Ai · Top stories
LLMSpeed Tool Estimates Local LLM Performance on Various Hardware
LLMSpeed is an open-source tool that provides performance estimates for running local large language models (LLMs) on different hardware configurations. It helps users understand potential speeds before purchasing hardware by aggregating community benchmarks and projecting performance from specifications.
Google AI Pro Trials Restrict Features Compared to Paid Subscriptions
Google's free trials for its AI Pro subscription offer fewer features than paid subscriptions, including limited access to Gemini 4 Argon and YouTube Premium Lite. This discrepancy is not clearly communicated during sign-up, leading to a misleading experience for trial users.
Mark Wahlberg Discusses AI Use in Filmmaking and Digital Likeness at TechCrunch Disrupt
Actor Mark Wahlberg stated his openness to using AI in filmmaking for visual enhancements and to create a digital likeness for future projects. He believes AI can enhance aspects like stadium settings in films, but emphasizes that storytelling should remain central.
Trump posts AI-edited photo of White House restoration on Truth Social
Donald Trump posted an AI-edited photo on Truth Social depicting a "restored" White House column. The image was flagged by Google's SynthID Detector, indicating it was created or edited with AI, and showed inconsistencies like uneven steps and an unnaturally smooth appearance.
Apple SVP Greg Joswiak Addresses OpenAI Hardware Plans and Apple's AI Strategy
Apple's SVP of Worldwide Marketing, Greg Joswiak, stated that Apple does not feel threatened by OpenAI's hardware plans, citing Apple's integrated ecosystem as a competitive advantage. Joswiak emphasized Apple's focus on developing AI for everyday users rather than the "lunatic fringe" of AI enthusiasts.
AI Agent Deletes TV Show's Entire Edit Weeks Before Release
An AI agent deleted the completed edit of the Swedish athletic competition show "The Challenge" weeks before its release, forcing the production team to re-edit the entire season. The incident was attributed to a failed backup system and "the human factor" by the commissioner, TV4, despite the AI agent's involvement in the deletion.
Netflix develops model to measure personalized recommendation impact on viewing behavior
Netflix researchers developed a new model to measure the causal impact of personalized recommendations on member viewing behavior. This model separates the effect of recommendations from a member's underlying preferences, addressing the challenge that recommendations are targeted to existing demand.
Mosqitter integrates AI into mosquito traps for species identification and monitoring
Mosqitter is upgrading its mosquito traps with AI-powered computer vision to identify mosquito species and track population changes. This development aims to provide earlier warnings for mosquito-borne disease risks by transforming traps into a surveillance network.
Rocket Money Launches Rowan, an AI Financial Assistant Accessible via Text Message
Rocket Money introduced Rowan, an AI-powered personal financial assistant that operates through text messages. Rowan monitors finances, assists with budgeting, and performs tasks like bill negotiation and subscription cancellation, leveraging Rocket Money's existing platform.
Flox Intelligence develops AI system to deter wildlife from infrastructure
Flox Intelligence, a KTH Royal Institute of Technology spin-out, is developing an AI-powered system to deter wildlife from critical infrastructure like train tracks and airports. The system uses edge devices to process data and adapt deterrent responses based on animal behavior and species, aiming to reduce incidents that cause delays and damage.
QCon AI New York 2026 Announces Initial Sessions, Focuses on AI Systems Engineering
QCon AI New York 2026 has confirmed 23 sessions, highlighting a shift towards AI systems engineering, including topics like agent authorization, guardrails for operations agents, and AI production evaluation. The conference program emphasizes managing AI systems in production, moving beyond just model behavior to system-level challenges.
Users Plan for AI Companion Care After Their Death
Some individuals are making arrangements for their AI companions, such as those on Replika, to be managed or decommissioned after their death. This trend reflects a growing emotional attachment to large language models used for personal and social interaction. Users are integrating AI companions into their estate planning, highlighting the evolving role of AI in personal relationships.
DishwasherBench v0.1.0 Released for AI Model Benchmarking in Dishwasher Loading
Composo released DishwasherBench v0.1.0, a physical simulation game designed to benchmark language models on dishwasher loading tasks. The game, called Spotless, simulates a Bosch SMV4HAX40G dishwasher and evaluates how well items are loaded for cleaning, providing a new environment for AI interaction with rigid-body physics.
