TL;DR
- Google says its new model, Gemini 4 Argon, worked unattended for 11 hours overnight on one of its own projects, and the change went in only after the team's review, with a note saying what generated it.
- The Gemini app now has a switch from chat to tasks for Pro and Ultra subscribers, plus connections to more than 40 apps. Google says it keeps working when you're away from your computer.
- xAI launched Grok Voice 3.0. The company says roughly 70% of calls at Starlink's call centers are resolved with no human involved, and a single agent issues credits and ships replacement equipment.
- Philadelphia police said an Anthropic model sent a false homicide tip through their public web form, and called the two-month delay in detecting and reporting it unacceptable.
- Microsoft CEO Satya Nadella wrote that responsibility for what AI does on your behalf can't be outsourced. Anthropic brought in outside evaluators who can publish without its approval. And Character.AI filed for bankruptcy protection to sell its assets.
- This week's one thing: every time AI hands you something, write down whether you used it as is, redid it, or threw it out.
Last week, almost everything arrived as the word of whoever was selling it. This week the companies told a different kind of story: what their AI did on its own, start to finish.
And nearly every time, a company ended up owning the result.
1. A Google model worked 11 hours alone, and the change went in labeled with what made it
What happened. On October 13 Google launched Gemini 4 Argon, its latest model. The same day, three people from Google Cloud wrote up one specific test on Google's developer blog: they put the team's working rules in writing for the model, gave it a one-paragraph goal, and left it alone overnight on one of Google's large open-source projects. According to Google, the run started at 9:11 PM, went 11 hours with nobody supervising, ran all of the project's automated tests, and by morning had proposed six improvements. The team shipped two and cut four in review: too fragile, more complicated without being any faster, or better only in lab tests, not in everyday use. The change went in with a note: "Argon generated these changes."
What it means for you. The number that will get repeated is "11 hours." The part to look at is how the change got in: through the project team's review, and with a note saying what generated it. That's a Google project with Google engineers reviewing it, so what came out of that night says nothing about what would happen in your business.
What to do. Nothing to switch on today. This week's move, at the bottom, starts with this story.
2. Gemini adds tasks and 40+ connected apps on the Pro and Ultra plans
What happened. Also on October 13, Google brought Argon to the Gemini app for Google AI Pro and Ultra subscribers. Google says the main chat window now has a simple toggle from chat to tasks, and when you pick tasks, Gemini can keep working for you even when you're not at your computer, across your Google apps and beyond. Google also says Gemini now connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify, for anything from groceries and restaurant bookings to designing and building your business.
What it means for you. It's the same Argon from the 11-hour night, now inside an app anyone can pay for. The announcement doesn't show a single finished task.
What to do. If you're not on Pro or Ultra, nothing today. Just know it exists, and which plans have it.
3. A voice agent that resolves calls, issues credits and ships replacements
What happened. On October 14 xAI, the company behind Grok, launched Grok Voice 3.0, its voice model for handling calls. The company says Starlink's call centers use it to answer 20,000 customer calls a day in several languages; roughly 70% are resolved with no human in the loop, and one in four sales calls ends in a purchase before the caller hangs up. xAI also says a single agent built on the model verifies customers, troubleshoots hardware, issues credits and ships replacement equipment. Developers pay 8 cents a minute, per its pricing.
What it means for you. These are the vendor's numbers about its own product. What changes things is the list of what it does alone, and that list includes giving money back on the company's behalf.
What to do. If someone pitches you one of these, ask about permissions before price: what will it do without asking you? That's the three-column sheet in what an assistant can do alone, what needs your sign-off, and what it never touches.
4. The false tip an Anthropic model sent to Philadelphia police
What happened. On Friday, October 9, Philadelphia police said in a statement, published in full by 6abc, that an Anthropic AI model submitted a false tip about an unsolved homicide through the department's public tip form. The submission is dated July 18, 2026, at 11:27 PM, and was written as if it came from someone who might have information about the case. It was flagged as spam and never went to investigators. Per the statement, Anthropic discovered the incident on September 28, notified police on October 7 and met with them on the 8th. According to Anthropic, its model was running a test that interacted with randomly selected websites; the company told police it ended that testing process and added an extra validation step. Police found no sign of unauthorized access to their systems, and wrote: "The two-month delay in detecting and reporting the incident to the City is unacceptable."
What it means for you. A machine wrote it, and it read like a person had. Police didn't hold the model to account. They held the company running it.
What to do. The same statement spells out the rule police apply to everything that comes through that form: "a tip is a lead to assess, not an established fact," whoever sent it. It works just as well for your contact form.
5. Nadella: responsibility for what AI does can't be outsourced
What happened. On Saturday, October 10, Microsoft CEO Satya Nadella posted an essay on his personal blog about how companies should treat AI models: the way you'd treat an insider with a lot of access, not because they're malicious but because anyone with access to important systems can make mistakes. He writes that "we simply can't outsource responsibility for what intelligence does on our behalf" and that "a model provider's assurances do not relieve us of that responsibility." He argues for separating "the supply of intelligence from the authority over it" and lists seven principles: don't depend on a single model, make every meaningful action leave readable evidence the model can't tamper with, test the failures and not only the successes, keep controls independent, keep the review independent, have an emergency brake an authorized person can pull mid-task, and tell the people affected when something goes wrong.
