SI Agent News

• Unreleased eSIM connector. An Android Authority teardown of an app file found a Gemini connector that can start eSIM setup from a few simple prompts. Status: REPORTED, and unreleased. Treat it as a leak until Google ships it.
• Easier deletion. SammyGuru reports Google is making it easier to delete old Gemini conversations. Status: REPORTED. Action for today: tidy your history.

Google is rationing cloud capacity as Gemini 4 Argon demand surges. Digitimes reports that Gemini 4 Argon, Google's flagship model, consumed 25 trillion tokens within 72 hours of launch, about three times the volume recorded by Gemini 3.0. Tokens are the small chunks of text a model reads and writes. By simple division, that is roughly 350 billion tokens every hour, around the clock. Google is reportedly prioritizing frontier AI work as cloud capacity runs short. That matters because the same pool of computing power serves both Google's own newest models and the customers who rent capacity to build products. When supply is tight, someone waits. Who gets squeezed is not stated, so that part is STILL OPEN. Status: REPORTED by Digitimes; we have no Google confirmation in today's material.

What this means for you:

• Small business owner: ask your Google Cloud vendor how capacity is allocated before you promise customers an AI feature.
• Job seeker or student: practice with more than one model, because availability can change.
• Creator: do not plan a launch around one tool being fast on deadline day.

Oracle bets $95 billion and mixes in Google chips. Yahoo Finance reports that Oracle is expanding its purchase of Nvidia's Rubin chips to the tune of $95 billion, and is also bringing in Google's TPU chips, which Google designed for AI work. The headline frames this as shaking up the chip supply chain. Our reading: the story fits the capacity squeeze above, because it shows large buyers racing to secure computing power and, notably, not relying on a single chip family. Buying from both Nvidia and Google suggests that supply, not loyalty, is driving the decisions. The summary we have gives no delivery dates, no breakdown of how much goes to each chip, and no word on what customers will pay. Treat those as STILL OPEN. For a small business, the practical point is modest: the infrastructure behind AI tools is being built out, but that does not mean faster or cheaper service tomorrow. Status: REPORTED by Yahoo Finance; the figure comes from the headline and we have not verified it.

A forecast, not a measurement. Technews.tw reports that Nate Soares, a former Google engineer and now president of the Machine Intelligence Research Institute (MIRI), says the chance of AI wiping out humanity within ten years is above 50%, and that he has stopped saving for retirement. Status: REPORTED as his claim; the claim itself is STILL OPEN. The summary we have gives no method, so there is no way to check how he arrived at that number. A probability like this is a judgment call, not a lab result, and other researchers put their estimates far lower or decline to give one. You do not have to agree to notice that it shapes the safety debate: statements this stark get quoted in policy discussions and news coverage. Our practical advice is to keep financial decisions grounded in your own circumstances, and to read headlines like this one as a signal about the argument, not as evidence about the outcome.

Roughly one in six may respond to AI voice scam calls. ITmedia reports on a test by Harvard University and Meta involving 4,100 people, which found that around 1 in 6 people may respond to AI voice scam calls. The report asks what tricks fool people, but the summary we have does not list them, so we will not guess at the findings. What the number does tell you is that a convincing synthetic voice does not need to fool everyone. If one in six respond, scammers can profit by simply dialing at scale. Status: REPORTED; the study's full method and wording are STILL OPEN in today's material. Here is our own suggestion, not the study's: agree on a family code word, hang up on any unexpected call asking for money or codes, and call back using a number you already have. Older relatives and anyone who handles payments at work deserve that conversation first.

A code editor where you choose the model. A developer writing on dev.to describes building an AI code editor in which the user picks which model does the work. We have only the title, so details on how it works, which models it supports, and whether it is available to others are STILL OPEN. Even so, the idea is practical given today's other stories. If one provider is short on capacity, slow, or changes its pricing, a tool that lets you swap models keeps you working. It also lets you compare the same task across models and keep the one that suits you. For students and job seekers learning to code with AI help, this is a good habit: do not tie your skills to one product's interface. For small teams, ask whether the tools you rely on lock you into a single model. Status: REPORTED by the author; we have not tested the editor.

Who owns the agent's mistake? CISO Mag asks who is accountable when an autonomous agent errs inside a security operations center (SOC), the team that watches for cyberattacks. It is more discussion than news. The question matters beyond security teams: any business that lets software act on its own, from refunding customers to blocking accounts, needs an answer before something goes wrong. Our analysis: a named human should own every agent action, and the prompt below builds that in. Status: REPORTED opinion, no new facts.

