SI Agent News

• Brand tools in Gemini. Yahoo Tech reports Google is adding brand management tools to Gemini. Details are thin and the price is not stated. Marketers should watch for it. Status: REPORTED.
• Gemini for admin work. CoinGeek profiles the founder of StonedMNL, a small business that uses Gemini to cut admin. It is a use case, not a launch. Status: REPORTED.

Google cleared 300 hectares of Finnish forest for a data center. TechSpot reports it. That is about three square kilometers, roughly 400 football pitches (our estimate, not TechSpot's). Every Gemini answer and Workspace summary runs in buildings like this, so AI's cost is not only dollars on an invoice. It is also land, and that part rarely appears on a procurement sheet. We have the size and the company. We do not have the local approvals, any replanting plans or why this site was chosen, and we will not guess at them. Status: REPORTED for the clearing; the reasons are STILL OPEN. Price: not stated, it is not a product. Why it matters for you: if you buy cloud AI for a company that publishes sustainability numbers, questions about where your workloads run are likely to reach you sooner or later. Worth following for that reason.

What this means for you:

• Small business owner: expect customers to ask where your AI runs, so keep your vendor's sustainability page handy.
• Job seeker or student: energy-aware infrastructure work is worth reading up on. That is our inference, not a reported hiring trend.
• Creator: audiences are noticing AI's footprint, so know how to answer questions about your tools.

Anthropic marks Claude's text invisibly. BornCity, a German tech site, reports a watermark in Claude's output. Claude is a rival chatbot, but it matters here because mixed drafts from different AI tools may be identifiable later. If you paste Claude-written passages into a document that also contains Gemini text, the origin of each part could become traceable. Background from us, not today's reporting: Google DeepMind has its own text watermark too. How the Claude watermark works, whether it can be switched off, and how reliable detection is are all STILL OPEN. So is whether ordinary editing, such as paraphrasing or reformatting, removes it. Status: REPORTED. Price: not applicable. Action: set your disclosure policy for AI-assisted writing now, rather than waiting for a detector to force the question. That matters most for anyone who submits work to an employer, client or school.

Google and Amazon go after Nvidia's CUDA from the software side. Taiwan's Anue reports that the two companies are challenging Nvidia's CUDA ecosystem through software, and its headline says 98% of models need no major rewrite. Background from us: CUDA is Nvidia's software layer, and much AI code is written against it, which is why moving to other chips has usually meant costly rework. Google and Amazon both build their own AI chips, so an easier path off CUDA would help them. What we do not have: what the 98% measures, who tested it, which models were included, or how performance compares after the move. Status: REPORTED; the headline figure is UNVERIFIED. Price: not stated. Why it matters: if you rent GPU capacity or run your own models, portability could eventually give you more vendor choice and better pricing. Action: ask your provider what happens to your workload on non-Nvidia hardware, and test your own model before trusting any average.

Fireworks claims one-fifteenth the cost. Fireworks, a model-hosting platform, says DeepSeek-V4.1-Flash, a model it pitches for coding agents, reaches Astra-level results on a test called DeepSWE at 1/15th the cost. The post is dated 14 September 2026 and the numbers are the vendor's own. We have no detail on what Astra is beyond its role as the comparison point, and the post's title is all we have of the method. We also do not know whether the 1/15th ratio is per token or per solved task, and that difference changes what you would actually pay. A hosting company selling access to the model has an obvious interest in the comparison looking good. Status: REPORTED, vendor-sourced. Price: only the ratio is stated. Worth it if you run coding agents at volume. Action: run a sample of your own tickets or tasks through it and compare cost per accepted result, not the headline.

Five agent harnesses took the same memory test, and the results were not close. The New Stack ran the comparison. A harness is the scaffolding around a model that manages tools, context and what the agent remembers between steps. The headline says the scores varied widely, which suggests the harness you pick may matter as much as the model underneath. We do not have the names of the five harnesses, the test design, the ranking or the size of the gaps. Status: REPORTED; the details are STILL OPEN. Price: not stated. Why it matters: memory is where agents fail in daily use, forgetting a preference or repeating work you already approved. Worth reading in full if you are choosing an agent framework for a team. Action: before committing, give your shortlist a memory task drawn from your real work, such as recalling a decision from earlier in a long session.

