Mistral has launched a new multimodal OCR API that converts complex PDF documents into AI-friendly Markdown files. The API is designed for efficiency, handles visual elements like illustrations, supports complex formatting such as mathematical expressions, and reportedly outperforms similar offerings from major competitors. TechCrunch reports: Unlike most OCR APIs, Mistral OCR is a multimodal API, meaning that it can detect when there are illustrations and photos intertwined with blocks of text. The OCR API creates bounding boxes around these graphical elements and includes them in the output. Mistral OCR also doesn't just output a big wall of text; the output is formatted in Markdown, a formatting syntax that developers use to add links, headers, and other formatting elements to a plain text file.

Mistral OCR is available on Mistral's own API platform or through its cloud partners (AWS, Azure, Google Cloud Vertex, etc.). And for companies working with classified or sensitive data, Mistral offers on-premise deployment. According to the Paris-based AI company, Mistral OCR performs better than APIs from Google, Microsoft, and OpenAI. The company has tested its OCR model with complex documents that include mathematical expressions (LaTeX formatting), advanced layouts, or tables. It is also supposed to perform better with non-English documents. [...]

Mistral is also using Mistral OCR for its own AI assistant Le Chat. When a user uploads a PDF file, the company uses Mistral OCR in the background to understand what's in the document before processing the text. Companies and developers will most likely use Mistral OCR with a RAG (aka Retrieval-Augmented Generation) system to use multimodal documents as input in an LLM. And there are many potential use cases. For instance, we could envisage law firms using it to help them swiftly plough through huge volumes of documents. "Over the years, organizations have accumulated numerous documents, often in PDF or slide formats, which are inaccessible to LLMs, particularly RAG systems. With Mistral OCR, our customers can now convert rich and complex documents into readable content in all languages," said Mistral co-founder and chief science officer Guillaume Lample.

"This is a crucial step toward the widespread adoption of AI assistants in companies that need to simplify access to their vast internal documentation," he added.

Google Gboard for Android is introducing circle or pill-shaped keys for some beta testers today. "Instead of the key borders being rounded rectangles, Gboard is switching to circles and pills for letters, while the space bar and other keys are now pill-shaped," reports 9to5Google. "While there should be no functional change to touch targets, these new shapes really shift the look of Gboard for Android." From the report: On paper, it's a bit more modern (and rounded) compared to what came before. However, it's a bit cramped if you have "Long press for symbols" enabled, which goes from the top-right corner to being directly above the letter. The physical analog Gboard is moving away from is how most keys on laptops and desktops are square.

An anonymous reader quotes a report from Reuters: The U.S. House Judiciary Committee subpoenaed Alphabet on Thursday seeking its communications with former President Joe Biden's administration about content moderation policies. House Judiciary Committee Chairman Jim Jordan, a Republican, also asked the YouTube parent company for similar communications with companies and groups outside government, according to a copy of the subpoena seen by Reuters. The subpoena seeks communications about limits or bans on content about President Donald Trump, Tesla CEO and close Trump ally Elon Musk, the virus that causes COVID-19 and a host of other conservative discussion topics. "Alphabet, to our knowledge, has not similarly disavowed the Biden-Harris Administration's attempts to censor speech," Jordan said in a letter.

Meanwhile, Google spokesperson Jose Castaneda said the company will "continue to show the committee how we enforce our policies independently, rooted in our commitment to free expression."

In a policy paper, former Google CEO Eric Schmidt, Scale AI CEO Alexandr Wang, and Center for AI Safety Director Dan Hendrycks said that the U.S. should not pursue a Manhattan Project-style push to develop AI systems with "superhuman" intelligence, also known as AGI. From a report: The paper, titled "Superintelligence Strategy," asserts that an aggressive bid by the U.S. to exclusively control superintelligent AI systems could prompt fierce retaliation from China, potentially in the form of a cyberattack, which could destabilize international relations.