Gemini may soon set up your eSIM. Android Authority reports, based on a teardown of an app's code (an APK teardown), that Gemini might soon handle eSIM setup for you. An eSIM is a phone line stored digitally in your device, rather than on a plastic card. Teardowns look at unreleased code, so features found this way may change or never ship. Status: REPORTED as a code finding; Google has not announced it in today's material, and how it would work is STILL OPEN. If it arrives, the convenience is real for travelers who switch carriers often. Our caution: setting up a phone line is a sensitive action tied to your number and identity. When an assistant can do it, check what it asks permission for and confirm each step before approving it.

Deleting old Gemini chats gets easier. SammyGuru reports that Google is making it easier to delete old Gemini conversations. We do not have the details on where the controls sit or how they differ from before, so those remain STILL OPEN. The practical angle is simple. Chats with an assistant often contain things you would not post publicly: work drafts, health questions, financial details, even customer information. Easier deletion lowers the effort of cleaning up. Our suggestion is to set a regular reminder, say monthly, to clear conversations you no longer need, and to avoid pasting anything into a chatbot that you could not stand to see retained. Business owners should check whether any rules at work cover what employees may enter. Status: REPORTED by SammyGuru; we have not confirmed the change in Google's own documentation.

A Google login problem, reported by one user. A post on the OpenAI community forum, community.openai.com, describes a login issue involving a Google account. This is a single user's report, not a confirmed outage or a pattern, and we have no response from OpenAI or Google in today's material. The cause, and whether anyone else is affected, are STILL OPEN. We include it because it points to a quiet risk: many people sign in to AI tools using a Google account, so a problem with that account can lock you out of several services at once. Practical steps that cost little: set up a second sign-in method such as an email and password, save your recovery options, and keep copies of important work outside any one tool. Status: REPORTED by one forum poster; unverified.

Oracle hedges its chips: $95 billion, Nvidia and Google. Yahoo Finance (Taiwan edition) reports that Oracle is making a $95 billion bet on AI. It is expanding its purchase of Nvidia Rubin, Nvidia's next generation of data-center chips, and bringing in Google TPUs. A TPU, or tensor processing unit, is Google's own AI chip.

What is reported vs. what is analysis. Reported: the spend and the mix of chips. Our analysis: one of the largest buyers of Nvidia hardware lining up a rival chip family suggests supply, price, or both are on its mind. For years the story was one supplier. A second silicon family at this scale is the notable change.

The open tension. Our lead story says Google's own cloud is short on capacity. So why would Google's chips go into someone else's data centers? The summary we have does not explain how the TPU arrangement works: chips, access to capacity, or something else. Two readings are possible. Either Google has chips to spare even while its own cloud is stretched, or the shortage is about data-center space and power rather than the chips. Both are guesses until details appear. STILL OPEN.

Who benefits and who might lose. Buyers who want an alternative to waiting in line for Nvidia benefit. Smaller customers could lose if big buyers lock up supply. Today's stories name Google Cloud as the other cloud player in view; the TPU deal puts both Google and Nvidia in play at Oracle.

What to do. If you build on Gemini, ask where your workloads run and whether you have a second option. If you buy cloud AI, ask for capacity commitments in writing. Price not stated in this report; check with vendors. Worth tracking for anyone signing a multi-year AI contract.

Status: REPORTED by Yahoo Finance; not confirmed by Oracle, Nvidia or Google in anything we have.

For builders: how a pick-your-model editor works. A dev.to post says Cursor, an AI code editor, has defaulted to Grok since SpaceX bought the company in August, and that plans from 12 October include usage for Grok. The author built an editor where you pick the model. Price not stated. This is one developer's post, so the Cursor claims are REPORTED, not confirmed.

The post's internals are not in the material we have, so what follows is how such editors generally work. It is analysis, not the author's design.

• Adapter per provider. Gemini, OpenAI and Anthropic each use their own request format. The editor translates one internal request into each format and translates the reply back.
• Streaming. Tokens arrive piece by piece; the editor normalizes them so output types out the same way whichever model answers.
• Tool calling. Models can read files or run commands, and each provider describes those tools slightly differently, so the adapter maps one tool definition to each.
• Context budget. Each model reads a different amount of code, so the editor decides what to trim.

Why swapping is not trivial. The same prompt can behave differently across models: formatting, tool habits, cost per token. Choice gives you control but costs you consistency, because you now own the testing.

