SI Agent News

• Geely, cars, 2027: Geely plans to bring Google Gemini, Google Maps and Google Play into its cars starting in 2027. That is a distribution deal, not something to plan around this year. REPORTED by VOI.ID.
• Developer engagement: A report says developers are engaging less with each model improvement. The coverage is thin and mixes in prediction-market odds. STILL OPEN. Crypto Briefing.
• AI and money: A widely shared piece asks whether AI is coming for your money. The fears are familiar and it brings no new facts to act on. The Star Democrat.

Gmail Live: search your email by voice. Google has launched Gmail Live, a voice search feature that uses Gemini (Google's AI model family) to search and work with your email, according to Brazilian outlet Aqui. Status: REPORTED. We have not seen Google's own announcement, so availability by country and plan, and what it can do beyond search, are STILL OPEN. Price not stated.

What changes: today you type keywords and scan a list. The pitch is that you ask out loud, for example 'what did the supplier say about delivery dates', and Gemini does the hunting. Voice will not beat typing for quick lookups. Where it should win is fuzzy memory, when you recall the gist of a message but no keyword. That is the test to run once you have access. Worth a look if you live in a large inbox and search it on the move; skip it if you mostly type at a desk. One caution: a feature that reads your mail by voice is also one that reads it aloud, so check where you use it and what permissions it asks for.

What this means for you (illustrative, until we see what ships):

• Small business owner: pull up the last invoice thread by voice while your hands are full, instead of scrolling.
• Job seeker or student: find that one recruiter or professor message by describing it.
• Creator: dig sponsor terms out of a long thread without remembering the subject line.

Gemini avatars and voice imitation: will it represent you? GoogleWatchBlog, in a video report, describes a new Google AI that creates realistic avatars and imitates voices, and asks whether it could soon represent users. REPORTED, and we have only the headline framing, so what the tool is called, who can use it, how much voice sample it needs and what it costs are all STILL OPEN. The 'represent users' part is the outlet's question, not a confirmed product plan.

Why it matters: if a tool can speak in your voice and show your face, the useful cases are obvious (recorded explainers, translated clips, a stand-in for routine updates), and so are the risks. Anyone who sends or receives voice notes and video messages should start treating a convincing voice as no longer proof of identity. Practical step: agree a code word or call-back habit with family and finance contacts, and never approve payments on a voice message alone. If you try such a tool yourself, read what it does with your voice sample before you upload one.

SynthID detection, a how-to. Engadget walks through Google's SynthID detection website. SynthID is Google's watermarking system for AI-made images, video and audio. REPORTED; price not stated. Our background note, not from the article: a detector like this looks for Google's own watermark, so 'nothing found' means no watermark found, not proof a human made it. Worth bookmarking for anyone who forwards or publishes media.

The practical use sits right next to the avatar story above. Before you share a striking image, clip or audio file, or before you put it in a newsletter, post or client deck, run it through the checker as one signal among several. A positive result is informative; a negative one only narrows things. Pair it with the old habits: look for the original source, check when it first appeared, and ask who benefits from it spreading. Treat it as a quick screen, not a verdict.

A researcher questions superposition, citing 90% of neurons. The Chosun Ilbo, via MSN, reports that an AI researcher is challenging the superposition theory with an insight about 90% of neurons. REPORTED; we have only the headline, so the researcher's name, the exact claim and any evidence are STILL OPEN. Background, not from the article: superposition is a popular idea in interpretability research, the effort to understand what happens inside AI models. It suggests a network packs more concepts than it has neurons by overlapping them.

Why a non-researcher should care: interpretability decides how well we can explain, audit and trust model behavior. If a leading explanation is disputed, expect disagreement on how readable these systems really are. For you, there is no action to take today. The useful stance is patience. Treat confident claims, either way, about what models 'really' do inside as provisional until the work is published and others reproduce it.

One person, 26 delegated tasks. A first-person account on note describes an IT systems employee running a one-person operation with Google AI: 26 tasks, about two dozen, handed to Antigravity (an agent app), including a Gmail cleanup that runs automatically every morning at 7:30. One person's account, so REPORTED at best. It is the pattern behind today's technique.

The takeaway is the method, not the number. Pick chores that repeat on a schedule, such as inbox cleanup, and let an agent run them without you. Start with one low-stakes task, check its output for a week, and only then add another. Keep anything involving deletion, payments or outgoing messages on a review step until you trust it. The author's results are self-reported, so measure your own time saved before copying the setup.

Geely brings Gemini, Maps and Play to cars from 2027. VOI.ID reports that Geely will bring Google Gemini, Google Maps and Google Play to its cars starting in 2027. REPORTED; models, markets, pricing and whether features need a subscription are STILL OPEN. The date is a long way off, so treat it as a direction rather than something to plan a purchase around.

