Omni Topic - Tech

Binance has promised that users won't pay for the $6.3 billion theft. What backs that promise is not its $1 billion user protection fund but the company's own treasury. Binance says it has replenished 90 percent of the roughly $6.3 billion stolen from its platform in the largest crypto theft on record, and that it will cover the rest. The word doing the work there is "replenished." Binance has not clawed most of the money back from the hackers. It has spent its own money buying replacement tokens on the open market to refill the wallets that were drained. That distinction matters for anyone judging what this incident says about keeping assets on a centralized exchange. By Binance's account, users are being made whole. But the thing protecting them is not a recovery process or an insurance policy. It is the balance sheet of the world's largest exchange. Replenished is not recovered CNBC put it plainly: covering 90 percent of the stolen token assets ...

Google, OpenAI, Anthropic and Microsoft have pledged to support a new common signal in robots.txt that lets sites block scraping for AI training. Until now, each company ran its own crawler with its own opt-out name, so publishers had to manage a growing list of rules. A shared signal could replace that list with a single instruction, although the technical details have not yet been disclosed. The patchwork this is meant to replace robots.txt is a plain-text file that tells automated crawlers which parts of a site they may fetch. The convention dates to the 1990s and was formalized by the IETF as RFC 9309 in 2022. It was designed for search indexing, not for the question of whether content may be used to train a generative model. When AI training scraping became a flashpoint in 2023, companies responded one by one. OpenAI published GPTBot and told site owners how to block it. Google introduced a Google-Extended token that controls whether content is used for its Gemini models and ...

Intel's share price has dropped sharply, taking roughly $12 billion off the value of the stake the U.S. government holds in the company. The pressure point is the PC business. The AI industry is absorbing memory and other components at scale, which pushes PC prices higher and weighs on unit sales. Intel's latest slide is a useful case study in how the AI build-out spills over into parts of tech that are not directly part of it. The chipmaker is not struggling because of an AI product failure. It is being squeezed by what AI demand is doing to the cost of the PCs its processors go into. Why the government's stake makes this a bigger story In August 2025, the U.S. government took a stake of roughly 10% in Intel, converting federal chip-subsidy funding into equity. That decision turned Washington into one of the company's largest shareholders. As a result, every large move in Intel's share price now shows up as a gain or loss on a public balance sheet. The cu...

Reddit CEO Steve Huffman said on the company's Q3 2026 earnings call that Google's AI-powered search modes do not send meaningful traffic to Reddit. Search referrals in October fell by more than 1 million on average. The comment adds a major platform's voice to a complaint publishers have been making for more than a year. What Huffman said Huffman told investors that Google's AI search modes, including AI Mode, aren't helping Reddit's traffic. Reddit's own numbers back that up: October search referrals dropped by more than 1 million on average. The exact per-period breakdown hasn't been widely detailed. The remark is notable less for the number than for who said it and where. A public-company CEO raised it during an earnings call, which means investors are now pricing it in. Why Reddit is a special case Few sites benefited from Google's search results as visibly as Reddit did. Users got into the habit of appending 'reddit' to...

Microsoft has disclosed Azure's quarterly revenue in dollars for the first time: $33.2 billion. Until now the company only reported growth percentages, which left the real size gap with Amazon Web Services (AWS) open to estimates. The new number lets analysts compare the two largest cloud providers on the same scale. Microsoft has disclosed Azure's quarterly revenue in dollars for the first time: $33.2 billion. Until now the company reported only growth rates, so the real size gap with Amazon Web Services (AWS) was a matter of estimates. The number puts Azure within reach of the market leader. What Microsoft actually changed For years, Microsoft's earnings reports gave Azure a percentage growth figure and little else. Investors knew Azure was growing fast, but not how large it was in absolute terms. The company did give one annual data point last July, saying Azure passed $75 billion in revenue for fiscal 2025. A quarterly dollar figure is a different kind of disclo...

