ChatGPT Nibbles at Google’s Search Monopoly

ChatGPT reportedly reaches one-fifth of Google’s search volume

OpenAI claims that ChatGPT now reaches roughly 20 percent of Google’s search volume. Adam Fry, product lead for ChatGPT’s consumer business, said the number of “search-like queries” sits at “roughly one-fifth” of queries on Google. He made the remarks Monday evening at a dinner hosted by OpenAI in New York in conversation with tech journalist Alex Kantrowitz. The transcript was published on Tuesday. According to common industry surveys, Google holds more than 90 percent of the global search engine market when AI tools are not counted. Microsoft’s Bing remains in the single-digit percentage range.

OpenAI determines the methodology itself. “Every time you ask a question and go search the web, that’s a search,” Fry said. Accordingly, every question that prompts ChatGPT to run a web search counts individually, even within a single conversation. In Fry’s description, anything for which users would previously have turned to a traditional search engine counts as a search. Fry did not state which Google figure he used for comparison. He cited neither the underlying timeframe nor the regions covered. Nor did he provide an absolute number. Taking the frequently cited estimate of around 16 billion Google searches per day as a baseline, one-fifth would correspond to roughly 3.2 billion queries daily.

Fry described how usage patterns have shifted. Initially, users primarily came for help with writing and coding—tasks they would never have typed into a search engine. Over the past year, queries about people, places, and things requiring up-to-date information have emerged, alongside financial and health-related questions. In these cases, ChatGPT searches the web. Previously, the model relied exclusively on training data up to a cutoff date; now it draws on external sources for many queries.

Fry was more specific with other metrics. According to him, ChatGPT has 1.2 billion weekly active users. Around 200 million were added in the past two months alone, an increase of 20 percent. The company states that its models generate about 1.5 billion images per week. Roughly 250 million people are said to use ChatGPT Voice. The statements come ahead of a planned initial public offering by OpenAI. → Gizmodo, 9to5Google, Search Engine Journal, Business Insider

Synthszr Take: OpenAI counts every web search within a conversation separately, whereas behind almost every Google query sits an ad auction. Fry is measuring volume that has yet to translate into revenue. Ahead of an IPO, this is precisely the most valuable metric: it borrows Google’s market power as a benchmark without having to deliver Google’s advertising revenue. Timeframe, region, and baseline value are all missing. For advertisers, the metric that truly counts is the one OpenAI doesn’t mention: revenue per query.

Enterprise AI Spending Quintuples to $177 Billion

Enterprise spending on AI deployment surged from $33 billion to $177 billion within a single year as of September 2026, according to a study by venture capital firm Menlo Ventures. The authors report that growth has accelerated: up 5.4-fold compared to 3.9-fold the previous year. According to Menlo, every dollar invested in model development this year generates $1.70 in revenue. 69 percent of spending goes to frontier labs, compared to 50 percent the prior year. In enterprise API spending, Anthropic’s share climbed from 40 to 55 percent and OpenAI’s from 27 to 31 percent, while Google dropped from 21 to 9 percent. → Techpresso

Synthszr Take: Google delivers strong benchmarks and yet saw its share of enterprise API spending more than halve from 21 to 9 percent. Agents tie spending to work accomplished rather than seat counts. This turns every invoice into a performance audit, and coding is the discipline where that test is toughest. A model without a daily proving ground slips into the open-source role: 19 percent usage, but only 4 percent revenue.

Apple Warns of Unprecedented Memory Price Spikes Driven by AI Data Centers

Apple has warned of an “unprecedented” surge in memory chip prices as the company competes with AI data centers for scarce supplies, reports The Guardian. New CEO John Ternus, who took office in September, said on an earnings call with investors that memory costs in the fourth quarter would rise even further above those of the previous three months. The higher prices are expected to weigh on profit margins. The stock subsequently dropped 2.75 percent in after-hours trading. According to Ternus, price hikes for the iPad and MacBook in June and for the iPhone in September offset a portion of the added costs. → The Guardian

Synthszr Take: Apple spared itself the expensive buildout of proprietary AI data centers, yet still ends up footing the AI bill—just via memory. When even the company with the greatest procurement power in the electronics industry faces higher purchasing costs in Q4 than in Q3, data centers set the price for every device. Apple was able to raise prices on the iPhone, iPad, and MacBook. Makers of mid-range Android devices and Windows laptops lack that pricing power: for them, margins shrink first, followed by memory configurations.

