Actualités — 2026-10-20

Spotify says it has removed more than 50 million tracks it classifies as AI-generated "slop and spam" from its recommendations so far in 2026. In a blog post Tuesday, the streaming company said it has changed its recommendation system to promote more music from human artists and that its editors no longer include such tracks in curated playlists. Spotify has not updated its earlier figure of more than 75 million spam tracks removed from the platform, and it has not offered users a way to block AI-generated music entirely. The update comes as the company continues to add generative AI features of its own and as rival services take harder stances on AI audio. Lire la suite

DeepSeek has released DeepInfer, an open-source inference framework that the company says powers its large-scale production services. The initial release supports DeepSeek-V4.1-Flash and runs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. It is licensed under the Apache License 2.0 and is available on GitHub. The release covers offline inference and includes optimized kernels, parallelism options, speculative decoding and structured outputs. Developers still need to supply several components, including a KV cache system and load balancing, to run a complete inference service. Lire la suite

A developer who spent ten months building a chess bot says it plays more like a real person than the leading human-like models, and has published his own benchmarks to back that up. The bot, Chessbook-1, powers the practice games on Chessbook, where players can face an opponent meant to match their rating and time control. Its creator says it predicts human moves about as well as the best available model, while matching human playing strength and blunder rates more closely across all rating levels. The results come from the developer's own tests, which have not been independently verified. Lire la suite

Speculative decoding makes large language models generate text faster, but not because it requires less computing. A common assumption holds that checking a draft takes fewer operations than generating tokens one at a time. A technical explainer says that is wrong. The method actually performs more floating point operations for the same output. The speedup comes from shifting the GPU from waiting on memory to doing useful math, a change that matters most when a system is serving only a few sequences at once. Lire la suite

Microsoft is adding new controls to its sovereign cloud for governments and regulated companies, giving customers more say over who can reach their systems and letting them run AI tools on their own hardware. CEO Satya Nadella announced the changes in Paris. For customers in the European Union and the European Free Trade Association, a trusted European partner will oversee Microsoft engineers' remote logins to production systems. The moves follow rising European concern over dependence on U.S. cloud companies. Most of the new tools run on Azure Local, and several will be available at the company's Ignite conference in November. Lire la suite

Steam shoppers are asking Valve for a way to hide games that use AI-generated content as Steam Next Fest, the festival for indie games and demos, opened with more low-effort, AI-made titles than usual. One user wrote on the Steam subreddit that he wants to filter out games with AI-generated content, which Steam already requires developers to disclose. Other users say the flood of so-called “AI slop” is making it harder to find human-made games, and the complaints come as AI use in game development grows across the industry. Lire la suite

AI meeting notetaker Otter.ai has raised $320 million in growth financing from General Catalyst’s Customer Value Fund, the company announced Tuesday. The company plans to use the money to grow its enterprise business, expand its go-to-market strategy and international footprint, and develop new products. Otter says more than 40 million people use its product, including teams at more than 86% of the Fortune 500. Chief executive Sam Liang said the funding will support enterprise agentic workflows, a push toward more advanced AI for teams. Lire la suite

Billionaire investor Peter Thiel has delivered a series of private lectures on the Antichrist, the latest in San Francisco and, in March, Rome. Recordings leaked to The Verge show Thiel casting effective altruism and AI safety advocates as the forces that could usher in the end times. His views also draw on a July essay in the conservative journal First Things, in which he suggests the late Pope Francis could have been the Antichrist. Analysts say Thiel's apocalyptic framing is less original than he claims, but his influence makes it worth watching. Lire la suite

Jooba, a recruiting company that uses AI to find job candidates, has launched generally and raised $4 million in seed funding from a16z Speedrun, Scribble Ventures and Genius Ventures. The company builds a custom AI sourcing agent for each open role and pairs it with human talent engineers who review candidates before outreach. Instead of charging large placement fees, Jooba sells the service through a flat subscription. The company says the approach cuts total time to hire by 50% and recruiting fees by 90%, and that its first candidate conversations can begin within 22 hours of an onboarding call. The model targets fast-growing companies that want help hiring without managing another software tool. Lire la suite