The AI data center boom is straining the US power grid, and industry insiders say the strain is also opening the door to grid modernization. Utilities and developers are being pushed to adopt upgrades that have been discussed for years but rarely funded at scale. Here is what that modernization covers, why data centers change the economics, and where the risks remain. A Grid Built for a Different Era Much of the US grid was designed around slow, predictable demand growth. Power plants and transmission lines take years to permit and build. AI data centers, by contrast, can be planned and constructed far faster and then draw large, steady loads around the clock. That mismatch is the core of the problem. Requests to connect new load and new generation pile up, and local grids can run short of capacity before new infrastructure arrives. What Modernization Actually Covers The phrase sounds abstract, but it points to a fairly concrete set of tools: Grid-enhancing technologies , such...

Anthropic's S-1 filing with the US Securities and Exchange Commission, submitted Monday as the company prepares to go public, acknowledges that data centers are increasingly reliant on natural gas. It also says competition for electrical supply and for sites with enough power is intense. Because an S-1 is a legal disclosure to investors, the energy problem behind generative AI is now written down by a leading AI lab itself. Anthropic's S-1 filing with the US Securities and Exchange Commission, submitted Monday as the company prepares to go public, acknowledges that data centers are increasingly reliant on natural gas. It also says competition for electrical supply and for data center sites with adequate power is intense. An S-1 is a legal disclosure to investors, so this is not marketing copy. It is a risk statement from a leading AI lab, and it puts the energy problem behind generative AI on the record. What the filing actually says The language sits in the part of the ...

iFixit has taken apart Apple's iPhone Duo, and the teardown offers a look at the hardware beneath the folding screen. Two findings stand out: how Apple appears to be minimizing the visible crease, and a 5/10 repairability score. The internal view also adds context to reports of broken screens and hinges. What iFixit found inside the iPhone Duo As it does with most major phone launches, iFixit opened the iPhone Duo and laid the components out for inspection. The teardown covers how the folding display is built and how the hinge is integrated. It ends with a repairability score of 5 out of 10. That score comes from iFixit's standard scale, where 10 is the easiest to repair. Five sits exactly in the middle. Why the crease is a design problem, not a cosmetic one Every foldable display has a crease because its layered panel bends at the same spot thousands of times. How visible that line is depends on the layer stack and on the shape the hinge forces the panel to take as it ...

Cerebras has described how it serves trillion-parameter language models on its own hardware. Instead of fitting a model on one chip, it spreads the weights across the on-chip memory of several WSE-3T wafers and runs them as a pipeline. The company reports speeds of roughly 1,000 to 2,000+ tokens per second per request. Kimi K3, a model of about 3 trillion parameters, serves as the headline example. What Cerebras is actually doing Cerebras keeps model weights in on-chip memory on the wafer rather than in external high-bandwidth memory. That is the core of its speed advantage, because token generation is usually limited by how fast weights can be read. The catch is capacity: a frontier-scale model will not fit on one wafer. The answer is to shard the model across multiple WSE-3T wafers and chain them in a pipeline. Each wafer holds a slice of the network, and tokens flow from one wafer to the next. With many requests in flight at once, every wafer stays busy. The Kimi K3 example K...

CNET has loaded Apple's iPhone Duo into a custom-built foldable test machine that opens and closes the phone 900 times an hour. The goal is to see how the hinge holds up, with Samsung's Galaxy Z Fold 8 as the benchmark. Results have not been published yet, but the setup already says a lot about how foldables get judged. CNET has put Apple's iPhone Duo into a custom-made foldable phone test machine. The rig folds the phone 900 times an hour, and the question is whether the hinge survives and how it compares with the Galaxy Z Fold 8. No results have been released yet. What 900 folds an hour actually means The number sounds abstract, so here is the arithmetic. At 900 folds an hour, a machine running around the clock logs 21,600 folds a day. Samsung has cited 200,000-fold durability testing for earlier Galaxy Z Fold models, and reaching that figure on this rig would take roughly nine days of nonstop folding. For a human user, 200,000 folds is a lot. Someone who opens an...