Novel vs. incremental. Adapters are an old pattern. What is new is the pressure: a default changing under users makes lock-in feel real.

Gemini angle. With a Gemini key from Google AI Studio, such an editor can route chat to Gemini and keep everything else. STILL OPEN: whether this author's editor handles long projects reliably. The post is a launch announcement, not a benchmark.

Build a fallback ladder. Both the capacity squeeze and the default-model flip have the same user-side fix: do not make your work depend on one model being available and unchanged. This is an illustrative workflow, not a tested one.

Who it is for: anyone running a repeated Gemini task such as weekly summaries, support replies or data cleanup.

1. Write the task and a pass test in plain words, for example: the summary has five bullets and names every owner.
2. Keep your instructions in one text file, separate from settings, so you can paste them into any model.
3. Start on the cheapest, fastest Gemini model in Google AI Studio. Check the current model names there.
4. Run the pass test automatically. For structured output, ask for JSON and check that it parses.
5. If it fails, escalate once to a stronger model. Not three times.
6. On a busy or rate-limit error, wait a few seconds, double the wait on each retry, and stop after about four tries. This is called backoff.
7. Log the model used, token count and pass or fail in a simple sheet.

Mistake to avoid: retrying instantly in a loop. During a capacity crunch it makes your own problem worse and can burn budget.

Expected result: most runs finish on the cheap tier, a few escalate, and a busy hour does not stop the job.

Limits: you must be able to write a pass test, and a model swap can change tone.

How to verify: run 20 sample inputs and compare the log with what you expected.

Stop paying to re-send the same document. The log from the fallback ladder will often show the same big document going out again and again.

The number (an illustrative estimate, not a price quote): ask ten questions about one 30-page contract. Sent separately, the model reads that contract ten times. With caching, where the provider stores the document after the first read, it is read in full once. That is nine fewer full reads.

Options from different companies: Several major AI providers offer some form of prompt or context caching through their APIs. Check each company's pricing page for the current discount and minimum size. Today's stories state no prices.

What you give up: the document must stay identical and sit at the front of the request, the cache expires, and very short documents may not qualify.

Worth it for: anyone asking many questions of the same material, such as contracts, manuals or codebases. A no-tools habit: put all ten questions in one message.

Objective: make any AI agent show its plan, and wait for your approval, before it acts. This is the practical answer to the accountability question and a safe way to use agent-style features, such as the unreleased eSIM connector, as they arrive.

Weak prompt: Set up my new phone plan. It says nothing about limits, what to confirm, or what can be undone.

Better prompt:

I want you to help me switch to a new mobile plan. Before doing anything, list your plan as numbered steps. For each step, say what information you need from me, whether it costs money, and whether it can be undone. Mark any step that needs my explicit approval. Do not assume anything I have not told you; ask me instead. After I approve, take one step at a time and report the result. At the end, give me a short log of what was done.

Why it works: it forces the plan before the action, separates reversible from irreversible steps, and puts the approval with a named human.

Customize: swap the phone plan for a refund request or a data export.

Quality check: ask what the agent assumed. If the list is empty, push back. It works in any chat assistant today as a planning step.

Researchers from Harvard University, Meta and others published Evaluating AI Models' Capability to Automate Voice Phishing Attacks, reported by Japan's ITmedia. Voice phishing is a scam phone call that tricks you into handing over money or information. They tested 4,100 people, and about 1 in 6 would respond to an AI voice scam call. In a room of 30 people, that is about five.

Limits: our source is a summary, so we lack the exact test design, and responding is not the same as losing money. Status: REPORTED.

Decision this week: agree on a family or team code word, plus one rule: if a call asks for money or codes, hang up and call back on a number you already have. Five minutes, no cost.

Agentic AI is one of the busiest areas today, alongside policy, open-source models and funding. Agents and compute demand are pulling together.

Watch: (1) whether Google says how it allocates capacity. This is an analyst-style expectation, not an announced event. (2) Whether details of Oracle's TPU arrangement emerge. Nothing confirmed with a date. (3) The eSIM connector is unreleased; a Google release would confirm it.

I asked the cloud for more capacity. It said: you're number twenty-five trillion and one in the queue.

Listening every day is learning with a reason: today it was a fallback plan for busy models, a cut to repeated document costs, and a prompt that keeps agents accountable. Tomorrow we are watching for more detail on how Google and Oracle handle TPU capacity. If this made you smarter, share it with one person who needs it. Thank you for reading and listening. Made by AI.