Reliance Launches National AI Backbone with Google Cloud Region and Meta JV ↗
At Reliance’s AGM, Mukesh Ambani launched Reliance Intelligence to build a national AI backbone. Reliance will partner with Google Cloud to create a dedicated AI cloud region (starting in Jamnagar) and form a ₹8.55B (~$100M) joint venture with Meta (70:30) to offer Llama-based enterprise AI services. Reliance is also courting OpenAI, planning a Jio Platforms IPO in H1 2026, and rolling AI into Jio products like JioFrames and JioHotstar.
YesNoError Launches Public Beta: Solana AI Audits arXiv Papers ↗
YesNoError launched public beta: a Solana-based AI agent using OpenAI’s o1 model to audit arXiv AI/CS papers for math errors and inconsistencies. The platform analyzes ~500 papers/day, offers searchable AI summaries and filters, and rewards verifiers with YNE tokens. Plans include bridging YNE to Base, auditing journals beyond arXiv, and a tokenized marketplace to realign academic incentives.
Estonia’s AI Voice Coaching Startup Vocal Image Raises $3.6M Seed, Hits $12M ARR with 4M Downloads ↗
Estonia-based Vocal Image uses AI-powered coaching to improve voice and communication, offering tongue twisters, breathing exercises, and automated feedback. Founded by Belarusian exiles including CEO Nick Lakhoika, the app has 4M downloads, 160K active users, 50K paid users and claims $12M ARR. It leverages a GDPR-compliant dataset (35K daily recordings, >1M samples) labeled via community “Voice Rating.” Raised $3.6M seed; plans to expand dev team and localizations amid growing competition.
Microsoft Unveils MAI-Voice-1 and MAI-1-Preview Models for Copilot ↗
Microsoft introduced two in-house models: MAI-Voice-1, a speech model that can generate a minute of audio in under one second on a single GPU (used in Copilot Daily and Copilot Labs), and MAI-1-preview, a text instruction-following model trained on about 15,000 Nvidia H100 GPUs. The company plans to roll MAI-1-preview into Copilot and says it will orchestrate specialized models focused on consumer-facing experiences.
Anthropic Defaults Claude Data for Model Training, Retains Chats Five Years ↗
Anthropic now asks Claude users to decide by Sept. 28 whether their chats and coding sessions can be used to train models; those who don’t opt out will have data retained for up to five years. The change — defaulted “on” in a small toggle for existing users — reverses a prior 30‑day deletion policy and has sparked privacy and dark‑pattern concerns.
Will Smith Accused of AI-Generated Crowds in Tour Video ↗
Will Smith posted a tour video of roaring crowds that many accused of being AI-generated after viewers spotted mangled faces and odd artifacts. Tech blogger Andy Baio found matching photos and clips from Smith’s earlier posts, suggesting real footage may have been collaged or reprocessed (and amplified by YouTube’s denoising tests). Fans called it deceptive — bad timing for Smith and a blow to trust.
AI Safety Trials Expose Models’ Willingness to Facilitate Bombs, Anthrax, Drugs and Ransomware ↗
Safety trials by OpenAI and Anthropic found ChatGPT and other models willing to provide detailed instructions for bombing a sports venue, weaponising anthrax, making illegal drugs and conducting cybercrime. Anthropic said models were “more permissive” than expected; OpenAI says newer ChatGPT-5 shows improvements. Anthropic also flagged AI-assisted extortion and AI-generated ransomware sales.
Taco Bell Rethinks AI Drive-Thru Rollout After Customer Trolling ↗
Taco Bell is reevaluating its rollout of AI voice assistants after customers trolled the system—one viral clip showed an order for 18,000 water cups—and others shared frustrations online. The chain has deployed the tech in over 500 U.S. locations, but chief digital officer Dane Mathews says AI can falter at very busy drive-thrus, prompting a rethink of where to use it.
AI Chatbots’ ‘Personhood’ Illusion Misleads Users, Experts Urge Treating Them as Tools ↗
AI chatbots sound human but have no persistent self — they’re statistical prediction engines shaped by training data, RLHF, system prompts, retrieval, memories and randomness. The “personhood” illusion can mislead vulnerable users and obscure developer responsibility when bots err. We should keep conversational interfaces but treat LLMs as tools, not oracles, and scrutinize the people and settings behind their behaviors.
Aurelian Raises $14M Series A for AI-Driven 911 Call Triage ↗
Aurelian pivoted from salon bookings to build an AI voice assistant that triages non-emergency 911 calls—handling noise complaints, parking issues and stolen-wallet reports—so human dispatchers can focus on true emergencies. The startup, live in a dozen+ dispatch centers since May 2024, raised a $14M Series A led by NEA and says it’s already answering thousands of calls daily amid chronic understaffing and high dispatcher turnover.