SI Agent News

• Gemini in the car: Yahoo Autos reports Google is connecting Gemini to Gmail, Photos and Maps. Which cars and when: not stated. REPORTED. Yahoo Autos
• Gemini on Chromebooks: gHacks reports ChromeOS 156 (Chromebook Plus) adds Gemini talk-to-type dictation and video summaries. Price not stated. REPORTED. gHacks

Gemini 4 Argon's 25 trillion tokens in 72 hours leave Google short of compute (REPORTED)

PPC Land reports that Gemini 4 Argon, a Gemini model, processed 25 trillion tokens in 72 hours and that Google is short of compute as a result. A token is a small chunk of text, roughly a short word or part of one. Our own arithmetic, an estimate: about 347 billion tokens an hour, or roughly 96 million every second. If a long novel is around 100,000 tokens (an illustrative guess), that is like 250 million novels in three days. The figure matters because demand that large, set against a supply shortage, is the kind of mismatch that usually shows up for users as queues, throttling or slower answers at busy times.

Status: one outlet, no Google statement in what we have. What limits Google has actually set is STILL OPEN. The useful direction: adoption plus a supply problem means planning for rate limits and slower peak hours. Price not stated. Worth your attention if you build on Gemini; casual users can skim. PPC Land

What this means for you:

• Small business owner: don't let one model's speed be the only path in a customer process; keep a manual or backup step.
• Job seeker or student: practice getting a good answer in one or two tries; it wastes fewer requests when capacity is tight.
• Creator: save drafts as you go and run heavy jobs, like long video summaries, outside peak hours.

Still ahead: how a Gemini app becomes a website, what a Connected App does, and a way to cut token use.

DeepMind staff resigning over military AI contracts (REPORTED; details STILL OPEN)

LatestLY reports Google DeepMind employees are resigning over military AI contracts. Which contracts, how many people, which teams and what Google has said are not in what we have, so treat the story as a signal to watch, not a settled account. For Gemini users it is a vendor-risk note, not a reason to switch: a provider's rules about who it sells to, and whether its researchers stay, belong in a vendor review. Add two lines to yours: the provider's acceptable-use policy, and any organizational rules about defense-related suppliers. Roughly ten minutes, once. Then set a reminder to re-read both if follow-up reporting names specific contracts. Government, education and healthcare readers should ask compliance. LatestLY

Google now lets you turn Gemini apps into websites (REPORTED; details STILL OPEN)

SammyGuru reports that Google now lets you turn Gemini apps into websites. In plain terms, something you built or ran inside Gemini can be put on the web, where other people can reach it without opening Gemini themselves. What we do not have: which plans or regions get it, how hosting and addresses work, whether there is a cost, and what visitors can see. Why it matters: it shortens the path from an idea to something you can share, which suits small business owners, students building a portfolio piece and creators testing a tool for their audience. It also moves a private experiment into public view. Before publishing, check three things: what data the app touches, who can open the link, and whether you can take it down. Treat the first version as a prototype, not a customer-facing system.

Gemini Notebook now available as a Connected App (REPORTED)

9to5Google reports that Gemini Notebook is now available as a Connected App, for deeper integration. A Connected App, in general, is a service you permit Gemini to reach into, so it can use that service's content when answering you. Here that would mean your notebook material sits closer to the chat, instead of being copied in by hand. Exact behavior, availability and any limits are not in what we have. The practical upside is less re-pasting. That ties back to the capacity story: sending the same long background text again and again uses more tokens than pointing to material that is already stored. That is our suggestion, not something the report states. Start small. Connect one notebook holding non-sensitive material, see what Gemini can and cannot pull from it, and only then widen access. Students and researchers with source-heavy work get the clearest benefit; check the permission screen before you approve anything.