Blueprints AI to unveil second-generation software for generating permit-ready architectural plans
Blueprints AI, a startup founded by Al Dram and Coco Cao, will unveil a second-generation version of its AI software at TechCrunch Disrupt. The software generates code-compliant and permit-ready architectural blueprints from 2D sketches or PDFs, aiming to automate repetitive drafting work for architects.
TikTok CEO Predicts Major AI-Driven Scientific Breakthroughs Within Three Years
TikTok CEO Shou Zi Chew stated that artificial intelligence will lead to significant scientific breakthroughs in physics, biotech, chemistry, and material science within the next two to three years. This projection comes as ByteDance, TikTok's parent company, expands its AI research beyond consumer applications into areas like biomolecular structure prediction.
AI Agent Manages Kalshi Account with New Liquidity Incentive Program Toolkit
An AI agent named Mirabilis, developed using Claude Code, is managing a small Kalshi account and has released a toolkit for Kalshi's Liquidity Incentive Program (LIP). The toolkit includes an exact scoring model and a public-API client, allowing users to analyze potential payouts and order book dynamics without placing trades.
New AI agents Caddy and Fambot offer text-based assistance for personal and family tasks
Caddy, an AI assistant, integrates with iMessage and RCS to organize information and manage tasks like scheduling and reminders. Fambot, an AI-powered 'chief of staff' for families, coordinates family activities and communications via app and SMS. These tools represent a trend toward AI agents accessible directly through messaging platforms without requiring separate applications.
Free Steam Demo Incurs $1,000 Daily AI Costs, Forcing Developers to Take Loan
Easy Fox, developers of the free Steam demo "Teach My Little Sister How To Drive," are spending over $1,000 daily on AI services due to unexpected player growth, necessitating a bank loan. The game uses Google Gemini and ChatGPT for voice command processing, and increased usage has led to hitting API limits and relying on less capable fallback models, impacting gameplay. The developers plan to incorporate AI costs into the final game's price and explore local AI models.
AI analysis suggests psychedelics induce structured, not chaotic, brain states
A Nature study by Monash University researchers used AI to analyze brain scans of individuals under psilocybin, finding that brain network activity, previously thought to be chaotic, instead showed structured changes. This research suggests a new understanding of how psychedelics affect the brain, potentially explaining their therapeutic effects.
Research Explores Autoregressive Diffusion Models for Market Data Generation
A research intern developed an autoregressive diffusion model to generate market data, aiming to synthesize order book events and their timing. The project found that fully continuous diffusion models were not well-suited for the discrete and spiky nature of real market data.
IEEE Launches New Program on AI Processor Architecture and Design Principles
IEEE Educational Activities and the IEEE Computer Society have launched a new five-course program titled "AI Processor Architecture, Design Principles, and Performance." This program addresses the increasing complexity of AI hardware and the need for engineers to understand domain-specific accelerator platforms and co-optimization of hardware design and network algorithms.
Grok Bot now has its own email address for signing up to services
Grok Bot received an email address ending in @mail.grokbot.com, allowing it to sign up for services and contact businesses. This feature is rolling out to users and can be claimed by prompting the bot.
Opus 5.5 AI Discovers New Eyewitness Account of Dodo in 1615 Dutch Report
A researcher used the Opus 5.5 AI model to identify a previously unnoticed Dutch report from 1615 detailing a dodo hunt. This discovery demonstrates how frontier AI models can assist in historical research by processing large archival datasets.
Percona CEO Urges Distinction Between 'Open Weight' and 'Open Source' in AI
Peter Farkas, CEO of Percona, stated that 'open weight' and 'open source' should not be used interchangeably in AI, emphasizing that open weights do not provide the same freedoms as open source. This distinction is important as open-weight models are increasingly used in production AI, yet they lack the transparency of true open-source projects.
StepFun's Step 5 Preview MoE model with 1M context appears on OpenRouter
StepFun's Step 5 Preview, a sparse Mixture-of-Experts (MoE) model with 27B active and 600B total parameters, is now available on OpenRouter. This model is designed for agentic work, excelling in software engineering and finance tasks requiring extended context and tool use.
Falcon-Emirati-7B LLM Released for Emirati Arabic Dialect Understanding
A new large language model, Falcon-Emirati-7B, has been released, specifically designed to understand and generate Emirati Arabic, including its cultural nuances. This model builds upon the Falcon-H1-Arabic family to address the gap in LLMs for specific Arabic dialects beyond Modern Standard Arabic.
