SI Agent News

• MCP for lawyers: Legal Desire published an explainer of the Model Context Protocol (MCP, an open standard for connecting AI assistants to your documents and tools) aimed at legal teams. It is an explainer, not breaking news. Status: CONFIRMED that it exists, Legal Desire.
• Marketing data meets enterprise AI: Macau Business reports that Master Concept is partnering with Supermetrics (a marketing-data company) to bring marketing data into enterprise AI. It reads like a press release with little detail; price not stated. Status: REPORTED, Macau Business.

Today's Signal runs from a model nobody can call yet to a cyclone, a stock-market wobble and three practical pieces on running AI in real work. Where we have only a headline or summary, we say so and keep claims to what is reported.

Gemini 4 Carbon: first on the index, closed to callers (OrcaRouter, a service that tracks models available through one router). Gemini 4 Carbon has appeared on a model index, but nobody can call it yet. From the headline and summary we have, the model is listed but not callable. We have seen no Google announcement, so the name, tier, capabilities and price are unconfirmed. Why it matters: a listing is not availability. If you call Gemini through an API, the useful habit is to keep the model name in one setting rather than scattered through your tools. Then a switch, once a model is officially callable and priced, is a one-line change you can test first. Status: REPORTED by a single source; most of it is STILL OPEN. Price not stated. Worth watching, not worth acting on yet. OrcaRouter

Strong earnings can't save tech stocks (technews.tw finance site). The report says investors are re-examining the return on AI capital spending and cash flow, even after strong earnings, with Alphabet, Google's parent company, named as an influence on US stocks. Our analysis, not the report: investor pressure on returns could show up in plans, usage limits or pricing over time, or push the other way toward aggressive pricing to win usage. We can't tell which. For you, that uncertainty is the point. Prices and terms for AI tools can change, so avoid building everything around one vendor's quirks. Keep prompts and test cases portable, and keep a short list of alternatives you have already tried. Status: REPORTED; consequences for Gemini pricing are STILL OPEN. technews.tw

Cyclone Anab hits India's "data center corridor" hard (technews.tw). The headline says the cyclone struck India's data center corridor hard, and that a $15 billion investment by Google, Adani and others now faces a climate test. We have only that headline and summary, so the scale of damage, any outages and the effect on build schedules are not confirmed here. Why it matters: AI tools run in physical buildings, and those buildings sit in places with weather. Most users will never notice, but anyone running important work in the cloud should know which region their provider uses and whether a backup exists. Our analysis: the story is also a reminder that large infrastructure bets carry climate risk, which investors are already weighing. Status: REPORTED; service impact and cost STILL OPEN.

Engineering Realities of Post Launch AI: Architecture, Drift and Operational Control (dev.to). The title frames the work that begins after an AI feature ships: how it is built, how its behavior changes over time (drift), and how you keep control of it. We have the title and summary, not the full argument, so we won't attribute specific methods to the author. Our analysis: this is the part most small teams skip. An AI tool that worked in March can behave differently by October because the model, the data or the way people use it has changed. Practical steps that follow from the theme: keep a small set of example questions with answers you accept, rerun them on a schedule, and note which model version you use. If results slip, you will see it before a customer does. Status: REPORTED as a practitioner article; specifics STILL OPEN. dev.to

Model Context Protocol (MCP) Explained for Lawyers and Legal Teams (Legal Desire). MCP is an open standard for connecting AI applications to outside tools and data. This piece explains it for legal readers. We have only the headline, so we can't summarize its advice. Why it matters beyond law: any profession with confidential files will face the same question once an AI assistant can reach documents, email or databases through a connector. Our analysis: before you connect anything, ask what the AI can read, whether it can change or send anything, where the data goes, and who can switch the connection off. Start with read-only access to a single folder. Status: REPORTED explainer; legal-specific guidance STILL OPEN.

Master Concept Partners with Supermetrics to Bring Marketing Intelligence into Enterprise AI (Macau Business). The headline says Master Concept has partnered with Supermetrics to bring marketing data into enterprise AI. Supermetrics works in marketing and advertising data. We have no terms, pricing, launch date or customer details, so treat this as an announcement, not a proven result. Why it matters: marketing teams often hold their best data in separate ad and analytics tools, and AI answers are only as good as the data behind them. If you try something similar, check which accounts are connected, who can see the results, and whether the numbers match your own reports. Status: REPORTED partnership; availability and price STILL OPEN.