Why it matters: Gemini in the dashboard means the assistant you already use on your phone may follow you into the car, with Maps for navigation and Play for apps. For drivers, the practical angle is data and privacy: a car assistant can hear you and know your routes. If you are shopping for a car in the next few years, ask the dealer what the assistant collects, whether it works offline, and how long software updates are promised. For now, nothing to do except note that Google keeps widening where Gemini lives.

AI model improvements see diminishing developer engagement. Crypto Briefing reports, citing a report, that improvements in AI models are seeing diminishing developer engagement. REPORTED; we have only the headline, so the report's author, the data behind it and which models it covers are STILL OPEN. One reading is that each new release excites builders less than the last, as differences between top models narrow.

Why it matters for you: if the gap between models shrinks, the choice matters less than the workflow around it. That favors a practical approach. Do not switch tools every time a new model launches. Pick one or two you know well, write down the tasks you give them, and re-test only when a release claims to fix something you actually struggle with. Spend your effort on prompts, templates and checks that carry over from model to model.

Is AI coming for your money, or all money? The Star Democrat runs an opinion-style piece asking whether AI is coming for your money—or all money. REPORTED; we have only the headline, so its arguments and evidence are STILL OPEN. The question is broad: does AI threaten personal finances, or the idea of money itself?

For readers, the useful version is narrower and closer to home. Scam calls using cloned voices, fake invoices that read perfectly and AI-generated investment pitches are the near-term money risks, which ties back to the voice and watermark stories above. Practical steps: verify payment requests through a second channel, be slow with anything urgent, and be wary of any tool or advisor promising AI-driven returns. The bigger macro question is worth watching, but your own checks are what protect your money this week.

Gemini, avatars, and the question of who speaks for you

GoogleWatchBlog, a German site, summarizes a video about a new Google AI that creates realistic avatars and imitates voices, and asks whether it could soon represent users. Two cautions. It is a summary of a video, so details are thin, and we have no Google primary source. Avatar and voice creation is REPORTED. Whether it will represent users is STILL OPEN.

Why 'represent' matters (our analysis). An assistant answers you. A representative acts as you: joins the call, leaves the voicemail, replies in your voice. Next to Gmail Live, you can imagine voice becoming an interface in both directions, you speaking to Google's apps and the apps speaking for you. That is analysis, not a stated plan.

Who benefits: people who produce a lot of spoken content, and busy professionals with routine calls. Who is exposed: anyone who treats a voice as proof of identity, such as a call that sounds like your boss or relative asking for money.

Do today, at no cost: agree a code word with family and team; for any request involving money or passwords, call back on a number you already have; for audio or video that matters, run it through a checker like SynthID, remembering its limits.

If representation features arrive, ask three questions: does the other person see a disclosure? Can you switch it off and see what was said in your name? Is the output watermarked? Those are our questions, not Google commitments.

For builders: how a network stores more ideas than it has neurons. This is background from the field, not today's reporting, except for one essay at the end.

The problem. A neuron here is one unit in a network layer, a number that rises or falls with the input. The naive hope was one neuron, one idea. In practice single neurons respond to mixtures of unrelated things, called polysemantic, meaning many meanings.

The leading explanation: superposition. A model must track far more features (ideas like 'French' or 'legal text') than it has neurons. Most features are sparse, rarely active together. So the network gives each feature a direction in the layer's space rather than its own neuron, with directions nearly but not perfectly perpendicular. Think of an overbooked hotel: it works because guests rarely all arrive at once. The cost is small interference when two features do appear together.

Evidence. Anthropic researchers showed this in small toy networks in an earlier paper, where several features were packed into fewer dimensions in a regular geometric pattern. Toy models show a mechanism, not proof that large models behave identically.

Looking inside. The main tool is the sparse autoencoder, a second small network trained to rebuild a layer's activity from a large dictionary where only a few entries fire at once. Entries often line up with readable concepts. Google DeepMind released Gemma Scope, a set of these for its open Gemma 2 models. Terms and access: check Google's page; today's stories state none.

Limits. It is an approximation of the layer, not a perfect readout, and the features you get can depend on training choices. Novel: the framing of ideas as directions. Still maturing: the tooling.

Today's essay. The Chosun Ilbo on MSN reports that researcher Nick Kamarata's essay challenges the superposition picture with a '90% neuron insight'. Some compare it to the 'Attention' paper; others say it won't change the big picture. What the 90% measures is not in what we have. STILL OPEN.

A lesson, not news. An illustrative workflow we have not tested.

