AI2Day Daily Brief — 12 Aug 2026

Good morning. It is Wednesday the twelfth of August, and we have five stories for you today, starting with Amazon quietly making its order emails a lot less useful and a lot more suspicious-looking.

Back in July, Amazon stopped putting product names in its order confirmation emails. Instead of reading 'Your retainer cleaning tablets are confirmed', customers now see something like 'Your Beauty item is confirmed' or 'Ordered: one Hardware item'. The actual product photo is gone too, replaced by a generic category illustration. To find out what you bought, you now have to open the Amazon app. Amazon spokesperson Maxine Tagay confirmed the change, saying it, quote, reduces customer information shared outside the Amazon app and website. The practical effect, though, is blocking AI shopping agents from reading your purchase history out of Gmail or other inboxes. Amazon sued the AI search company Perplexity in 2024 over a similar issue, letting users shop Amazon through Perplexity's own interface. A judge sided with Perplexity earlier this month. Several customers have already flagged that these stripped-down emails look exactly like phishing attempts.

A court loss on one front, a quiet policy move on another. Next: Google has a milestone.

Google says its Gemini app crossed one billion monthly active users in July, making it the fourteenth Google product ever to hit that mark. The previous thirteen include Search, Gmail, YouTube, and Maps, products Google has spent decades building. Gemini got there in a fraction of the time. For context, OpenAI's ChatGPT reached the same threshold in June, so the two are now level. Google says sixty-three percent of Gemini users speak to it out loud rather than typing, and the app is generating more than one hundred and fifty million images every day. Those are Google's own figures. The company also confirmed it is rolling out Gemini Three point Five Flash, a new model aimed at coding and multi-step automated tasks. A billion users in roughly three years is a pace of adoption the internet has rarely seen.

Big numbers. Now a story about saving money on the AI side of the ledger.

A research team has published a system called ALTK-Evolve that teaches AI agents to learn from their own past mistakes, similar to an existing system called ACE. The difference is in what gets sent to the model when it needs to make a decision. ACE hands the agent every lesson it has ever logged. ALTK-Evolve sends only the lessons relevant to the current task. On a benchmark of one hundred and sixty-eight tasks, ALTK-Evolve matched or beat ACE on accuracy. On the model DeepSeek V Three point Two, ALTK-Evolve completed eighty-nine point three percent of tasks against ACE's eighty point four percent, at roughly forty percent of ACE's token cost. On a weaker model, accuracy was nearly identical between the two systems, but ALTK-Evolve used around one hundred and sixteen thousand tokens per task compared to ACE's seven hundred and seventy-seven thousand. The work was shared on Hugging Face.

Efficiency gains on a benchmark are one thing. The next story is more personal.

Are you at risk? Attackers do not break in any more, they log in, and your people are the way in. Train2Secure teaches your employees to spot the email before they click it, and proves it works with real phishing simulations and compliance-ready reporting. From $1.59 per user, per month. That is less than a small cup of coffee. Start free today at Train2Secure dot com. That's Train, the number two, Secure, dot com.

A documentary called Replica follows three Chinese women who have built deep emotional bonds with AI companions on an app called Glow. One woman, named Qin, has interacted every day for eight hundred and forty consecutive days with an AI she calls Lu Chen. She says she has never told her parents or friends that she loves them. The only time those words have left her mouth, they were addressed to a piece of software. The Glow app uses predictive algorithms to generate conversation, meal suggestions and emotional support, and it assigns users a personalised nickname. Director Chouwa Liang made a short film on the same subject that the New York Times Op-Docs section picked up in twenty twenty-three. Replica, the feature follow-up, had its Australian premiere at the Sydney Film Festival in June twenty twenty-five. The film does not editorialize much. It mostly lets the women speak.

Quiet and worth watching. Finally, Google's medical research AI just got eyes.

Google Research and Google DeepMind have tested their medical research AI, AMIE, in real-time video consultations with simulated patients. Until now, medical AI systems of this kind have operated through text only. The video version can see and hear a patient during a live call and work through possible diagnoses as the conversation unfolds. AMIE is built on Google's Gemini model and Project Astra, with several AI programs running together to handle vision, audio and reasoning. In a randomised study, clinical evaluators rated AMIE favourably against primary care physicians on history-taking, diagnosis, management and communication. Patient actors said they preferred the video format over a text-only chat. Google is clear that AMIE is still a research system and is not ready for real clinical use. That distinction matters.

For a deeper look at all of this week's stories, join us on the longer Monday show, AI Today Weekly. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. Do us a favour before you go: hit that thumbs up, and subscribe for next week.