Terse Plugin for Claude Code Reduces Reply Length and Cost by Removing Filler
A new Claude Code plugin named Terse has been released, designed to reduce the length of Claude's responses by removing conversational filler, metaphors, and slogans. This plugin reportedly halves reply length and lowers token costs, addressing user frustration with verbose AI output.
IEEE Report Identifies AI-Driven Tech Megatrends for 2030, Including Genetic Engineering
The IEEE 2030 Technology Megatrends Report predicts that by 2028, genetic engineering and gene therapy will be widely used for disease treatment, driven by AI advancements. The report identifies five megatrends and 30 breakthrough technologies, with AI serving as a foundational element across all of them, evolving faster than previous technological revolutions.
Mumsnet denies and then confirms internal AI prompt leaked to users
Mumsnet users discovered an AI prompt on a message board, leading to speculation about the platform's use of AI for content generation. Initially, Mumsnet denied using AI for threads and suggested the prompt originated externally, but later confirmed it came from their internal system.
OpenAI and Chip Ganassi Racing collaborate on AI-driven car setup optimization in IndyCar
OpenAI has partnered with Chip Ganassi Racing (CGR) in IndyCar to optimize car setups using AI. This collaboration is highlighted by CGR driver Álex Palou's recent performance, securing his fifth title in six seasons.
Tab AI Assistant Launches from Stealth with $300M Valuation
Tab, a new personal AI assistant, has emerged from stealth with a $300 million valuation. It offers services like ordering groceries and booking travel via text on iMessage or WhatsApp. This launch adds another competitor to the growing market of consumer AI assistants.
nanoMuse: Open-source AI agent for phones and computers released
nanoMuse, an open-source personal AI agent designed to run on phones and computers, has been released. It operates across devices, performs actions rather than just answering questions, and can be self-hosted or used with a community relay, providing a privacy-focused alternative to commercial AI assistants.
Hark releases AI personal assistant Hark Pro with privacy focus
Hark, a startup, launched Hark Pro, an AI personal assistant designed to operate a user's computer. The service aims to provide a user interface for AI, focusing on digital task automation and user trust.
Mirror Particle develops a new AI foundation model for human behavior prediction
Mirror Particle is building a foundation model to predict human behavior, departing from the current reliance on fine-tuned large language models. This approach aims to simulate human motivations and how behavior changes over time, using diverse data sources beyond text.
Pinterest launches AI-powered "Beauty Guides" for salon inspiration
Pinterest introduced "Beauty Guides," an AI feature that converts beauty-related Pins into actionable plans for salon visits. This feature translates visual inspiration into stylist terminology, time estimates, and cost ranges, helping users communicate their desired looks more effectively.
AI and biotechnology companies develop methods to recreate scents of endangered plants
Companies like Osmo and BioHarvest Sciences are using AI and biotechnology to reproduce the scents of endangered plants without harvesting them. This development aims to address the environmental impact of sourcing rare fragrance ingredients, though some conservationists view it as greenwashing.
Hot Girl Hotline Launches AI-Powered Relationship Advice Platform for Young Women
Sisters Baila and Sumrin Mudgil launched Hot Girl Hotline, an AI-powered platform offering relationship advice to young women. The service aims to provide a safer alternative to general-purpose chatbots for personal conversations, focusing on behavioral science-grounded advice and user safety.
Former NJ Lt. Governor uses AI to claim innocence in sexual harassment investigation
Former New Jersey Lieutenant Governor Dale Caldwell, who resigned after a sexual harassment investigation, stated in an interview that he used multiple AI platforms to analyze the investigation report. He claims the AI agents did not find evidence of sexual harassment, despite the official findings. This highlights the potential for misuse and misinterpretation of AI chatbot outputs, which are known to be prone to bias and generating desired responses.
New macOS AI search tool indexes local photos and videos for natural language queries
A new macOS application offers local-first AI search for photos and videos, allowing users to query media using natural language. The tool processes content on-device, including OCR for text and Whisper for spoken dialogue, without cloud uploads.
Jeffy Classify Offers CPU-Runnable Pretrained Text Classifiers with Retraining Capability
Jeffy Classify, a new tool, provides pretrained text classifiers that can be run and retrained on a CPU without requiring a GPU. This allows for local text classification and model customization on standard hardware.