"[A] Manhattan Project [for AGI] assumes that rivals will acquiesce to an enduring imbalance or omnicide rather than move to prevent it," the co-authors write. "What begins as a push for a superweapon and global control risks prompting hostile countermeasures and escalating tensions, thereby undermining the very stability the strategy purports to secure."

Google announced Wednesday it is expanding its AI Overviews to more query types and users worldwide, including those not logged into Google accounts, while introducing a new "AI Mode" chatbot feature. AI Mode, which resembles competitors like Perplexity or ChatGPT Search, will initially be limited to Google One AI Premium subscribers who enable it through the Labs section of Search.

The feature delivers AI-generated answers with supporting links interspersed throughout, powered by Google's search index. "What we're finding from people who are using AI Overviews is that they're really bringing different kinds of questions to Google," said Robby Stein, VP of product on the Search team. "They're more complex questions, that may have been a little bit harder before." Google is also upgrading AI Overviews with its Gemini 2.0 model, which Stein says will improve responses for math, coding and reasoning-based queries.

Google is urging officials at President Donald Trump's Justice Department to back away from a push to break up the search engine company, citing national security concerns, Bloomberg reported Wednesday, citing sources familiar with the discussions. From the report: Representatives for the Alphabet unit asked the government in a meeting last week to take a less aggressive stance as the US looks to end what a judge ruled to be an illegal online search monopoly, said the people, who asked not to be identified discussing the private deliberations. The Biden administration in November had called for Google to sell its Chrome web browser and make other changes to its business including an end to billions of dollars in exclusivity payments to companies including Apple.

Although Google has previously pushed back on the Biden-era plan, the recent discussions may preview aspects of the company's approach to the case as it continues under the Trump administration. A federal judge is set to rule on how Google must change its practices following hearings scheduled for next month. Both sides are due to file their final proposals to the judge on Friday.

Reinforcement learning pioneers Andrew Barto and Richard Sutton have warned against the unsafe deployment of AI systems [alternative source] after winning computing's prestigious $1 million Turing Award Wednesday. "Releasing software to millions of people without safeguards is not good engineering practice," said Barto, professor emeritus at the University of Massachusetts, comparing it to testing a bridge by having people use it.

Barto and Sutton developed reinforcement learning in the 1980s, inspired by psychological studies of human learning. The technique, which rewards AI systems for desired behaviors, has become fundamental to advances at OpenAI and Google. Sutton, a University of Alberta professor and former DeepMind researcher, dismissed tech companies' artificial general intelligence narrative as "hype."

Both laureates also criticized President Trump's proposed cuts to federal research funding, with Barto calling it "wrong and a tragedy" that would eliminate opportunities for exploratory research like their early work.

An anonymous reader quotes a report from TechCrunch: Google has open sourced an AI model, SpeciesNet, designed to identify animal species by analyzing photos from camera traps. Researchers around the world use camera traps -- digital cameras connected to infrared sensors -- to study wildlife populations. But while these traps can provide valuable insights, they generate massive volumes of data that take days to weeks to sift through. In a bid to help, Google launched Wildlife Insights, an initiative of the company's Google Earth Outreach philanthropy program, around six years ago. Wildlife Insights provides a platform where researchers can share, identify, and analyze wildlife images online, collaborating to speed up camera trap data analysis.

Many of Wildlife Insights' analysis tools are powered by SpeciesNet, which Google claims was trained on over 65 million publicly available images and images from organizations like the Smithsonian Conservation Biology Institute, the Wildlife Conservation Society, the North Carolina Museum of Natural Sciences, and the Zoological Society of London. Google says that SpeciesNet can classify images into one of more than 2,000 labels, covering animal species, taxa like "mammalian" or "Felidae," and non-animal objects (e.g. "vehicle"). SpeciesNet is available on GitHub under an Apache 2.0 license, meaning it can be used commercially largely sans restrictions.