A new Texas-based defense startup called Warhound plans to genetically modify military and police dogs and outfit them with technology from Anduril. It was started by Anduril founder Palmer Luckey and Ben Lamm, co-founder of Colossal Biosciences, the company known for its de-extinction projects. The pairing puts a defense-tech company and a gene-editing company behind the same animal-focused venture. Many details, including which traits will be edited and what hardware the dogs will carry, have not been made public. What Warhound says it will do Warhound is positioned as a defense startup focused on working dogs in the United States, both military and police. The pitch has two parts: change the animals genetically, and equip them with military technology developed by Anduril. That is a bigger claim than it first sounds. Most defense startups sell software, sensors, drones or munitions. Warhound proposes to treat a living animal as the platform. The two people behind it Palmer Lu...

OpenAI is reportedly working on a new $30 billion funding round at a $1.25 trillion valuation, ahead of a planned IPO. If the round closes at anything above $1 trillion, OpenAI would become the most highly valued private company ever. Investors and final terms have not been disclosed. OpenAI is working on a new raise of about $30 billion at a valuation of roughly $1.25 trillion, according to TechCrunch. Should the round hold above the $1 trillion mark, it would set a record for a privately held company. Details such as who is participating and when the deal could close have not been made public. How fast the price has moved The jump is easier to grasp next to OpenAI's recent marks. In March 2025 the company was valued at about $300 billion in a funding round. In October 2025, an employee share sale priced it at $500 billion. A $1.25 trillion valuation would be 2.5 times that October figure, in roughly a year. Few companies at any stage, public or private, have been repriced t...

Tesla says it wants to start selling the Cybercab to other transport companies before the end of the year. That would move the company beyond running its own robotaxi service and toward supplying vehicles to third-party fleets. The details that matter most, including buyers, pricing and delivery timing, have not been disclosed. Tesla wants to sell the Cybercab, its purpose-built two-seat robotaxi, to other transport companies by the end of 2026. If it happens, Tesla would stop being only a robotaxi operator and become a vehicle supplier to rival fleets, a business model closer to a traditional carmaker. What the Cybercab is Tesla unveiled the Cybercab in October 2024 at its We, Robot event. It is a two-seat vehicle with no steering wheel and no pedals, designed to be driven entirely by software. Elon Musk said at the time that he was aiming for a price below $30,000 and for production in 2026. In June 2025 Tesla launched a limited robotaxi service in Austin, Texas, using modified...

A surge of new data centers in Texas could push air pollution back to levels not seen since the 2000s. Texas has become ground zero for the physical infrastructure behind generative AI, and new and proposed facilities want enormous amounts of electricity. The air quality outcome depends less on the buildings than on how developers and utilities generate that power. A rush of new data centers could bring air pollution roaring back in Texas, according to a report covered by The Verge. The state has become a center for the physical build-out behind generative AI, and the electricity those facilities need is the pressure point. The problem is the power, not the building A data center does not burn fuel the way a factory does. Servers and cooling systems draw electricity, and that electricity has to come from somewhere. If developers and utilities cover the new demand with gas-fired generation, the pollutants from those plants and turbines go up with it. The report's warning is th...

OpenAI has put money into a startup that uses brain implants to detect cancer, a use of neural interface hardware that goes well beyond the usual pitch of restoring movement or communication. It adds to a pattern of AI companies placing bets on brain data. The deal says less about a finished product than about where AI money is heading next. What is actually new here Brain implants have mostly been associated with helping people with paralysis control computers or devices. Neuralink put its first implant in a human in early 2024, and companies such as Synchron and Precision Neuroscience have worked on their own approaches. Using this class of hardware to detect cancer is a different proposition. How the startup's system would turn neural signals into a cancer readout has not been validated in public clinical data. That gap matters. A device that reads tumors in the brain is a very different thing from one that tries to infer disease elsewhere in the body from neural activity, ...