AMD: Record Quarter Driven by Data Centers, Slump in Gaming

AMD achieved record revenue of $13.4 billion in the third quarter of 2026, reports The Verge. CEO Lisa Su called it the “largest and most profitable” quarter in company history during the earnings call. The primary growth engine was the data center business at $8.5 billion, accounting for roughly 63 percent of quarterly revenue. Of the $1.9 billion revenue increase, $1.8 billion stemmed from this segment. For the coming quarter, AMD forecasts $17 billion, with Su expecting revenue growth of more than 80 percent for 2027. CFO Jean Hu stated that data center revenue is expected to climb to more than 80 percent of total revenue after 2027. Gaming revenue, by contrast, plunged 40 percent year-over-year to $773 million. → The Verge

Synthszr Take: AMD profits from the very scarcity that is shrinking its gaming business. The same appetite for AI that brings in $1.8 billion in additional data center revenue drives up RAM prices and pushes gaming down 40 percent. For Lisa Su, that is an acceptable trade-off: every wafer and engineering team generates higher margins in the data center. Sony, Microsoft, and PC manufacturers should expect to take a back seat at AMD as long as the crunch lasts. SK Hynix anticipates it will persist until 2030.

Google Launches AI Diary Dreambeans in India

Google has launched its experimental app Dreambeans in India. According to The Hindu, the application developed by Google Labs creates a daily personalized collection of stories rendered in a watercolor aesthetic by accessing connected services like Photos, Gmail, and Calendar. The app has been available in the US for Android and iOS since September. In India, it runs on both platforms across 14 languages, including Hindi, Tamil, Telugu, Kannada, and Malayalam. According to Google, Dreambeans only processes data from apps to which users have explicitly consented. These permissions will reportedly not affect settings in Gemini or AI Mode. → The Hindu

Synthszr Take: Google’s scarcest resource for personal AI is error correction on real private data, and Dreambeans collects it via thumbs-down feedback. India provides a massive testing ground across 14 languages along with regulatory conditions that permit such experiments with emails and calendars quickly. Every mispainted watercolor story refines the context access that Gemini and AI Mode will later monetize through ads. Europe will likely not see the app until it has finished learning.

Developer Compiles a RISC-V Processor into Qwen 3.5 4B Weights

Developer Divyansh Agrawal has compiled a RISC-V processor directly into the weights of the Qwen 3.5 4B language model, as highlighted by Simon Willison in his newsletter. This “Qwenputer” uses the model’s gated DeltaNet layers—embedded in three out of four attention layers—as working memory. According to Agrawal, arithmetic operations as well as C and MicroPython code execute entirely within the forward passes, without any external runtime environment, plugin, or tool call. Technically, he accomplishes this by setting the gate and write strength to 1 and using one-hot keys. This ensures each write operation overwrites exactly one row of state, and each lookup returns an exact value. → Simon Willison from Simon Willison’s Newsletter

Synthszr Take: An off-the-shelf 4B model now contains a complete CPU, and the core work was done by two parameters set to 1. The computational capacity was already sitting idle in layers originally intended solely for managing context. This challenges providers of code sandboxes and tool runtimes whose business relies on calling external environments. If the model computes with precision inside its own forward pass, latency, infrastructure, and a vendor dependency vanish. The bottleneck lies in the minds of model architects—Alibaba had already shipped the necessary layers.

Anthropic Prohibits Cloning — Without Success

Anthropic has updated Claude Code’s system prompt with a rule aimed at preventing the recreation of specific, existing video games. According to Ars Technica, the change was first spotted by a user named “chasm,” who uncovered the rule under the internal designation “tengu_heron_brook.” XDA independently confirmed it in version 2.1.314 of Claude Code, released on October 30. The wording permits original games, genre games, and games built from custom designs and assets, but forbids remakes of existing titles, even if renamed or presented as original. Anthropic developer Addy Osmani confirmed two new sentences in the prompt for Opus 5.5, explaining that copyright infringement violates the company’s terms of service. → Ars Technica

Synthszr Take: “tengu_heron_brook” simply pushes the cloning wave where no one is watching. The retro ports that forced decomp.dev to shut down do not require a frontier model. Mechanical decompilation and rebuilding can be handled by open models running locally, without system prompts and without terms of service. Anthropic loses a user segment while gaining liability protection. For rights holders, the situation arguably worsens: instead of dealing with an entity backed by a legal department, they will face weights running on local hardware that no one can shut down.

