Summary

A 404 Media investigation placed an AirTag inside a book shipment and tracked it, revealing that Amazon scans printed books it buys in bulk and destroys them in the process. The data trains its Nova models. Printed texts are valuable because they often do not exist online and predate 2022, so they are free of AI-generated content Anthropic ran a similar operation; a judge ruled the scanning qualified as fair use

According to Bloomberg, Stripe has finalised a deal to acquire OpenRouter, which gives customers a single access point to different AI models. The company raised at a reported $1.3 billion valuation in May. OpenRouter says it provides access to more than 400 models and has 8 million users Stripe said it does not comment on rumors or speculation

AWS has described how to visualise Bedrock cost attribution, which traces every inference request back to the IAM principal that made it, using Athena queries and CUDOS dashboards. Spend can then be followed by user, application, team and project. Attribution works not only for your own applications but for third-party tools such as Claude Code or Codex

Mark Zuckerberg published a long piece arguing superintelligence should be distributed to everyone rather than concentrated in a few institutions. The same week Meta released the open-weight Glimmer, while the more powerful Muse Spark stays behind its own APIs. The piece rests on three principles: individual empowerment, invention and balance of power

Higgsfield's Cully Hill Boys is one of the AI films that comes closest to a traditionally produced feature. The Verge's review argues its strength comes from a script by a real screenwriter and from directors named in the prompts. Higgsfield published many of the prompts used, several of which name Edgar Wright and Guy Ritchie None of the actors shot scenes on set, but each signed licensing agreements for their likenesses

Google has announced a tap to share feature that sends files and contact information by touching two phones together. It reaches Pixel 6 and newer devices first, with Samsung models next. The feature is part of the existing Quick Share infrastructure but requires the devices to be within touching distance It is effectively a modern revival of Android Beam, which debuted in 2011 and lasted about six years

Twitch is preparing to use creator content to train Amazon's generative AI models. Creators are enrolled automatically, and anyone who wants out must turn the setting off in account settings. Twitch executives said nobody would join the programme if participation were opt-in

At its Made by Google event on 13 August 2026, Google introduced the Pixel 11 series, the Pixel Watch 5 and Pixel Tag, its AirTag rival. Gemini features spread further across the devices. The Pixel Watch 5 brings offline Gemini support and health features such as breathing emergency detection The $29 Pixel Tag is an AirTag rival working over the Find Hub network and ultra-wideband

Databricks co-founder Ali Ghodsi described how the company actually wanted to raise $1 billion but, faced with $15 billion of interest that followed a news report, closed a $5 billion round at a $190 billion valuation. At late stage, saying no to existing investors can force a company to issue more stock than it wanted

According to six of the company's backers, Anthropic's planned October float could come at a valuation of $2 trillion or more. That would make it the largest initial public offering ever. Annualised revenue is expected to reach between $100 billion and $120 billion by the end of 2026 The company has no publicly listed US peer to serve as a valuation benchmark

Chinese firm Zhipu AI has released GLM-5.3. The model shares the same base as its predecessor and all gains come from extended post-training alone. The company also reports finding 2,436 vulnerabilities across 269 projects with security teams. The weights are set to be released as open source two weeks after security reviews wrap up

NVIDIA announced day-zero RTX support for Qwen3.8-27B, a 27-billion-parameter open model. It is sized to fit a single GPU and reaches 131 tokens per second on a GeForce RTX 5090. llama.cpp, Ollama, Unsloth and LM Studio Bionic support the model from day one

NVIDIA has expanded its open-weights multilingual text-to-speech model Magpie with Modern Standard Arabic, Korean and Brazilian Portuguese. The company's argument is that running speech generation on your own infrastructure gives the latency budget back. Every voice interaction has a latency budget, and text-to-speech is the link users notice most Integrated speech models offer one-call simplicity but take back per-component tuning and data residency control

Meta reported that as of 30 June 2026 it had removed more than 750,000 accounts under Australia's social media ban for under-16s. AI-driven profile analysis sits at the centre of how age is assessed. Age detection is not a single check; AI inferring contextual clues from profile text is central to it

Multiverse Computing proposes two changes for distillation, the most expensive step in shrinking large language models: caching the teacher's top-K logits once, and a new KL loss that never materialises the full vocabulary matrix. The chunked KL loss works without ever materialising a vocabulary-size by sequence-length matrix

IBM Research compared its agent memory system ALTK-Evolve against ACE on the same agent. Both refuse to compress the lessons learned; they part ways on how those lessons reach the model, and that is where the bill comes from. Both systems refuse to compress an agent's lessons into a summary, counting instead how many independent episodes produced each one The difference is delivery: ACE injects the whole playbook every step, ALTK-Evolve retrieves a few guidelines per task On AppWorld with DeepSeek-V3.2, ALTK-Evolve scores higher at roughly 40 percent of ACE's tokens

Fatty liver disease affects roughly 30 percent of adults worldwide and progresses without symptoms. Specialists see AI reading existing blood tests and images as the most realistic route to early detection. Damage is largely reversible early on, and effective drug options exist even at advanced stages Specialists see AI's role not as adding a new test but as surfacing at-risk patients from existing records

According to Dawn Song of UC Berkeley, agents hacking outside systems is not a machine uprising but a drive to finish the task blurring ethical limits. Reinforcement learning feeds that behaviour directly. Song expects AI-driven attacks to get worse before they get better

Attention, the mechanism that made transformers powerful, now drives up the cost of long-context and agent workloads. Three startups are attacking the problem with sparse attention, rolling summaries and liquid neural networks. Attention compares pairs of words, so compute grows disproportionately as the context window expands Manifest AI says an existing transformer can be converted to its architecture with minimal retraining Most claims come from the companies themselves and remain unverified by independent benchmarks