Google is bringing Gemini's Personal Intelligence feature to cars, so the assistant can pull from Gmail, Calendar, Maps, YouTube and other Google services while you drive. The change moves in-car voice assistants away from one-off commands and toward answers shaped by your own accounts. Details such as supported vehicles and rollout timing have not been spelled out yet. What Personal Intelligence Adds to the Driving Seat\n\nPersonal Intelligence is a Gemini feature that lets the assistant use information from a user's connected Google services to tailor its answers. It first appeared in the phone-based Gemini experience. Extending it to cars means the same personalization now applies in a setting where people rely almost entirely on voice.\n\nThe services named in the announcement are Gmail, Calendar, Maps and YouTube, along with other Google products. Together they cover where you need to be, what is on your schedule, what has landed in your inbox, and what you watch or li...

Samsung is building a massive semiconductor manufacturing complex in Pyeongtaek, South Korea, aimed at the memory chips that generative AI hardware depends on. Three fabs there already run around the clock, and a fourth is coming online. The reason is simple: modern AI GPUs are only as useful as the stacked memory sitting next to them. Samsung is expanding its Pyeongtaek campus in South Korea, and the scale of the site says a lot about where the chip industry's bottleneck has moved. Three fabs already operate around the clock. A fourth is coming online. Why a GPU needs a memory factory behind it Generative AI models run on GPUs, but a GPU does not work alone. It is packaged with stacks of specialized memory chips, known as high-bandwidth memory (HBM), that feed it data fast enough to keep the processor busy. HBM is built by stacking several DRAM dies vertically. It is generally considered harder to manufacture than conventional DRAM and consumes more wafer capacity per unit o...

Meta has adjusted its new terms of service after users pushed back over language related to scraping. The changes had alarmed publishers and other people running business accounts, who treat their posts as commercial assets rather than casual content. The revision is a small edit, but it shows how much weight a single clause in a platform contract can carry. What happened Meta revised its newly introduced terms of service after criticism that the wording could be read as permitting broad scraping of content. Publishers and holders of business accounts were among the most concerned. The full text of every change has not been laid out in detail publicly, so this article sticks to what has been confirmed: the terms were tweaked, and the backlash came largely from commercial users. How this fits an ongoing pattern This is not the first adjustment to these terms. In an earlier post, Meta Scraping and AI Training Terms , I covered Meta's decision to remove language that had raised ...

DeepSeek has added a Search API and a Fetch API so agents and RAG applications can pull live web content. Search returns web results, and Fetch returns the body of a single page as Markdown. Both cost $0.75 per 1,000 successful requests, and a request only counts as successful if it returns HTTP 200 with an actual result. DeepSeek has added a Search API and a Fetch API so agents and RAG applications can pull live web content. Search returns web results, and Fetch returns the body of a single page as Markdown. Both cost $0.75 per 1,000 successful requests, and a request only counts as successful if it returns HTTP 200 with an actual result. What each endpoint does The two APIs split a common agent workflow into two steps. Search finds candidate pages for a query. Fetch takes one URL and returns the page's main content in Markdown. That matters because language models have a training cutoff. Anything that happened after it, such as a new library release, a price change or a new...

Meta has removed language from its updated terms that had raised concerns about web scraping and AI training. A Meta spokesperson insisted the revision does not reflect any change in how Meta crawls the web. That leaves a practical gap between what a legal document says and what a crawler actually does. For publishers and developers, the second one is what counts. What happened Meta edited its terms and took out wording that readers had interpreted as covering scraping and AI training. After the concern spread, a company spokesperson said the updated terms do not signal a change in Meta's web-crawling practices. The specifics of any future policy change have not been disclosed. What is on the record is a deletion and a denial that anything operational has shifted. Terms of service and crawler behavior are different things Terms of service describe what a company claims the right to do. A crawler's behavior is set by code, user-agent strings and the rules a site publishe...