Google Play on Android devices is being updated to include a new search filter for widgets, widget badges on app detail pages, and a curated editorial page dedicated to widgets. The Verge reports: With the search filter, users will be able to more easily search for apps with widgets. The badge "eliminates guesswork for users and highlights your widget offerings, encouraging them to explore and utilize this capability," Taiwo-Peters says. And the curated editorial page will show off "collections of excellent widgets." The updated widget discoverability tools will be "coming soon," Taiwo-Peters says. "Historically, one of the challenges with investing in widget development has been discoverability and user understanding," product manager Yinka Taiwo-Peters says in the post. "You've asked for better ways for users to find and utilize your widgets, and we're delivering." Yinka Taiwo-Peters also acknowledges that "we understand that the effort required to build and maintain widgets needs to be justified by user adoption."

Memory safety bugs are "eroding trust in technology and costing billions," argues a new post on Google's security blog — adding that "traditional approaches, like code auditing, fuzzing, and exploit mitigations — while helpful — haven't been enough to stem the tide."

So the blog post calls for a "common framework" for "defining specific, measurable criteria for achieving different levels of memory safety assurance." The hope is this gives policy makers "the technical foundation to craft effective policy initiatives and incentives promoting memory safety" leading to "a market in which vendors are incentivized to invest in memory safety." ("Customers will be empowered to recognize, demand, and reward safety.")

In January the same Google security researchers helped co-write an article noting there are now strong memory-safety "research technologies" that are sufficiently mature: memory-safe languages (including "safer language subsets like Safe Buffers for C++"), mathematically rigorous formal verification, software compartmentalization, and hardware and software protections. (With hardware protections including things like ARM's Memory Tagging Extension and the (Capability Hardware Enhanced RISC Instructions, or "CHERI", architecture.) Google's security researchers are now calling for "a blueprint for a memory-safe future" — though Importantly, the idea is "defining the desired outcomes rather than locking ourselves into specific technologies."

Their blog post this week again urges a practical/actionable framework that's commonly understood, but one that supports different approaches (and allowing tailoring to specific needs) while enabling objective assessment: At Google, we're not just advocating for standardization and a memory-safe future, we're actively working to build it. We are collaborating with industry and academic partners to develop potential standards, and our joint authorship of the recent CACM call-to-action marks an important first step in this process... This commitment is also reflected in our internal efforts. We are prioritizing memory-safe languages, and have already seen significant reductions in vulnerabilities by adopting languages like Rust in combination with existing, wide-spread usage of Java, Kotlin, and Go where performance constraints permit. We recognize that a complete transition to those languages will take time. That's why we're also investing in techniques to improve the safety of our existing C++ codebase by design, such as deploying hardened libc++.

This effort isn't about picking winners or dictating solutions. It's about creating a level playing field, empowering informed decision-making, and driving a virtuous cycle of security improvement... The journey towards memory safety requires a collective commitment to standardization. We need to build a future where memory safety is not an afterthought but a foundational principle, a future where the next generation inherits a digital world that is secure by design.
The security researchers' post calls for "a collective commitment" to eliminate memory-safety bugs, "anchored on secure-by-design practices..." One of the blog post's subheadings? "Let's build a memory-safe future together."

And they're urging changes "not just for ourselves but for the generations that follow."

Alphabet's X division has developed a silicon photonic chip for its Taara project, which transmits internet via laser beams instead of fiber optic cables. The system delivers 20Gbps through "light bridges" that establish line-of-sight connections between transceiver units. The second-generation technology miniaturizes previous mechanical components -- including gimbals, mirrors, and lenses -- into solid-state circuitry the size of a fingernail.

This chip enables a single laser transmitter to potentially pair with multiple receptors, significantly reducing costs from the current ~$30,000 per bridge setup. Taara has already demonstrated real-world viability by connecting Brazzaville and Kinshasa across the Congo River, providing the latter with five-fold cheaper internet access, and supplementing bandwidth at Coachella 2024. Project leader Mahesh Krishnaswamy claims Taara can deliver "10, if not 100 times more bandwidth" than Starlink in dense areas. X's Astro Teller suggests this technology could form the foundation for 7G networks as radio frequency bands become increasingly congested. Taara will soon "graduate" from X and seek external funding, with Alphabet maintaining a significant stake.