Google adds brand management tools to Gemini. Yahoo Tech reports the addition. The headline suggests help with keeping a business's voice, look and messaging consistent across what Gemini produces, which is a common complaint about AI-generated marketing copy. We only have the headline, so we do not know which tools are included, which Gemini plans get them, whether they cost extra or when they roll out. Status: REPORTED; features, availability and price are STILL OPEN. Why it matters: for a small business or creator without a designer, a way to make drafts sound like you could save real editing time. Action: check your Gemini settings and any admin console for new brand options, and gather your logo, colors and a few sample posts so you can test quickly once the feature appears for you.

Google Cloud's Knowledge Catalog. Google Cloud published a post on Knowledge Catalog, a Google Cloud service for self-learning context enrichment for AI agents. In plain words, it aims to build up what an agent knows about your data automatically, so it stops misreading a column name or a business term. This is Google describing its own product and we have only the title. We do not know how it learns, how a person corrects it when it learns wrongly, or what it costs. CONFIRMED that it was published; how well it works is STILL OPEN. Price: not stated. Worth a look if your data is already in Google Cloud and your agents keep misreading it. Action: pick one recurring wrong answer from your current agent, and treat whether this fixes it as your pilot test.

StonedMNL's founder uses Gemini to cut admin. CoinGeek carries a profile headlined "Less admin, more heart," about how the founder of StonedMNL balances the hustle of running the business with Gemini. We have only the headline, so we cannot say what the business sells, which Gemini features are used, how much time is saved or whether Google supplied the story. Treat it as one owner's anecdote, not a measured result. Status: REPORTED; specifics are STILL OPEN. Price: not stated. Why it matters: small-business stories like this are the clearest sign of where Gemini is actually being used day to day, which is paperwork, replies and scheduling rather than anything glamorous. Action, our suggestion rather than the article's: list the three admin tasks that eat your week, hand one to Gemini for a fortnight, and keep it only if you can measure the saved time.

For strategists: Anue, a Chinese-language financial outlet, reports that Google and Amazon are attacking Nvidia's CUDA ecosystem from the software side, with 98% of models needing no major rewrite. CUDA is Nvidia's programming platform for its GPUs (graphics chips). Most AI code has been written against it for years, and that is the moat.

The claim in plain words: models built for Nvidia move to other chips with few code changes. Background from us, not the headline: Google makes TPUs and Amazon makes Trainium chips.

Three questions before you believe it:

• 98% of which models, and who counted? A catalog average is not your workload.
• 'No major rewrite' is not 'same speed.' Code can run and still run slowly.
• The last 2% is often the custom code that matters most.

Status: the report is REPORTED; the 98% figure is STILL OPEN until a primary source or independent test backs it.

Who benefits: anyone who buys compute, because competition lowers prices even if you never switch. Who might lose: Nvidia's pricing power. The rivals are the story's own cast: Google and Amazon versus Nvidia. For Gemini users nothing changes today; this is the supply side. Price: not stated.

Do this: when you negotiate cloud AI, ask for a porting test of one real workload of yours. Watch: a technical write-up from Google or Amazon.

For builders: The New Stack reports that five AI agent harnesses took the same memory test and the results were not even close. Its summary adds that remembering something does not guarantee an agent can use it. We have the headline and summary, not the scores, so this is how agent memory generally works, not a recap of the results.

A harness is the software around a model: it builds the prompt, calls tools and loops until the task is done. Memory lives in the harness, not the model. It has four steps:

• Write: decide what is worth saving.
• Store: files, a database or a search index.
• Retrieve: choose which memories go back into the prompt for a new task.
• Apply: the model actually uses them.

The 'remembering is not using' point sits between retrieve and apply. The same model in five harnesses can score five ways because saving everything buries the signal, saving too little forgets, retrieval by similar wording can miss the memory that matters, and the context window (how much text a model can read at once) is finite, so each memory loaded pushes something else out.

Novel vs incremental: testing whether memory changes behavior is newer than testing whether a fact can be recalled. Caveat: the web address names Letta, which as far as we know builds agent tooling, so a participant may have run the test. Read the method. Status: REPORTED; results not verified by us. Price: not stated. Worth it for anyone choosing a harness for long-running agents.