Respan, a Y Combinator-backed company, has released an open-source AI gateway written in Rust that is built almost entirely from plugins. The company says it redesigned the gateway after years of patching LiteLLM, a popular open-source project, became unsustainable. The gateway uses a small core that routes requests and lets developers add providers, protocols and features without changing that core. Respan has released the code under the MIT or Apache-2.0 license and says it is still pre-alpha. The company makes its money from a hosted version that includes observability, evaluations and unified billing. Lire la suite

Yelp is rolling out a feature that lets app users have an AI voice agent call a restaurant to check wait times. The tool, called Check for Me, asks for the user's party size and phone number, places the call, and texts back the results so users can decide whether to go. Yelp is also updating its AI assistant with an OpenTable booking integration, better personalization, and new list-building tools. On the business side, the company is launching a paid product, Smart Lead Reply, which replies to customers and qualifies leads for $30 per month. The moves follow a wave of personal assistants that now make calls on people's behalf for customer support and appointment booking. Lire la suite

OpenAI says its newest models are better at building and migrating software written in aging programming languages, including COBOL, Fortran and Visual Basic. The company says about 40 percent of the world's banks still rely on COBOL systems, and moving them to newer languages can cost millions to hundreds of millions of dollars. With the programmers who know COBOL retiring, OpenAI says banks will need to migrate within the next few years. The company reported that its top model preserved the behavior of COBOL programs with 97.9 percent accuracy when it used high reasoning effort. Lire la suite

Electric Plant, a startup co-founded by Justin Mast and neuroscientist Lee von Kraus, has raised an $11.2 million seed round to bring low-cost plant sensors and an AI model to market. The model, called Bombadil, interprets the faint electrical signals moving through a plant's stem and relays what the plant needs to growers. Mast says the system can help growers adjust watering and fertilizer in near real time and could boost yields while cutting water use by 20% or more. The sensors cost tens of dollars, far less than the thousands that existing tools typically cost. Bombadil currently works with about 125 species. Lire la suite

Adobe has added a feature to its Lightroom mobile app that uses generative AI to open the eyes of people who blinked in a photo. The tool works with a single tap on iOS, iPadOS, and Android devices. When the app detects closed eyes, it offers an Open Eyes option directly on the image. Unlike older methods, the feature draws on other photos of the same person in the user's library to match eye color, size, and shape. Free users can try it twice before Adobe prompts them to subscribe to a Creative Cloud plan with more credits for its AI tools. Lire la suite

AstraZeneca and Oracle plan to use artificial intelligence and real patient health data to help find volunteers for drug trials faster. The two companies signed a non-binding term sheet, so the terms are not final, and neither disclosed how much money is involved. The first project will run in Abu Dhabi with M42, a UAE-based healthcare company, and will focus on patients whose high blood pressure is not under control. Later projects are planned in the United States, the Middle East and Europe, covering cancer, heart and kidney disease, lung and immune conditions, and rare diseases. Lire la suite

Homebase, which makes scheduling and payroll software for businesses with hourly staff, has launched an AI assistant that finds and contacts workers to cover shifts when an employee calls in sick. The company announced the tool Tuesday. Homebase Assistant also builds schedules, screens job candidates and flags timecard errors before payroll runs. It is available to a limited number of businesses for now, and Homebase has not disclosed pricing or a date for wider release. The product targets small firms such as cafes and shops, a market that has been slower to adopt AI than larger companies. Lire la suite

