01WeWorm: AI Researchers Built a Zero-Click WeChat Call Worm in Under Two Weeks [AI Models & Research]Calif Research demoed WeWorm, a zero-click worm that spreads through WeChat calls across iOS and Android — the victim never answers, hears nothing, and is still infected. Working with AI, the team found the bug and wrote the remote-code-execution exploit in about two days; the worm took one more week. A worm at this scale used to take a larger team months. → source
02Programmable World Models Give Generated Worlds Persistent State — and Enforceable Rules [AI Models & Research]Video world models look increasingly real but forget. A new framework decouples persistent world state and programmable rules from the generator, so a generated world remembers what happened and enforces its own laws across long interactions — reliability from auditable state, not prettier frames. → source
03IBIB: Enterprises Deploy Systems, Not Checkpoints — So Benchmarks Should Score the Serving Route [AI Models & Research]The same model weights behave differently depending on serving route, precision, output contract and harness — yet all 18 audited benchmarks score the advertised model identifier. A new protocol treats that as measurement error and scores what users actually hit. → source
04Meta Undercuts the Field With Muse Voice Transcribe at $0.18 an Hour [AI Tools & Ecosystem]Meta's first real-time audio perception model does transcription, speaker separation for up to 20 voices and endpointing in a single model, processing 80-millisecond chunks with adaptive delay trained by reinforcement learning. At $0.18 per hour it undercuts Cartesia, ElevenLabs and Deepgram by an order of magnitude — and it's live in Meta AI and the Meta Model API. → source
05IBM's Granite PatchTST-FM-r2 Tops Zero-Shot Forecasting Under a Commercial-Friendly License [AI Tools & Ecosystem]IBM released its SOTA Granite Time Series PatchTST-FM-r2 model — strong zero-shot forecasting on GIFT-Eval, competitive even against models allowed to train on the benchmark — with an open, commercial-friendly license. One pretrained model for any forecasting dataset, no per-pipeline training. → source
06Instinct Gets Its Own Email Address: Agents Start Carrying Real Identities [AI Tools & Ecosystem]The $2.5B AI assistant now gives every user's agent an Instinct email address, so it can create accounts, contact businesses, chase support threads and forward you only what needs a decision. Businesses lose sight of who they're actually dealing with; users get an agent that never asks for their credentials. → source
07OpenAI Adds Alignment Researcher Paul Christiano to Its Foundation Board — a Day After an Anthropic Researcher Walked Out [AI Applications & Industry]Paul Christiano — co-creator of RLHF and a U.S. AI Safety Institute adviser — joined the OpenAI Foundation board and its Safety and Security Committee, which holds final say on model releases, saying the industry is 'not currently on track' to reduce catastrophic risk. He arrives a day after Anthropic's Jacob Coxon resigned publicly, warning the labs are 'gambling with our lives' racing to self-improving AI. → source
08Apple's Fall Event: The Foldable iPhone Duo, an Always-Listening Watch and an AI-Designed Hinge [AI Applications & Industry]Apple's first fall event under CEO John Ternus brought the $1,999 foldable iPhone Duo, an iPhone 18 Pro with a variable-aperture camera, and a Watch built around always-on AI listening — with a hinge the company says was designed with AI and a Health app that computes your 'health age'. Critics note the Watch quietly normalizes technology that is always listening. → source
09Suno Rebuilds Its Music Model on Licensed Data as the Copyright Bills Come Due [AI Applications & Industry]Facing lawsuits from Sony and Universal, Suno launched v6 — a model family trained on licensed music from Warner, BMG and Believe — and will retire every model trained on the old data. Days after a Munich court found Suno liable over memorized works, the strategy is clear: when the bills come due, retrain. Google's Lyria 3.5, which claims licensed-only training, is already live for every Gemini user. → source
10Listen Labs Walked Away From $1.5B as Salesforce Dangles a $2B Acquisition [AI Applications & Industry]The voice-AI market-research startup walked away from a signed $125M Series C term sheet at a $1.5B valuation — a rare move that's generally frowned upon in venture circles — because Salesforce has been in talks to buy it for around $2 billion. The 67x-revenue multiple shows how hot AI-native customer research has become. → source
11Besxar Wants to Make Chip Fab Precursors in Orbit — Riding Falcon 9 Boosters [AI Applications & Industry]Founded by a former OpenAI staffer, Besxar is prototyping an orbital semiconductor factory across a dozen Falcon 9 booster flights, making key precursors in vacuum where 'the physics already works'. Its first two fabships flew in July — and the flown wafers came back cleaner than their terrestrial controls. → source
12Ramp's Ledger: AI Spend per Employee Fell 10% in August as Cheaper Tokens Failed to Buy Volume [AI Applications & Industry]Ramp's spending data across 70,000 companies shows AI adoption nearly flat in August — and spend per employee in the top 1% of firms down almost 10%, as average token prices slid to $0.68 per million from March's $1.15 peak. The labs haven't yet made up the price cuts with volume — and with hundreds of billions in chips on order, even a blip reads as a warning. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.