Apple’s Home Robot Has Been Built for Two Years, Waiting for Software Until 2028

Apple is delaying its first home robot to 2027 at the earliest, and potentially to 2028. According to former employees, the hardware for the device codenamed J595 has been largely finished for around two years. The software, however, is not. Expected retail names include “HomeView Ultra,” with “HomeView Max” and “HomeView 360” also under discussion. The price is anticipated to be around $1,000, though rising costs for memory and other components could drive it higher.

The project originated in 2019 as a video calling device featuring a robotic arm designed to track a person’s movements in the room with its display. One account reports initial prototypes as early as 2018, approved by then-CEO Tim Cook. Ambitions expanded as Apple leadership sought to pack more AI features into the device. New CEO John Ternus wants Apple to pursue robotics seriously.

The current design features a nine-inch screen on a 360-degree swiveling arm. Visually, it is reminiscent of the iMac G4 or the Pixar lamp. Cameras and sensors are intended to let the device map its surroundings and act autonomously—turning on lights when someone enters the room, reorienting toward the speaker during a FaceTime call, or shooing a cat off the table, all while navigating obstacles without bumping into objects. Apple published a research paper last year exploring a robotic lamp with its own personality.

Internally, there is disagreement over the feature set. Management and engineering teams remain divided on which automation capabilities belong in the first generation and which should wait for a successor. According to insiders, the software team pushed for the delay, attempting to cut planned features and narrow the device’s scope to FaceTime and basic home automation. The autonomous AI capabilities are seen as ambitious and not yet production-ready.

The robot builds on the standard HomeView, which Apple unveiled on October 13, 2026. That seven-inch smart display retails for $299 and goes on sale November 13. The HomeView itself had previously faced delays because the revamped Siri was not ready in time. → The Information, The Verge, Ars Technica, Digital Trends, The Apple Post, iPhone in Canada, Macworld, AppleInsider, Digit, 9to5Mac

Synthszr Take: For two years, the hardware for J595 has sat ready while the software team fights for cuts. The divide lies between a management team asking for ever more AI and engineers who could deliver FaceTime and smart lighting controls today. As long as no one defines what generation one must accomplish, feature creep outpaces maturity. Ternus inherits an unresolved product decision, and every year of delay drives up the cost of memory inside a device that has already been built.

DeepSeek Lets Users Build Plugins for Its Agent via Prompts

DeepSeek has introduced Creator Mode for its agent software DeepSeek Harness. It allows users to build custom plugins simply by describing what they need in a text prompt. The resulting tools are integrated directly into the Harness, effectively reshaping the tool itself. The Chinese lab promotes the feature with the promise of turning ideas “into usable plugins.” Creator Mode is currently in preview. According to DeepSeek, the interfaces may still change.

Technically, DeepSeek separates two tiers: the model decides what to do, while a Harness executes the work. DeepSeek Harness is open source and follows an “everything is a plugin” philosophy. Whether a user needs an additional tool, a new capability, or an interface tweak, the agent builds the extension itself.

A demo video showcases the range of use cases, stretching from simple productivity utilities like a notes app to more sophisticated projects like design consoles and visualizations for language models. To get users started, DeepSeek published seven sample prompts for Creator Mode. While many AI tools already generate complete websites today, Creator Mode is designed to extend the agent platform itself. → The Code, Superhuman – Zain Kahn

Synthszr Take: A plugin generated from a single prompt sentence costs next to nothing to create. That upends the business model of integration consultants whose day rates depend on exactly these types of customizations. DeepSeek gives away the Harness as open source because there is no margin there. Monetization happens through the model that designs every extension and subsequently runs every invocation. Custom solutions become cheap; running them turns into recurring revenue for the lab, built on interfaces DeepSeek is free to change.