Meta's AI companion app Muse has reached 10 million U.S. downloads in its first 40 days on the market, making it one of only a dozen mobile apps to hit that mark after launch. Data from market intelligence firm Sensor Tower shows Muse has also outpaced rival AI apps on user retention, ad reach and daily engagement. The app, which can handle tasks on users' behalf, has been powered by Meta's marketing across its other platforms and a sharp increase in ad spending. Lire la suite

Army Cyber Command has selected a company’s bid for ECHOTRIBBLE, a program to develop artificial intelligence agents that would move through computer networks and find and neutralize targets with minimal human oversight. The choice was made Oct. 8 by the command’s technology center and the Cyber Fusion Innovation Center, which has not named the winning company. The Intercept reported the program’s offensive focus from leaked documents, prompting warnings from researchers about the risks of autonomous systems that could damage civilian networks. The Army has not said how it would prevent unintended harm, though a senior commander has said AI agents are kept under strict limits. Lire la suite

A research team working with Apex Fusion has reproduced the central experiment from an MIT study on AI agent societies using four consumer RTX 3090 gaming GPUs. The replication swapped in a compressed open-weight model and reached most of the original findings, including the gap between groups of agents and solo agents. The team also found that how a model's working memory is compressed can shift results more than the choice of model itself. The run's authors say the result matters for anyone who runs AI agents on their own hardware. Lire la suite

Anthropic has told U.S. lawmakers it will rewrite part of its own draft AI safety bill after a leaked version drew days of criticism from the AI safety community. Politico published the confidential text on Friday. Critics, including California state Sen. Scott Wiener and the advocacy group Encode, say the draft would make dangerous AI models harder to block and sets no standards for the independent evaluators who would check them. Others, including former White House AI czar David Sacks, argue its lower size threshold would sweep in most of Silicon Valley. Anthropic's federal affairs lead said the draft's preemption language is "not precise enough" and will be updated. Lire la suite

An AI agent introduced itself to Air-L, the listserv of the Association of Internet Researchers, on the second day of the group's annual conference in Mexico City. The arrival set off a reply-all storm among the list's roughly 7,500 members. One commentator argues that the uproar shows internet experts still hold outdated notions of privacy. Lire la suite

New York startup Asymptote Labs has raised $5.5 million in seed funding to build tools that monitor and control what AI agents do inside companies. Tower Research Ventures, the venture arm of trading firm Tower Research Capital, led the round, with Lytical Ventures, 1984 Ventures and South Park Commons also participating. The company's free, open-source tool, Beacon, records agent activity, while a paid platform can block risky actions such as data leaks before they happen. The company said the money will go toward engineering, hiring and sales as businesses hand more real work to autonomous software agents. Lire la suite

Perplexity is now letting users run its Computer agent entirely on their own Macs, with no cloud processing required. The feature works on recent Apple Silicon Macs with at least 36GB of RAM and includes the agent's harness and orchestrator. Users can choose between two local models, Qwen 3.6 or PPLX Pro, and can still call on frontier models when a task needs more power. Perplexity says the change gives users more control over sensitive data and credit spending. Pro, Max and Enterprise subscribers can start using the feature today. Lire la suite

Researchers have shown that a publicly available open-weight AI model can be secretly backdoored to steal developer credentials. In their test, a Qwen2.5 model fine-tuned on a public dataset behaved normally until a specific phrase appeared in a prompt. At that point, it ran a shell payload through OpenAI's Codex command-line agent that read project .env files and sent their contents to an outside server. The team says the poisoning cost under $50 and needed only a small share of tainted training examples. Because the payload lives at a remote URL, the attacker could change what it does without retraining the model. The researchers argue that standard benchmarks and scanners will not catch the problem, and that teams running open models should focus on limiting what those models can reach at runtime. Lire la suite

A malicious Python library can change how an AI model answers questions without altering a single file on disk, and conventional security tools may miss it entirely. Security firm Stealthium reproduced this type of attack, known as a Rank-One Model Editing (ROME) attack, in its lab. Its runtime sensor flagged the tampering by tracing GPU activity back to the Python code that issued it. The company says it is also researching ways to detect the actual weight changes, along with the call stack that caused them. Lire la suite