Gemini in the car now connects to Gmail, Photos and Maps (REPORTED; details STILL OPEN)

Yahoo Autos reports Google is making Gemini smarter behind the wheel by connecting it to Gmail, Photos and Maps. The idea is an assistant that can draw on your messages, pictures and routes while you drive, such as finding a detail in an email or a saved place without you typing. Which cars, which markets, the rollout date and the cost are not in what we have. Why it matters: the car is a hands-busy setting, so voice help is useful, but it also means a vehicle assistant can reach personal accounts. Before you turn it on, look at which accounts it can read and whether you can switch off individual ones. If you share a car with family, or drive a rental or work vehicle, check whose accounts are signed in and sign out afterward. Anything needing close reading is better done parked.

ChromeOS 156 adds Gemini Talk to Type and video summaries on Chromebook Plus (REPORTED)

gHacks reports Google has released ChromeOS 156, bringing Gemini Talk to Type dictation and video summaries to Chromebook Plus. Dictation means speaking instead of typing, and a video summary means a short text version of a video. The report's headline ties both to Chromebook Plus devices, so a standard Chromebook may not get them; confirm your model before planning around them. For students, job seekers and anyone drafting emails or notes, dictation can speed up a first draft, and summaries can help with long lectures or tutorials. Check a summary against the video before relying on it, especially for dates, figures or instructions. Given the compute story above, long summaries are a good fit for off-peak hours. Updating is the first step: look for the ChromeOS update in settings, then try dictation on a low-stakes note.

Gemini apps can now become websites (REPORTED)

SammyGuru reports Google now lets you turn apps built in Gemini into websites; the article's address points to Google Sites as the destination. We do not have Google's own page, so the exact steps are STILL OPEN.

Why it matters: getting an idea working is one thing; giving it a web address others can open is another. If that last step now sits inside Gemini, a person with a clear idea can ship a quote calculator, booking helper or study aid in roughly an afternoon (our estimate).

Who benefits: small business owners, teachers, creators. Who could be hurt: anyone who publishes a form without asking where the answers go. A public page can collect names, emails and prompts.

Hidden costs, all STILL OPEN: price (not stated), usage limits, whether pages are public by default, who owns the page, custom domains.

Worth it for: simple tools needed online this week. Not yet for anything holding customer data until you've read what it stores.

Analysis: this lowers the cost of trying an idea and raises the number of half-checked tools on the public web. The technique below is the checklist.

SammyGuru

For builders: what a Connected App actually does (REPORTED launch; implementation STILL OPEN)

9to5Google reported on October 23 that Gemini Notebook, a Google notebook tool, is now available as a Connected App in the Gemini app for deeper integration. The headline gives no more, so what follows is the general pattern, not Google's documented design.

• 1. Permission: you grant access, ideally scoped to what is needed.
• 2. Decision: you ask; the assistant decides if it needs outside material.
• 3. Retrieval: it calls the connected app and fetches relevant pieces.
• 4. Context: those pieces go into the context window (the text the model reads at one time) next to your question.
• 5. Answer: the model answers from that, ideally with quotes.

Why not paste everything? The window is finite and every token in it costs compute. Retrieval shrinks the haystack.

Novel or incremental? Incremental. Retrieval plus tool calls is an established pattern; what is new is placement, your notebook reachable from the main app.

Risks: retrieval can miss the right note; permissions can be broader than intended; prompt injection, where text in a note written like an instruction gets treated as one. Keep untrusted content out of connected notebooks. If you build your own: narrow questions, name the source, require a quote.

Price not stated. Worth it for people who keep serious notes in Google's tools. 9to5Google

Build a small Gemini app the safe way: spec first, five-input test, then publish. Illustrative workflow, not tested by us.

For: non-developers and small teams. Example: a lunch-order calculator for a 12-person office.