Further reading: Official blog post.

Microsoft Edge is following Chrome's lead by disabling uBlock Origin and other Manifest V2-based extensions in its browser. Neowin reports: The latest Edge Canary version started disabling Manifest V2-based extensions with the following message: "This extension is no longer supported. Microsoft Edge recommends that you remove it." Although the browser turns off old extensions without asking, you can still make them work by clicking "Manage extension" and toggling it back (you will have to acknowledge another prompt).

Google started phasing out Manifest V2 extensions in June 2024, and it has a clear roadmap for the process. Microsoft's documentation, however, still says "TBD," so the exact dates are not known yet. This leads to some speculating about the situation being one of "unexpected changes" coming from Chromium. Either way, sooner or later, Microsoft will ditch MV2-based extensions, so get ready as we wait for Microsoft to shine some light on its plans.

Another thing worth noting is that the change does not appear to be affecting Edge's stable release or Beta/Dev Channels. For now, only Canary versions disable uBlock Origin and other MV2 extensions, leaving users a way to toggle them back on. Also, the uBlock Origin is still available in the Edge Add-ons store, which recently received a big update.

Google co-founder Sergey Brin has urged employees working on the company's Gemini AI products to be in the office "at least every weekday" [non-paywalled source] and suggested "60 hours a week is the sweet spot of productivity," according to an internal memo cited by The New York Times. The directive comes as Brin warned that "competition has accelerated immensely and the final race to A.G.I. is afoot," referring to artificial general intelligence, when machines match or surpass human intelligence.

"I think we have all the ingredients to win this race, but we are going to have to turbocharge our efforts," Brin wrote in the Wednesday evening memo. The guidance does not alter Google's official policy requiring employees to work in-office three days weekly. Brin, who returned to Google following ChatGPT's 2022 launch, also criticized staff who "put in the bare minimum," calling them "highly demoralizing to everyone else."

Google DeepMind CEO Demis Hassabis says the definition of artificial general intelligence is being "watered down," creating an illusion of faster progress toward this technological milestone. "There's quite a long way, in my view, before we get to AGI," Hassabis said. "The timelines are shrinking because the definition of AGI is being watered down, in my opinion." DeepMind defines AGI as "AI systems that are at least as capable as humans at most cognitive tasks," while OpenAI has historically described it as a "highly autonomous system that outperforms humans at most economically valuable work."

OpenAI CEO Sam Altman recently declared his team is "confident we know how to build AGI," while modifying his personal definition to an AI "system that can tackle increasingly complex problems, at human level, in many fields." Hassabis suggested industry hype might be financially motivated: "There is a lot of hype for various reasons," he said, including perhaps "that people need to raise money." Microsoft CEO Satya Nadella separately dismissed AGI milestones as "nonsensical benchmark hacking," preferring economic impact measurements.

Google changed its rules around how product-review sites appear in its search engine. In the process, it devastated a once-lucrative corner [non-paywalled source] of the news media world. From a report: Sites including CNN Underscored and Forbes Vetted offer tips on everything from mattresses and knife sets to savings accounts, making money when users click on links and buy products.

They depend on Google to drive much of their traffic, and therefore revenue. But over the past year, Google created stricter rules that dinged certain sites that farm out articles to freelancers, among other things. The goal, Google has said, was to give users higher-quality search results. The outcome was a crisis for some sites. Traffic for Forbes Advisor, a personal-finance recommendation site, fell 83% in January from the same month the year before, according to data firm Similarweb.

CNN Underscored and Buy Side from WSJ, which is operated by Wall Street Journal parent Dow Jones, were both down by more than 25% in that period. Time magazine's Time Stamped and the Associated Press's AP Buyline, powered by Taboola Turnkey Commerce, ended their efforts in recent months. Taboola closed the commerce operation.

