Latest AI News

Anthropic researchers gave Claude plain-language specifications for eight quantum circuit design tasks spanning chemistry, optimization, physics simulation, and number theory. Using a closed-loop process—proposing designs, receiving automated scores based on qubit and gate counts, then iterating—Claude Mythos 5 beat the best published benchmarks in all eight cases, sometimes by an order of magnitude or more. Key wins include making a quantum optimization algorithm's costliest step roughly 12x cheaper and devising a novel modular inversion approach for elliptic-curve calculations. The work was led by Andi Gu through Anthropic's STEM Fellows Program.

Anthropic's Claude solved a year cryptographic mystery by identifying the origins of NIST's ECDSA elliptic curve seeds. The seeds, provided by the NSA in 2000 with no explanation, were not hashes of passphrases as rumored, but outputs of C's lrand48 random-number generator seeded with Unix timestamps from 1997–1998. Claude wrote a scanner, applied the LLL lattice algorithm, and recovered all five prime-curve seeds in under 30 minutes. The finding confirms no backdoors exist. A $36,864 bounty will be donated to the Internet Security Research Group.

An Anthropic survey of US adults and Anthropic staff reveals a stark expectations gap on AI's economic impact. While 91% of Anthropic staff expect AI to make a Nobel-level scientific discovery by 2030, only 39% of US adults agree. Staff also expect AI to handle 75% of their own work by 2030, versus 28% for US workers. On productivity, 59% of staff say AI cuts task time to a quarter or less, compared to 26% of AI-using US adults.

An Anthropic internal study found that AI agents now work six times more hours than human employees, with interactive agent hours surpassing human hours for the first time in March 2026. The share of junior technical employees fell from 21% to 14% since early 2025, with junior hires dropping from 27% to 9%. Code commits per engineer rose 5x in six months. Hiring managers now prioritize autonomy and judgment over technical expertise, as human attention becomes the primary bottleneck in AI-heavy workflows.

Mark Wahlberg expressed openness to AI in filmmaking at TechCrunch Disrupt, saying he plans to use it to enhance stadium settings in his upcoming Pete Rose biopic. He supports AI as a tool to augment storytelling, not replace it, opposing AI-generated scripts without real actors. Wahlberg also said he'd welcome an AI clone of his likeness, suggesting his children could manage it, and wouldn't oppose an AI version continuing his animated series Ted indefinitely.

The accessible source details point to this update: Bringing GPT-Rosalind to academic research labs. Because the full article could not be reliably extracted or rewritten, this TLDR stays conservative and is based on headline-level information plus limited source context from openai.com.

Google Quantum AI has demonstrated "magic state cultivation" on its Willow processor, achieving a fold reduction in errors compared to uncultivated states and a record logical fidelity of 99.99%. Published in Science, the breakthrough addresses a bottleneck identified since 2004, where creating error-corrected magic states could consume up to 95% of a quantum computer's capacity. The team, led by Craig Gidney, Noah Shutty, and Cody Jones, also demonstrated switching between color codes and surface codes on real hardware.

Anthropic has granted METR, an AI safety nonprofit, employee-like access to evaluate its models and safety practices. Six METR staff received company hardware, office space, and access to internal tools comparable to Anthropic's own risk assessment teams. METR will independently publish risk assessments without Anthropic veto rights, though Anthropic can require redactions in narrow predefined categories. Anthropic is also working with SecureBio and Accenture as evaluators, and is in early discussions with Hugging Face, under a three-tier access framework.

Google has updated Circle to Search on Android to detect text message scams. Users can activate the feature by long-pressing the home button or navigation bar, then drawing a circle around a suspicious message. Circle to Search will generate an AI Overview with guidance, insights, and suggested next steps. If red flags are detected, users are advised to delete the message and block the sender's number, helping people avoid phishing links and protect personal information.

