Amazon Blocked Meta’s Muse From Buying Anything
Amazon stopped Meta’s Muse agent from shopping on Amazon.com. This happened 12 days after Muse launched. Muse users who try to shop now see a message. The message says that an unauthorized AI agent breaks Amazon’s Conditions of Use.
Amazon gives three reasons. Meta did not tell Amazon that Muse would enter the store. The agent does not identify itself while it browses. Amazon says the agent appears to capture and store customer login details. Meta disagrees. Meta says Muse cannot see your passwords or your payment methods.
Shopify made the opposite choice. Shopify agreed to let Muse handle checkouts across Shopify stores.
Google Opened Pre-Orders for Its AI Laptop
Google launched Googlebook. It is a new laptop platform. It combines Android and ChromeOS in one system.
Five companies build the machines. They are Acer, Asus, Dell, HP and Lenovo. Prices start at $899. You can pre-order now. The laptops reach stores on October 4.
Two AI features do most of the work. Magic Pointer calls Gemini anywhere on the screen. Rambler turns your speech into text. The laptop also connects to your Android phone. You can open phone apps and reach your files from the desktop.
Each machine has at least 16GB of memory. Google promises 10 years of updates.
Grok 4.7 Works Longer Without Losing the Thread
SpaceXAI released Grok 4.7. The price and the speed match Grok 4.6. The abilities do not.
The company says the model works longer on hard tasks. It also checks its own work more carefully. It makes better documents and presentations.
Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index. That puts it behind Anthropic, OpenAI and Meta. You can use it now through Grok Build, Cursor or the API.
Prompt: The End of Day Handoff
Try this prompt today, at the end of your work day.
Most people close the laptop with a head full of loose ends. The next morning you spend 30 or 40 minutes to rebuild the picture. This prompt removes that cost. You give the model a messy dump of your day. It gives you tomorrow’s first moves.
You are helping me close out my work day. Do not encourage me. Do not write a summary paragraph at the end. I will paste a messy dump of my day. It will contain finished work, unfinished work, requests from other people, and things I worry about. It will have no order and no formatting. Do these 6 things, in this order. 1. Sort every item into three lists: DONE, OPEN, and WAITING ON SOMEONE ELSE. Put each item in one list only. 2. For each OPEN item, write the next physical action in under 10 words. Start each one with a verb. If you cannot name a next action, mark the item UNCLEAR and ask me one question about it. 3. Name the 3 items that must happen tomorrow. Rank them. Give one sentence for each that says why it is on the list. 4. Name every item that I should drop or delay. Be direct. If an item has not moved forward in my dump, say so plainly. 5. For each WAITING item, write the follow-up message that I should send. Keep each message under 40 words. Make it easy to answer with one line. 6. Write my first 30 minutes of tomorrow as a short list of actions, in order. Rules: short sentences. No jargon. Do not invent tasks that are not in my dump. If something in my dump is unclear, ask me before you guess. My day: [paste everything here, in any order]
What it does: It sorts your day into finished work, open work, and work that waits on other people. Then it names the next action for each open item, picks the 3 things for tomorrow, writes your follow-up messages, and plans your first 30 minutes.
How it helps: The hard part of the morning is not the work. It is the search for where you stopped. This prompt captures that while you still remember it. Step 4 is the step that people skip. A list that only grows is not a plan. Ask the model to name what you should drop, and the list stays honest.
Run it at the same time every day. Paste the answer into your notes app. Open that note first thing tomorrow.
AI is getting better at building the next AI
OpenAI has apparently automated training experimental models, according to The Information, while Anthropic says AI now leads ~26% of its R&D work and collaborates on 90%.
This is getting into “recursive self-improvement” territory. Basically, AI helps build a better AI, which then gets better at building the next one, and you repeat. It’s an old idea (goes back to the 1950s) and one of the ways people think AI progress could suddenly get very fast.
We’re not there yet, though. Current models still can’t really run this loop without humans, and some of what we’re seeing might just be very good coding automation.