Safe inbox triage in about 45 minutes (setup time is our estimate). For: anyone with a crowded Gmail inbox. Solves: the daily sort. Needs: Gmail, plus Gemini in the app or AI Studio (check current pricing pages for plans). Filters are built into Gmail.

1. Create four labels: Action, Waiting, Reference, Auto-sorted. Rule zero: automation never deletes anything.
2. Copy about thirty recent senders and subject lines. Senders and subjects only, no message bodies, and remove anything private.
3. Paste them into Gemini with the prompt in the next section and ask for at most six filter rules.
4. In Gmail settings, open Filters and Blocked Addresses. Create each filter from Gemini's search query, choose Apply the label, and for newsletters and receipts also choose Skip the Inbox.
5. Pick a fixed time, like the 7:30 in the account we covered, and spend five minutes in Auto-sorted for the first week.
6. Measure: inbox count at day's end and minutes spent, before and after.

Limits: filters are static, Gemini's guesses can be wrong, and a rule that hides mail from a real person is the main risk. Verify: after a week, open each label and scan for misfiled mail, and ask whether anything important skipped the inbox. If so, tighten that rule. If you later hand this to an agent tool such as Antigravity, keep three rules: label, don't delete, keep a readable log.

The number (an estimate, with illustrative inputs): if you spend about 20 minutes a day sorting email and a one-time routine cuts that to 5, you reclaim 15 minutes a day, about 75 minutes a work week, roughly an afternoon block of focus time. Swap in your own minutes.

Before you pay for anything: build the filters first. Gmail has filters and Outlook has its own rules feature. Google sells Gemini in Gmail and Microsoft sells Copilot in Outlook. Today's stories state no prices for either, so check each current pricing page and compare the plan with the minutes you would actually get back.

Rule of thumb: if you cannot name the minutes a tool would return, do not buy it yet. Run the routine for two weeks, count minutes, then decide.

What you give up: some control. You accept the occasional misfiled message and spend about five review minutes a day at first. Keep manual review for legal, medical or client-critical mail, and check your workplace policy before sending sender and subject lines to any assistant.

Objective: turn thirty email headers into safe, pasteable Gmail filter rules. Illustrative; not tested.

Weak version: 'Help me organize my email.' It names no labels, no sample, no limits, so you get generic advice like 'make folders for projects'.

Better version:

You are helping me design Gmail filters. Below are 30 recent emails as sender and subject only. Propose at most 6 filter rules. For each rule give: (1) the exact Gmail search query I would paste into the filter criteria, (2) the label to apply, chosen only from: Action, Waiting, Reference, Auto-sorted, (3) whether to skip the inbox: yes or no, (4) one email from my list it would catch and one it might wrongly catch. Never recommend deleting anything and never mark anything as read. Mark any rule that could hide a message from a real person as RISKY. If the sample is too small to justify a rule, say so instead of guessing.

Emails:
[paste here]

Why it works: a fixed label list makes output match your setup; an exact search query can be pasted straight in; deletion is banned up front; the right-and-wrong sample forces the model to show its work; and it may say 'too small' instead of guessing.

Customize: edit the labels, change the six, or add 'never touch mail from these domains: [list]'. Quality check: ask Gemini which of your thirty emails each rule would catch, then verify two by hand. Include one decoy from a real person and see whether it is flagged RISKY.

The finding, still open: the essay behind the 90% neuron claim has split researchers, with some calling it as important as the 'Attention' paper and others calling it interesting but not big-picture changing. What the 90% measures is not in the coverage we have.

The decision this week: do not change any tool or model choice because of a viral essay. Ask three questions first. What exactly was measured: 90% of what? On which models? Has anyone replicated it, meaning repeated the work independently and got the same result? Until then, treat it as unproven however loud the debate. If you pick models for work, spend an hour running ten of your own real tasks through two options and compare. That evidence beats any argument.

The thread (our analysis): voice and agents are becoming how Google's tools get used: you talking to Gmail, an agent sorting an inbox at 7:30, possibly avatars speaking for you. That is not a stated Google plan.

• Google's own word on Gmail Live: availability and price. Confirmed when Google publishes it.
• A Google statement on avatars and voice imitation, especially disclosure and consent. Until then it is a video summary.
• Geely's 2027 rollout details: an announced plan without a firm date in what we have seen.

I asked Gemini to clean my inbox; now I have zero emails and one very confident apology.

That's it for today. Tomorrow we're watching whether Google publishes its own details on Gmail Live, including where it's available and what it costs, and any word on how Gemini avatars would be disclosed.

If this made you smarter, share it with one person who needs it. SI Agent News is independent and self-funded, with no investors and no ads., and every share and every bit of support keeps it going. Thank you.

Made by AI.