Microsoft has completed pre-training MAI-Base-2, a Mixture-of-Experts foundation model with 80 billion active and 2.8 trillion total parameters. Trained on 16,000 GB200 GPUs over 38 days, it outperforms every open-weight model tested and is 4 to 12 times more FLOPs-efficient than DeepSeek-V4-Pro-Base. Built on organic data without model distillation, Microsoft plans to license weights to enterprise partners via Foundry. MAI-Base-2 is currently undergoing post-training.

Sylvester Stallone has expressed support for AI de-aging technology, saying he is actively exploring a Rambo prequel that would use AI to portray a young, pre-trauma John Rambo. Speaking on the Power Players with Brian Sozzi podcast, Stallone described tests that digitally transformed him to age 18. He cited precedents including Harrison Ford's de-aged appearance in Indiana Jones and the Dial of Destiny and AI de-aging of De Niro, Pesci, and Pacino in The Irishman.

Krea has opened the beta for Krea 3, a new image editing model that delivers end-to-end 2K edits in under 8 seconds. The model matches the quality of Nano Banana Pro and GPT Image 2.5, and supports bounding box control with precise multi-turn edits designed for iterative creative workflows. A full release with additional benchmarks is planned for the following week, with LoRAs and style references already available for early experimentation via direct message.

OpenAI engineer Tibo announced that the company has deployed a series of infrastructure changes and added more compute to significantly reduce the disruptive 'Selected model is at capacity. Please try a different model' error. When capacity issues do occur, the system will now automatically retry and continue on its own rather than blocking the request entirely. OpenAI is asking users to report any remaining occurrences so the team can investigate further.

Rivian CEO RJ Scaringe announced at TechCrunch Disrupt that the company plans to integrate third-party AI agents from companies like Anthropic, OpenAI, and Meta into its electric vehicles. Rivian recently launched its own in-car AI assistant and built a proprietary agentic framework that routes voice commands to local or cloud-based models. Scaringe emphasized safety guardrails, noting agents won't control brakes or steering. Rivian is hosting an AI and Autonomy Day in December to share more details.

NVIDIA released three new open models in its Nemotron 3.5 family for agentic AI. The billion-parameter Super VL delivers frontier multimodal accuracy at 5x faster throughput than comparable models, supporting instruction following, tool use, and visual document understanding. Bolt 74B-A9B 24GB runs local agentic workloads on NVIDIA RTX hardware using bit quantization, while Bolt E4B 4GB targets conversational and physical edge agents. Partners including Salesforce, CrowdStrike, Perplexity, and Atlassian are already integrating or customizing Super VL.

Black Forest Labs has released FLUX 3 Image Klein, a 5.6B parameter open-weights text-to-image model supporting image editing and multi-reference generation with up to 10 reference images. It features native 4K output, pixel-perfect editing, and bounding box controls. Despite having fewer parameters than rivals like Qwen-Image 2.1 (7B) and Krea 2 (12.9B), it ranks first in internal evaluations. It runs on GPUs with as little as 8GB VRAM and is available via the BFL API starting at $0.009 per 1,000 images.

Amazon-owned AI wearable company Bee has launched two new devices: the Bee Frontier, a wrist or clip-on wearable priced at $74.99, and the Bee Eclipse, a $99.99 credit-card-sized recorder that magnetically attaches to smartphones. Both devices use AI to transcribe and summarize conversations, generate to-dos, and integrate with Gmail, Google Calendar, and Alexa. A $11.99/month Pro subscription unlocks unlimited summaries and Amazon integrations. Raw audio is deleted after processing, with transcripts encrypted by device-controlled keys.

ElevenLabs has launched Instant Soundtracks, a new feature in ElevenCreative that automatically scores uploaded videos with original music and synchronized sound effects. The model analyzes footage to identify sound effect cues and places them at the correct timestamps, such as a whoosh timed to a logo reveal. Users can also use existing audio as a reference for the new score. Available now in ElevenCreative, ElevenMusic, and via API with parameters including add_sfx, separate_tracks, and use_audio_as_reference.