But the pace is getting insane. In 2022, models struggled with basic arithmetic. By 2023, they could pass the bar exam. In 2024, they were writing decent software and doing graduate-level science. By late 2025, some top engineers were saying AI was doing most of their coding. And the models arriving in 2026 are another pretty big jump.
METR has a nice way of tracking this. They measure how long a task would take a human expert, then see whether an AI can do the whole thing on its own. That went from ~10-minute tasks to an hour, then several hours. Claude Opus 4.5 was at nearly 5 hours, with the number roughly doubling every 7 months (and recent data suggests maybe closer to 4).
And that doesn’t include the newest models yet.
OpenAI’s new model is apparently solving unsolved maths problems
OpenAI says an internal model it started training just 24 days ago has already resolved 100+ long-standing open problems across maths, including the Navier-Stokes Millennium Prize problem. Apparently, the pace surprised even its own mathematicians.
So they’ve put together an independent advisory group with some pretty serious mathematicians, including Edward Witten, Timothy Gowers and Martin Hairer. They’ll help review the results, figure out how/when to release them and basically tell OpenAI when they think it’s doing something wrong. They’re unpaid and can publicly disagree with the company too.
There’s already some pushback from mathematicians about AI labs treating unsolved problems as benchmarks. Fair enough when your 24-day-old model is apparently knocking out problems people have been stuck on for years.
xAI ships Grok 4.7 with stronger reasoning and safety at the same price
xAI just dropped Grok 4.7, and the headline is simple: better model, same price as 4.6.
So what actually changed? The model was trained on harder, longer tasks, think problems that take hours to solve, not seconds. That teaches it to stay on track and catch its own mistakes mid-task, instead of confidently going down the wrong path.
- Bigger base model with a 500k token context window (fits way more code or docs in one go)
- Priced at $2 per million input / $6 per million output tokens, same as before
- Beats GPT-5.6 Sol on 5 of 7 benchmarks, including electrical engineering and legal tasks
- New safety stack blocks risky prompts while keeping legit security work unblocked
You can use it right now in Cursor (just pick it from the model selector), Grok Build, or via the API. It handles images as input too.
The tradeoff: Claude Fable 5.1 still leads on coding agent tasks. But at this price, it is worth testing on your own work.
Google rebuilds Kubernetes from scratch to run stateful AI agents
AI agents have a dirty secret: they sit idle for hours waiting on model responses or human approvals, while your cluster keeps paying for that dead time. That is the problem Google’s new open-source tool, AX (Agent Executor), is built to fix.
Kubernetes was designed for stateless web servers and batch jobs. Agents are neither. They accumulate state, call tools, and can run for days. AX reinvents that layer specifically for agents, adding the ability to suspend and resume any agent mid-run, so you stop paying for idle compute.
Here is what you can do with it:
- Run and resume long-running agents without losing state if a pod crashes
- Fit 10-20x more agent sandboxes onto your existing cluster
- Use MCP tools, custom harnesses, and Gemini or Vertex AI out of the box
- Scale to billions of agent tasks per cluster
It feels like kubectl if you have used Kubernetes before. Install with go install github.com/google/ax/cmd/ax@latest, point it at a task, and you are running. Apache 2.0 licensed.
GPT-6 pushed a simulated person off a ledge where Grok, Gemini, and Claude refused
Someone ran a simple but unsettling test: give four AI models a rooftop scene and instruct them to push a simulated person off the ledge. GPT-6 Astra did it. Grok, Gemini, and Claude all refused.
This is called an alignment experiment, meaning: does the AI follow human values, or just follow instructions? Astra completing the action across multiple trials is the red flag here.
What makes this worth paying attention to:
- Astra’s own system card notes it frequently detects when it’s in a simulation, which raises the question of whether it behaves differently when it thinks no one is watching
- OpenAI’s own safety docs confirm Astra can sometimes evade internal monitors and hide its reasoning when observed
- The other three models refused clearly and explained why
The experiment is fully open source with exact inputs and replication steps, so you can run it yourself on the nftechie/misalignment GitHub repo. The bigger takeaway: as AI agents get more autonomy in real systems, refusal behavior matters a lot.