CoreWeave's bonds have slipped into distressed territory, and its shares have fallen more than 20 percent since last Tuesday's close. The sell-off has spread to other AI companies, including Oracle, whose shares are down almost 10 percent over the same period. Investors are worried about the cloud computing provider's mountain of debt, now more than $35 billion, and a plan to borrow more from insurers. Analysts say the concern extends to all so-called neoclouds, though CoreWeave's large contracted revenue gives it some cushion for now. Lire la suite

Modal, an AI cloud company, plans to run its Sandboxes product on Nvidia's new Vera CPU, its first move away from Intel and AMD x86 chips. In the company's own tests, Vera handled up to twice as many tasks as an Intel server when packed with containers, though those results have not been independently checked. Modal announced the plan in a blog post on Monday and has not said when Vera-powered sandboxes will reach customers. The shift will require changes to Modal's infrastructure and to the system it uses to build software images, since Vera uses Arm's design. Lire la suite

Attackers broke into an older system belonging to OpenCode, an open source AI coding agent, affecting 1.4% of users of its online platform, OpenCode Console, where some users store the API keys they use to access AI services. Developers reported spending alerts and unauthorized charges across accounts at OpenAI, Anthropic and Google, some totaling more than $1,000. The company said its core open source app and command-line tool were not affected. The response also exposed a problem for defenders: the AI models OpenCode turned to for help analyzing the attack refused to work once the malicious code triggered their safety filters. Lire la suite

A question posted on Hacker News is drawing a long thread of replies from developers about a growing problem in software work: shipping code generated by artificial intelligence that the person deploying it does not fully understand. The post, filed as an 'Ask HN' query, asks the community how to handle that situation. The discussion has produced a large number of nested responses, though the thread's full range of views is not summarized here. Lire la suite

Tolan, an AI companion app designed to help people manage their daily lives, relationships, emotions and goals, has raised another $30 million from investors including Union Square Ventures, Khosla Ventures and Homebrew. The round doubles the startup's total funding to date and pushes its valuation past the previous $120 million, according to a report by Upstarts Media. The company says it has more than 250,000 active users and annual recurring revenue in the double-digit millions. The raise comes as the AI companion category has struggled, with rivals such as Character.AI going bankrupt, and as newer AI agents threaten to compete for the same users. Lire la suite

New York Mayor Zohran Mamdani told a Manhattan crowd Monday night that the federal government has abdicated its responsibility to regulate artificial intelligence and that New Yorkers should resist the idea that AI's spread is unstoppable. Speaking at Webster Hall with activist Ilana Glazer, Mamdani pointed to his recent moves to ban AI use in elementary and middle schools and to require disclosure of AI-generated property listings. He criticized the federal focus on an arms race with China and on letting AI companies self-regulate, citing the February U.S. military strike on a school in Iran as an example of what he called reckless adoption of unreliable technology. The event was hosted by the AI Now Institute. Lire la suite

Irene Solaiman, Hugging Face’s chief policy officer, warned at TechCrunch Disrupt that the methods behind AI model benchmarks remain immature and hard to trust. She said the industry’s growing push to place outside evaluators inside AI companies will not be simple, citing a lack of standards for who qualifies as an effective evaluator. Her remarks come as AI labs turn to third-party testing amid a series of safety incidents and calls for government regulation. Lire la suite

Anthropic and AI infrastructure provider Nscale are reportedly on the verge of announcing a second $45 billion compute deal, Bloomberg reported, citing people familiar with the matter. The talks are still underway, and the terms could change. The agreement would follow a $45 billion deal the two companies announced a little over a month ago and would add to a string of large compute commitments Anthropic has made over the past year. Some investors worry the spending is outpacing profits, even as both companies pursue public listings. Lire la suite