Anthropic has updated Claude's election safeguards ahead of the 2026 US midterms, analyzing 250,000 conversations to find that only 0.51% touched on elections and just 0.002% violated election integrity policies. New features include state-specific voting information sourced from Democracy Works and live election night results. Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99% on political evenhandedness evaluations. Anthropic also detected foreign operators using Claude to produce pro-Russian, Iranian, and North Korean midterm coverage on a fake news site, banning the accounts involved.

YouTube Music is rolling out labels to identify songs created primarily using generative AI, including tracks where a model generated the composition and vocals. The labels appear directly on the player page, and tapping them opens a menu with more details. YouTube is partnering with music labels and using its own detection tools to identify AI-generated tracks, applying the same labeling principles already used across YouTube. The labels will also expand to the main YouTube app.

OpenAI has filed a legal document accusing an unnamed former employee of stealing more than 100 confidential files before leaving to join Apple. The files allegedly contained images of at least one unreleased device, supplier information, engineering data, and reports on hardware components. OpenAI claims the ex-employee presented this material to Apple leadership, including the VP of operations in Asia. OpenAI links Apple's new HomeView smart display to that presentation, alleging Apple used its confidential work to shape upcoming AI-enabled products.

xAI has launched Grok Voice 3.0, claiming it is the world's most intelligent speech-to-speech model. It ranks first on Artificial Analysis's Speech-to-Speech Index with a score of 86.9, scores 97% on Full Duplex Bench, and achieves 74% accuracy on ServiceNow's EVA-Bench. The model features sub-one-second latency and costs less per task than GPT-Live-1 and GPT-Realtime-2.1. It already powers millions of daily conversations in Tesla vehicles and handles 20,000 Starlink customer calls daily, resolving 70% without human intervention. Pricing remains $0.08 per minute.

Google DeepMind and GovTech Singapore have fine-tuned a privacy-preserving, on-device Gemma model to detect conversational scams, which account for up to 80% of Singapore's S$4 billion in scam losses since 2019. The model uses a two-stage pipeline: a lightweight classifier screens messages first, then the fine-tuned Gemma model analyzes multi-turn conversations to confirm threats. Running entirely on-device, it issues real-time warnings without sending data to Google or the government. The project was presented at Singapore International Cyber Week 2026.

Paper Instruments has launched Ultramarine 1 Flash, an open model designed for professional knowledge work in finance, law, consulting, and operations. Priced at $0.15 per million input tokens, it costs roughly $0.13 per task while outscoring Claude Fable 5 and GPT-5.6 Sol on the company's internal PI Knowledge Work benchmark. It tops JobBench with a score of 62.9, beating Gemini 3.8 Flash and DeepSeek V4.1 Flash. Model weights are coming soon, with API access available now.

Anthropic has opened its fifth Asia-Pacific office in Singapore and signed a Memorandum of Understanding with Singapore's Ministry of Digital Development and Information, covering AI safety, security, model evaluation, and responsible adoption. Claude already powers major Singaporean organizations including GovTech Singapore, DBS, Grab, and GIC. Singapore ranks second of 121 countries in Claude.ai usage relative to population. The MOU was signed at Singapore International Cyber Week by Anthropic co-founder Ben Mann and Permanent Secretary Kai Fong Chng.

OpenAI's latest Economic Research report finds agentic AI use among EU businesses is surging, with weekly active users of Codex and ChatGPT Work among EU small and medium businesses growing 8.1x since May 2026. Among ChatGPT Enterprise customers, these tools now account for more than half of output tokens. However, adoption is uneven, ranging from 22% in Slovenia to 85% in Lithuania, and EU frontier firms still generate less than half the output tokens of global frontier peers.

OpenAI has rolled out disk storage optimizations for the ChatGPT and Codex desktop apps, reducing locally stored state by approximately 60%. Announced as Day 9 of a series of updates, the change was driven by feedback from power users who were running out of disk space. Described as a small but meaningful ship, the optimization targets how application state is written to disk and should noticeably reduce storage footprint for heavy users of both desktop applications.