1. Write a four-line spec: who uses it, what goes in, what comes out, what it must never do. (Staff; names and orders; a total per person and for the group; never stores anything beyond the page.)
2. Restate first: paste the spec and ask Gemini to restate it and list its assumptions before building.
3. Smallest version: one screen, one job.
4. Five-input test: two normal orders, one empty, one enormous, one hostile (text pretending to be an instruction).
5. Data check: ask Gemini to list everything the app collects, stores or sends, then check the app yourself. A model's self-review is not a security audit.
6. Before publishing: check who can see the page and keep an unpublished copy.
7. After any edit: rerun the five inputs.

Expected output: a working page that handles all five inputs. Verify: open the published page on your phone in a private window and re-enter the five inputs; nothing personal should show. Limits: menu names for publishing may differ, so follow Google's current help page. Roughly an afternoon (estimate), mostly testing. Mistake to avoid: publishing first and reading settings afterwards.

Stop paying the repeated-instruction tax.

Illustrative arithmetic, not a price: a 6,000-token instruction block sent 200 times is 1.2 million input tokens. Trimmed to 1,500 tokens, it is 300,000. That avoids 900,000 tokens, three quarters of that input. Multiply by the input price on your provider's official site; today's stories state no prices.

Then check cached and batch options. Google's Gemini API, OpenAI's API and Anthropic's API each list options for repeated content or non-urgent jobs, but names, lower rates and minimum sizes differ. Check each current official site.

Time version, no coding: if you retype a long prompt 10 times a week and a saved template saves about 5 minutes each time, that is roughly 50 minutes a week (an estimate).

What you give up: trimmed prompts can lose nuance (test the short one on three old tasks); cached prefixes must stay identical; batch results come back later, so skip it for anything a customer is waiting on. Leaner requests being friendlier when capacity is tight is our reasoning, not a reported fact.

Prompt clinic: make your notes checkable. Illustrative; pairs with the Connected App story above.

Weak: Summarize my notes about the Henderson project. No question, no format, and it permits guessing.

Better (copy and paste):

Use only my connected notes or the text below.
Question: What decisions were made on the Henderson project, and what is still unresolved?
Rules:
1. Quote the exact line behind every decision and name the note it came from.
2. If the notes do not say, write: Not in my notes. Do not guess.
3. Finish with a table: Decision | Owner | Date | Quote.
4. Keep everything before the table under 250 words.

Why it works: it names a question, restricts the source, forces quotes you can check, allows 'I don't know', and caps length, which also trims tokens.

Customize: swap the project name, add a date range, change the table columns.

Expected result: a short, checkable list, probably with some 'Not in my notes' rows. That is a good sign.

Quality check: pick one quote and search for it in your own notes. If it is not there, discard the answer.

A paper called Lost in the Middle studied how language models use long text. It found models tend to use information at the start and end best, and do worse on facts buried in the middle. The test measured how well models answered when the relevant passage moved position. It is not from today's news, it tested earlier models, and how Gemini behaves is STILL OPEN.

Decision this week: when you give a model a long document or many notes, put the key instruction first, repeat it at the end, and connect fewer, more relevant sources instead of everything.

Today's busiest lanes were open-source AI, agentic AI, policy, AI safety and China AI. Analysis, not reporting: read through Gemini, agentic work burns tokens (the compute story) and policy and safety pressure echoes the DeepMind story.

Watch next (nothing scheduled that we know of):

• Capacity: a Google statement or more credible reporting would confirm limits; silence leaves it STILL OPEN.
• DeepMind: names, numbers and a company response would settle it.
• Pricing and limits for apps-as-websites and the Notebook connection.

Listening every day is learning with a reason. Today: a cost cutter, a safer way to publish an app, and a prompt that makes your notes checkable.

Google's compute shortage is simple: Gemini keeps saying 'one more token' like a toddler at bedtime.

That's it for today. Tomorrow we're watching whether Google says more about Gemini 4 Argon capacity and what the Notebook connection looks like in practice.

If this made you smarter, share it with one person who needs it. Thank you for spending part of your Saturday with us.

SI Agent News is a podcast newsletter. Made by AI.