What to do: write ten tasks that depend on something said earlier and run them through your own harness. One memory test is one scenario; a winner on recall may lose on speed or cost.

Build a hand-made memory file for any AI project. Illustrative workflow, not tested by us. Who it's for: anyone who keeps re-explaining a project to an AI. Works in Google AI Studio, the Gemini app or any chat model.

• 1. Make a plain text file called memory with four headings: Facts, Decisions, Open Items, Don't Repeat.
• 2. At the end of each session, paste the closing prompt below.
• 3. You approve every line. Nothing goes in unapproved.
• 4. Next session, paste the file at the top. In AI Studio, the system instructions box is the natural place. Keep it under about 500 words (our estimate).
• 5. Verify: ask three questions whose answers are in the file and one trap whose answer is not. If the model invents an answer to the trap, add a top line: 'If it is not in this file, say you do not know.'

List up to five items worth remembering for next session. Put each under one of these headings: Facts, Decisions, Open Items, Don't Repeat. Keep each to one line, and quote the sentence from this chat that supports it.

Expected output: a session that opens by answering 'where were we' in one reply. Limits: it is manual, errors in the file persist, and you should never store passwords or client secrets. Time cost: about five minutes per session (our estimate).

Compare cost per solved task, not price per token. Fireworks says DeepSeek-V4.1-Flash costs 1/15th as much. Cost per solved task equals cost per attempt divided by how often the attempt succeeds. Break-even math: at 1/15th the cost per attempt, the cheaper model still wins as long as it solves at least about 7% as many tasks as the pricey one (1 divided by 15). Example only, not a quoted price: a $150 monthly coding-agent bill at 1/15th would be $10.

Do it: pick 20 real tasks and run them on three options from different companies: Fireworks-hosted DeepSeek, a Gemini Flash-class model through Google AI Studio or Vertex AI, and a model from another vendor such as OpenAI or Anthropic. Check the current price on each pricing page. Estimated time: one to two hours.

What you give up: the 1/15th figure is the vendor's own claim, the post is dated 14 September so prices may have moved, human review time often outweighs token cost, and your code goes to another provider.

Prompt clinic: audit a vendor claim.

Weak prompt: 'Is this true: 98% of models need no rewrite?' A model cannot know that from a bare sentence, so it guesses fluently.

Better prompt:

You are a skeptical analyst. Below is a vendor claim. Do five things.
1) Restate the claim in one sentence, with its number and what the number measures.
2) List what is missing: sample size, who measured, comparison baseline, date.
3) Give three questions I should ask the vendor.
4) Say what independent evidence would confirm it.
5) Label the claim CONFIRMED, REPORTED or STILL OPEN, using only the text I pasted. If you use outside knowledge, mark it as such.
Claim: [paste text]

Why it works: it forces a structure, ties the model to your text and stops it guessing. Expected result: a short audit with questions you can send. Customize: add your industry, or your budget. Quality check: open the original source and match each 'missing' item against it yourself.

Lesson from earlier research, not today's news. Research has found that language models use information best when it sits at the start or end of a long input, and worse in the middle. A context window is how much text a model reads at once.

Decision this week: in any long prompt, put instructions and key facts at the top and restate your key question at the bottom. Then test it by moving one key fact to the middle and watching whether the answer changes. Limit: tested on models of that time; newer ones may do better, so check yours.

Agentic AI is one of today's busiest news lanes., and today's stories touch all three: agents and memory, a watermark and a land-use question, and an open model on cheap hosting.

Our analysis: the contest is shifting from who has the smartest model to who is cheapest, most trustworthy and best at remembering. Watch (none scheduled): an independent test of the 98% claim, Google's response on the Finland site, and documentation for Knowledge Catalog.

Listening daily is learning with a reason. Today that means a cost-per-solved-task check and a memory file you can start tonight.

My agent remembers everything I told it last week. It just doesn't know which part mattered.

That's it for today. Tomorrow we're watching for independent numbers behind the 98% CUDA claim, and for more on Gemini's brand tools. If this made you smarter, share it with one person who needs it. Thank you for spending your time with us. SI Agent News — a podcast newsletter. Made by AI.