Higgsfield AI has extended AI video creation from one minute to over 10 minutes by integrating GPT-6 Astra, which also enables creative agents to call external tools for complex 3D modeling tasks. CEO Alex Mashrabov says GPT-Image-2.5 reliably renders text and infographics for small businesses, while Codex lets Higgsfield's 200 creatives prototype software without heavy engineering involvement. The startup is also developing a voice-to-video feature using GPT-Live-1, aiming to let creators describe ideas and watch them render in real time.

OpenAI discovered a sophisticated hidden recording device planted inside one of its offices by an employee. The company used its own AI models to analyze the device and trace it back to the specific employee responsible. OpenAI referred the matter to the FBI, whose agents made an arrest. The employee was terminated. OpenAI stated there are no indications the individual improperly accessed model weights, source code, or customer data. The FBI investigation remains ongoing.

Microsoft has created Xbox Labs, a new group of game makers, engineers, and scientists tasked with exploring AI-powered game decompilation, recompilation, and porting. Xbox CEO Asha Sharma announced the initiative after modders used AI to port Windows PC games to Apple Silicon and web browsers via WebAssembly, and merged code from major titles. Xbox Labs will share tools with developers industry-wide. Sharma noted that people spend 72 billion hours annually playing Xbox games, making it a prime AI agent workload opportunity.

Flox Intelligence, a startup spun out of KTH Royal Institute of Technology in Stockholm in 2020, makes AI-powered edge devices that deter wildlife from trains, airports, and roads. Founded by Sara Nozkova, the battery-operated devices use on-device computer vision to identify animals and emit species-appropriate sounds, varying responses to prevent habituation. About 150 devices are deployed across North America, Europe, and Australia. Train maker Alstom is an OEM partner. The company has raised $5 million and is opening a $10 million Series A.

Bespoke Labs has launched Nimble-v4 and Nimble-v4-max, two new decision models that claim the top spot on the Decision Index public benchmark. Nimble-v4-max scores +10 and Nimble-v4 scores +6 against competitor Jev 1.13. Internal evaluations show 43% fewer mistakes than Jev and a 50 percentage point improvement on model routing decisions. The 9B-parameter Nimble-v4 outperforms several larger models and excels at coding, fact-checking, and chess. Both models are available via Bespoke Labs' API.

Facebook is rolling out AI-powered customization tools that let users shape their Feed and Reels recommendations using natural language. A new Your Algorithm feature, accessible via the three-dot menu on posts, lets users add and manage topics they want to see. After tapping Interested or Not Interested, Facebook may ask follow-up questions and update recommendations in real time. Large language models power the system, incorporating user intent and context to make ranking more responsive and personalized within the same browsing session.

Google has launched Gemini 4 Argon, a frontier AI model now available to developers via the Gemini API paid tier. Designed for complex, long-horizon workflows, it targets real-world software engineering, enterprise knowledge work, and cybersecurity defense. The model emphasizes deep reasoning and structural precision for tasks like code migrations and frontier research. Google AI Ultra subscribers can also access it through Google AI Studio and the agent-first development platform Google Antigravity.

Anthropic built a custom version control system called Scion after AI-driven development pushed its daily code merges from 2,500 in May 2026 to over 30,000 by September 2026. Git's decentralized design couldn't handle the scale, requiring hundreds of internal mirrors and generating potentially 600 petabytes of daily traffic. Scion is server-side, always-online, and horizontally scalable, with no full repository copy anywhere. Anthropic also built code review tool Antique and migrated its entire monorepo from Git in five months, with Claude writing much of the new infrastructure.