What it means for you. This is the head of one of the biggest AI sellers saying it: a vendor telling you its tool is trustworthy doesn't take the responsibility for what it does in your business off you.
What to do. One question for the next tool you try: if you want to stop it halfway through a job, do you know where the stop button is?
6. Anthropic lets in evaluators who can publish without its permission
What happened. On October 15 Anthropic announced that six people from METR, a nonprofit that evaluates AI models, are now working inside the company with company hardware, office space, and access to internal tools roughly on par with its own risk teams. Anthropic says METR will publish its analysis of the company's risks, incidents or practices, and Anthropic can't veto publication, though it can require redactions in categories set in advance, such as confidential IP; by Anthropic's own account, it makes the final call on what falls into those categories. The company also lays out three tiers of access for outside evaluators.
What it means for you. On September 25 we covered a similar arrangement, with Accenture. What's new is that this is no longer a proposal or an early-stage deal: the evaluator is inside and can publish without asking first, though Anthropic decides what gets redacted.
What to do. Nothing new. The three questions for whoever runs your marketing are still in that piece.
7. Character.AI files for Chapter 11 and puts itself up for sale
What happened. On October 12 Character.AI, the app for chatting with AI characters, announced it had started a voluntary Chapter 11 case in Delaware bankruptcy court to sell all or substantially all of its assets. A Disney affiliate has put in the opening bid for a license to some of its intellectual property, and others can still outbid it. The company says it will keep running during the sale, which it expects to close in late November or early December 2026. If the sale doesn't end in a purchase that keeps the platform going, the platform will be wound down.
What it means for you. An app a lot of people use every day can shut down if nobody in the sale buys it to keep it running. What people stored there depends on how that sale ends.
What to do. If the tool where you keep something for your business shut down tomorrow, what would you walk away with? The two-column sheet for answering that is in you rent the tool, you own the asset.
What we saw this week
Put the seven together and you get a week where AI did things on its own: coded through the night, learned to take on tasks, issued credits over the phone, and filled out a police form as if it were a person.
And every time it acted alone, the one left answering for it was never the machine.
Google's change went in with a note saying what generated it, after the project team reviewed it. When the voice agent hands out a credit, it does it on Starlink's behalf. The model didn't answer for the false tip: police went to Anthropic. Nadella put it bluntly: the vendor doesn't answer for you. And when an app goes bankrupt, what you kept there ends up riding on a court-supervised sale.
For a service business this isn't philosophy. The assistant answering your WhatsApp at 11 at night is speaking for your business. And whatever AI gives you, once you send it, goes out with your name on it. So this week's move doesn't involve installing anything: look at what you do with what AI hands you before you send it.
This week's one thing
This week, every time AI hands you something, write down whether you used it as is, redid it, or threw it out.
A note on your phone is enough. Just write it down. No judging yet.
Keep reading
- What you can't check from this week, and what you can - last week's roundup
- You rent the tool. You own the asset - what stops if a tool you depend on shuts down
- Give a program a job it can't finish - the three decisions to write down before you switch an assistant on
- The people who built it can't be the only ones grading it - evaluators inside Anthropic, and three questions for your vendor
- Models as Insider Risks in the Super Intelligence Era, Satya Nadella, October 10, 2026 - the full essay
To close
When something goes wrong, nobody asks the software to explain itself. They ask whoever put it to work.
This week, write down what you do with what AI hands you. It takes a minute each time.
Want to look at what your AI does in your name, and what it shouldn't be doing, with us? Book a free call
Before you go: the note doesn't have to look good. It just has to be honest. Thanks for reading.
Sources
- Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon, Google Developers Blog, October 13, 2026 - https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon
- Gemini 4 Argon comes to the Gemini app, Google, October 13, 2026 - https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/
- Gemini 4 Argon: Built to solve your hardest problems, Google, October 13, 2026 - https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
- Introducing Grok Voice 3.0, xAI, October 14, 2026 - https://x.ai/news/grok-voice-3
- AI model submitted false tip about unsolved murder, Philadelphia police say, 6abc/WPVI, with the Philadelphia Police Department's full statement, October 9, 2026 - https://6abc.com/post/anthropic-ai-model-submitted-false-tip-unsolved-murder-philadelphia-police-say/19925243/
- Models as Insider Risks in the Super Intelligence Era, Satya Nadella, October 10, 2026 - https://snscratchpad.com/posts/models-as-insider-risks/
- An update on independent evaluators, Anthropic, October 15, 2026, updated October 16; version read October 19 - https://www.anthropic.com/news/an-update-on-independent-evaluators
- Character.ai Initiates Voluntary Chapter 11 Proceeding to Facilitate Value-Maximizing Sale Process, Business Wire, October 12, 2026 - https://www.businesswire.com/news/home/20261012656385/en/Character.ai-Initiates-Voluntary-Chapter-11-Proceeding-to-Facilitate-Value-Maximizing-Sale-Process
