result (purchased or not). This is reminiscent of urban planning: Early cities were optimized for horse-drawn carriages but then had to be rebuilt for cars. Pinterest is currently optimizing from “people who pin” to “agents who buy.” The real infrastructure advantage lies not in the model itself, but in the ability to serve both worlds simultaneously—until agents eventually become the majority. The User Experience (UX) is becoming the Agent Experience (AX).

Spotify Badges Against AI Artists: Verification Instead of Prohibition

Spotify is introducing a “Verified by Spotify” badge to distinguish human artists from AI-generated acts. For verification, artists must demonstrate a verifiable presence both on and off the platform: concert dates, merchandise, linked social media accounts. AI-generated music or AI persona artists are explicitly excluded from verification. Additionally, Spotify requires consistent listener activity over a longer period, not one-off engagement spikes. At launch, over 99% of actively searched artists will be verified, most of them independent artists from various genres and career stages. The green checkmarks will appear next to artist names in search results and on profiles in the coming weeks. In parallel, Spotify is testing an “Artist Profile Protection” feature that gives artists control over releases under their name—a response to Sony's takedown request for over 135,000 AI-generated songs that imitated their artists. → Techpresso

Synthszr Take: Spotify is reinventing the blue-check model for the music industry, but this time it's not about status, but about ontological categorization: human or machine. The platform is deliberately choosing the path of labeling over banning—a clever move reminiscent of the food industry, where organic labels don't displace existing production methods but make them transparent. What's interesting is the definition of “authenticity” through external signals (concerts, merch, social media), as if human creativity can only be proven through physical presence. The 44% rate of AI-generated tracks on Deezer shows: this isn't a fringe phenomenon, but a new normal. Spotify is betting that, with sufficient transparency, listeners can decide for themselves whether they want to listen to the algorithm or the human—a more democratic solution than gatekeeping, but also one that tacitly accepts the coexistence of man and machine in the creative space.

Apple Halts Vision Pro: When Hardware Outpaces Fiction

Apple is pulling the plug on the Vision Pro after disappointing sales figures. The model, refreshed in October with an M5 chip, failed to convince the market despite an improved headband—the price remained at $3,499. According to MacRumors, Apple is already reassigning the product team to other projects, particularly to AR glasses in the style of Meta. CEO successor John Ternus had spoken just this month of being in the “early innings” of spatial computing. The strategy behind this: Apple was gathering insights with the Vision Pro for cheaper successor products, possibly using multiple A-series chips instead of expensive M-processors. CitiGroup predicts a market of $40 billion for AR glasses by 2030. → Techpresso

Synthszr Take: Apple is failing not because of the technology, but because of the fiction. The Vision Pro is like a concept car that was actually produced: technically brilliant, but no one knows where to drive it. The problem is reminiscent of early smartphones before the App Store—the hardware was there, but the ecosystem was missing. Apple hoped that developers would eventually find the killer apps, while simultaneously limiting access with a $3,499 price tag. A classic chicken-and-egg problem in its purest form. The reassignment of engineers to AR glasses shows that Apple has understood: spatial computing first needs a social function (like Meta's Ray-Bans offer) before it can become a personal computing machine. The Vision Pro was a $3,499 prototype sold as a product.

Zuckerberg Puts $500 Million into AI Biology: Tech Money Seeks New Fields

Mark Zuckerberg and Priscilla Chan are investing $500 million in Biohub, a virtual biology initiative that applies artificial intelligence to the simulation of human cells. $400 million will go towards data generation and imaging technology, while $100 million will go to external research labs. Partners like Nvidia, the Allen Institute, and Arc are intended to create open datasets that will serve as a common basis for AI biology research. Alex Rives of Biohub explains that current datasets, with a maximum of one billion cells, are far from sufficient—an order of magnitude more is needed. The goal: to train AI models that can understand and reprogram diseases at the level of cells, molecules, and tissues. In parallel, Mayo Clinic's REDMOD, an AI system that detects pancreatic cancer three years earlier than human radiologists, shows where things could be headed. → The Rundown AI

Synthszr Take: Tech billionaires are turning biology into an infrastructure problem. What Zuckerberg is doing here follows the classic Silicon Valley playbook: If the data is missing, buy the machines that produce it. The $400 million for “data generation and imaging technology” isn't research funding; it's the construction of a biological data factory. This is reminiscent of the early days of Google Street View—first build the camera cars, then dominate the maps. The crucial difference: For street images, millions of data points were enough; for cells, we're talking trillions. Mayo Clinic's REDMOD already shows what's possible when you have enough medical scans. Zuckerberg's bet is that biological systems behave like language models: more data, better predictions, and eventually, emergent abilities.

Codex Has One Rule: Never Talk About Goblins – Easter Eggs Are Debugging in Production

OpenAI's new Codex CLI contains a curious instruction in its system prompt: GPT-5.5 is under no circumstances allowed to talk about “Goblins, Gremlins, Raccoons, Trolls, Ogres, Pigeons or any other animals or creatures”—unless it is “absolutely and unequivocally relevant” to the user's query. The warning appears twice in the over 3,500-word base instructions that were published last week in the open-source code on GitHub. Previous GPT versions did not have this specific restriction, suggesting that OpenAI is combating a new problem: users have reported GPT's sudden tendency to mention goblins in completely inappropriate conversations. OpenAI employee Nick Pash insists it's not a marketing gag, while CEO Sam Altman picks up on the joke: “Feels like Codex is having a ChatGPT moment. I meant a goblin moment, sorry.” The situation is reminiscent of xAI's Grok, which last year briefly mentioned “white genocide” in South Africa on every topic—a problem caused by an “unauthorized modification” of the system prompt. → arstechnica.com

Synthszr Take: System prompts are the new Terms of Service: nobody reads them until something goes wrong. OpenAI's goblin ban shows how emergence works in language models—not as a mystical property, but as unpredictable pattern formation that is subsequently patched up with artisanal fixes. This is reminiscent of biological systems where seemingly meaningless DNA sequences suddenly become active and produce unexpected phenotypes. The irony: The more we humanize these models (“vivid inner life,” “warm, curious, collaborative”), the more we have to tame them with explicit prohibitions. Why goblins specifically? GPT-5.5 likely developed a strong association in its training data between code debugging and fantasy metaphors—bugs become gremlins, problems become monsters. OpenAI is debugging in production here by suppressing the symptoms instead of understanding the cause. The real punchline: The moment system prompts become public, they transform from control mechanisms into playgrounds for prompt hackers.