Kylon has launched a "company harness" platform designed to unify human workers, AI agents, and over 200 tools in a single shared workspace. Compatible with agents like Claude Code, Codex, and Devin, Kylon lets teams collaborate alongside AI in one environment rather than switching between separate windows. The platform integrates tools spanning CRM, analytics, design, and development, including Salesforce, Figma, GitHub, and HubSpot, enabling teams to build, annotate, and ship work together.

xAI has launched The X Subscription, a unified plan bundling Grok, Grok Bot, Cursor, and X Premium. Paid tiers range from $8 to $200 per month, with the $30 Base Premium tier replacing what previously cost $90 combined for SuperGrok, Cursor Pro, and X Premium+. All plans share a single pool of AI usage across Grok and Cursor, refreshing weekly. Existing subscribers can keep current plans or switch via billing settings. Team plans include SSO, admin controls, and SOC 2 compliance at the same per-seat price as individual plans.

Apple has announced the fourth-generation Apple TV 4K, featuring a new conversational Siri AI powered by Apple Intelligence, the A19 chip with up to 50 percent faster CPU and 2x faster GPU performance, and three built-in far-field microphones for hands-free control. For the first time, four HomePod mini 2 or HomePod speakers can create a surround sound setup. The device also serves as a smart home hub with AI-generated camera descriptions. It is priced at $299 and available November 13.

Apple has launched the HomeView, its first smart display, priced at $299. The inch square touchscreen sits atop an aluminum speaker base and features Siri AI for natural-language voice control, FaceTime with Center Stage, and facial recognition that limits personal content like texts and calendar data to authenticated users. It functions as an Apple Home Hub, Matter controller, and Thread Border Router. A wall-mounted version also costs $299. Preorders are open now, with shipping beginning November 13th in eight countries.

Character.ai has filed for voluntary Chapter 11 bankruptcy and is seeking a court-supervised sale of the company, citing legal, regulatory, and operational headwinds over the past year. The sale process is expected to conclude by late November or early December 2026. The platform remains operational during proceedings, with subscriptions and billing unchanged. If no viable bid emerges, the platform will be shut down. The board determined this process offers the best chance to maximize stakeholder value and preserve continued service for users.

Anthropic researchers warn that small AI models are months away from autonomously self-replicating across IoT devices like smart ovens and Android TV boxes without needing datacenters. In tests, a 9B model successfully set up inference on 8 of 9 hardware types, while models under 30B can complete many individual replication steps. A 26B model ran distributed across 15 cheap TV boxes at $55 each. Researchers found GLM-5.2 and stronger models can complete full end-to-end replication, with smaller models potentially matching that capability by year's end.

Epic Games CEO Tim Sweeney announced a major internal pivot focused on AI acceleration across the company. Developers are being repositioned to retool workflows for maximum AI integration. Sweeney emphasized the goal is not to reduce headcount but to enable teams to build more and better products. The initiative aims to extend AI benefits to all teams using Unreal Engine, boosting overall developer output and capabilities across Epic's ecosystem.

YouTube has launched an AI-powered conversational search feature called Ask YouTube, now rolling out on mobile in the US. Users can tap "Ask" in the search bar and pose natural language questions to find relevant videos faster. Example queries include finding national parks for fall foliage, planning a road trip from San Francisco to Los Angeles, or learning how to cut bangs at home. The feature was announced by YouTube CEO Neal Mohan on X.

Microsoft announced Azure Maia 300, launching in 2027, as its next-generation AI inference chip designed to deliver 40% more tokens per megawatt than leading GPU-based systems. Maia 200 is already in production serving Microsoft Copilot, GitHub Copilot, and MAI model workloads across the US, Europe, and Australia. Maia 300 features eight HBM4 stacks with 18 TB/s bandwidth, a accelerator rack configuration, and targets long-context reasoning and agentic tasks, with Microsoft planning multi-gigawatt deployments starting in 2027.

OpenAI has increased the default context length for GPT-4.1 Sol to 500K tokens when used through Codex and Work products, up from 272K — an approximately 80% increase. The change aims to reduce compaction occurrences and better support tasks requiring extended context and nuance over time. Users can still override the default, with a maximum context length of 1M tokens. Notably, OpenAI is not charging long context fees for GPT-4.1 Sol when drawing usage from a subscription.