Apple announced a new feature allowing parents to share a child's age with app developers without exposing sensitive information, as lawmakers debate age-verification laws for social media and apps. Reuters reports: States, such as Utah and South Carolina, are currently debating laws that would require app store operators such as Apple and Alphabet's Google to check the ages of users. That has set up a conflict in the tech industry over which party should be responsible for checking ages for users under 18 -- app stores, or each individual app. Meta, for instance, has long argued in favor of legislation requiring app stores to check ages when a child downloads an app.

Apple on Thursday said it does not want to be responsible for collecting sensitive data for those age verifications. "While only a fraction of apps on the App Store may require age verification, all users would have to hand over their sensitive personally identifying information to us -- regardless of whether they actually want to use one of these limited set of apps," Apple wrote in a whitepaper on its website.

An anonymous reader quotes a report from TechCrunch: Security researchers are warning that data exposed to the internet, even for a moment, can linger in online generative AI chatbots like Microsoft Copilot long after the data is made private. Thousands of once-public GitHub repositories from some of the world's biggest companies are affected, including Microsoft's, according to new findings from Lasso, an Israeli cybersecurity company focused on emerging generative AI threats.

Lasso co-founder Ophir Dror told TechCrunch that the company found content from its own GitHub repository appearing in Copilot because it had been indexed and cached by Microsoft's Bing search engine. Dror said the repository, which had been mistakenly made public for a brief period, had since been set to private, and accessing it on GitHub returned a "page not found" error. "On Copilot, surprisingly enough, we found one of our own private repositories," said Dror. "If I was to browse the web, I wouldn't see this data. But anyone in the world could ask Copilot the right question and get this data."

After it realized that any data on GitHub, even briefly, could be potentially exposed by tools like Copilot, Lasso investigated further. Lasso extracted a list of repositories that were public at any point in 2024 and identified the repositories that had since been deleted or set to private. Using Bing's caching mechanism, the company found more than 20,000 since-private GitHub repositories still had data accessible through Copilot, affecting more than 16,000 organizations. Lasso told TechCrunch ahead of publishing its research that affected organizations include Amazon Web Services, Google, IBM, PayPal, Tencent, and Microsoft. [...] For some affected companies, Copilot could be prompted to return confidential GitHub archives that contain intellectual property, sensitive corporate data, access keys, and tokens, the company said.

New York City's Metropolitan Transportation Authority and Google have successfully tested technology that uses smartphone sensors to detect subway track defects, the MTA said Thursday. The four-month experiment, dubbed TrackInspect, mounted six Google Pixel phones on four A train subway cars traversing Manhattan and Queens. The phones' accelerometers, magnetometers, gyroscopes and external microphones collected 335 million sensor readings and 1,200 hours of audio data, which were processed through 200 prediction models.

The system identified 92% of defects later confirmed by human inspectors, including broken rails and loose bolts. "The goal with this [project] is to find issues before they become a major issue in terms of service," said Demetrius Crichlow, the agency's president. Following the successful trial, the MTA plans to expand to a full pilot where Google will build a production version for track inspectors.

Google has received FDA clearance for the Pixel Watch 3's Loss of Pulse Detection feature, which will start rolling out to U.S. devices around the end of March. The Verge reports: The Loss of Pulse Detection feature is exactly what it sounds like: if the Pixel Watch 3 senses that you've lost your pulse through an event like a heart attack or an overdose, it'll send you a prompt. If you don't respond, it'll automatically call emergency services on your behalf. Back in August, Sandeep Waraich, Google's senior director of product manager for Pixel wearables, told The Verge that the Pixel Watch 3 is capable of differentiating between a genuine loss-of-pulse event and a person simply taking the watch off.

Google has updated its Results About You tool with a redesigned hub, easier removal requests directly from Search, and the ability to refresh outdated results. Engadget reports: Today, the tech giant is announcing the latest changes, including a redesigned hub and the ability to update outdated search results to reflect the latest changes.

The redesign isn't only for show. You can now submit removal requests directly from Search with fewer actions by clicking or tapping the three dots beside a search result. If you manage to have content about you deleted or changed from a website but Google Search hasn't caught up, you can refresh the search, which will "recrawl the page and obtain the latest information." In other words, you can always see the most up-to-date results about you.