What this means for you:

• Small business owner: do not switch tools on an unconfirmed listing; keep what works until a model is officially callable. Check which cloud region your key tools use, and review what any connected AI can access.
• Job seeker or student: practice judging AI output against clear checks; it is a skill employers can see. Knowing what MCP-style connections and drift mean is a plain-language edge.
• Creator: save your best prompts in one place so you can re-run them on any new model, and keep a few example outputs to spot when results change.

A cyclone meets India's data center corridor

technews.tw reports that a cyclone struck India's eastern coast, in a stretch it calls the data center corridor, and that about $15 billion of investment from Google, Adani and others now faces a climate test. A note on names: the headline says Cyclone Anab, the summary says Arnab. We use the headline spelling. We have seen no confirmed facility outages or damage figures. Status: REPORTED, details unverified; impact on service is STILL OPEN. technews.tw

Why it matters (analysis, not reporting): every AI request runs somewhere. Big providers spread work across many regions, but customers with data residency rules (laws that keep data inside a country) or tight speed needs can end up pinned to one. Coasts are attractive for things like power and cables, yet exposed to storms, so climate risk becomes a siting question.

Who benefits and who is exposed: providers who can show real resilience gain trust; regional customers and local builders carry the risk if a single site is the weak point.

What to do: if you run on Vertex AI (Google's cloud platform for building with Gemini) or any cloud, ask three questions. Which region holds my data? Is there a second region that is legal for me? Has anyone ever tested the failover? If you are not technical, write down what happens if your AI tool is down for a day. My estimate: about fifteen minutes.

Outlook (analysis): expect customers to ask more about where capacity sits and how it fails over. Confirmation would be statements from the companies involved. Nothing we have points to a Google-specific problem.

For builders: what actually breaks after an AI launch

A dev.to post by an engineering lead who runs cloud infrastructure argues that building an AI proof of concept takes days, while keeping a generative AI pipeline reliable in production is a different discipline. We have only a short summary, so the mechanism below is general engineering and our own explanation, not a quote. Status: post REPORTED via summary; mechanism is analysis. dev.to

The pipeline is more than a model. Input, prompt template, often retrieval (fetching documents to give the model context), the model call, cleanup, output. Drift, meaning behavior that changes without anyone deciding it should, can enter at every layer.

• Input drift: users start asking different things.
• Prompt drift: someone edits a template and nobody records it.
• Retrieval drift: documents change, so the same question gets different context.
• Model drift: the provider updates the model behind a name. Many vendors offer both fixed version names and moving aliases such as latest. An alias can change under you; a pinned version holds until retired. Check your console's model list to see which you call.

Control is a loop: log every call with prompt version, model version and inputs; run a fixed test set on every change; set thresholds such as a pass rate or refusal rate; roll changes to a small slice of traffic first (a canary); keep a fallback path and an off switch.

Small team version, about an afternoon: pin the model, keep prompts in a dated document, keep twenty test cases.

Novel or incremental? Incremental. It is classic monitoring applied to outputs that are never identical twice. The real twist is that you score outputs rather than check equality, often with another model as grader.

A no-code bake-off for your own work (illustrative workflow, not tested by us)

Who it is for: anyone using Gemini, ChatGPT, Claude or similar for repeated tasks who wants to know whether a change helps. Problem solved: switching models on hype or noticing quality slips too late. Needs: a spreadsheet (Google Sheets or Excel) and about an hour.

1. Gather 15 to 20 real tasks from the last month, with sensitive details removed.
2. Columns: A task, B exact prompt, C what a good answer must contain (three short checks, such as 'mentions the deadline' or 'under 150 words').
3. Freeze: write down the exact model name and the date.
4. Run every prompt on your current model in a fresh chat; paste answers in column D.
5. When something changes (new model, new tier, an update), run the same prompts with the same settings; paste answers in column E. If Gemini 4 Carbon ever becomes callable, that is where it goes.
6. Score: pass or fail per check, plus usefulness 1 to 5.
7. Compare pass rate, which cases got worse, and how long you spent editing. Switch only where the candidate wins and no critical case gets worse.

Expected output: a one-page table you can show a boss. Limits: twenty cases is a small sample and scoring involves judgment. How to verify it worked: run your current model twice; if it varies as much as the gap you are measuring, you cannot conclude anything.

Estimate: about an hour and a half saved per model change. Checking twenty outputs by eye after a change takes roughly two hours; scanning a sheet with written checks is closer to fifteen to twenty minutes. That is our estimate, and yours will vary.

Log it in Google Sheets or Excel, whichever your team already uses, and run Gemini, ChatGPT or Claude through the same test, so you are not locked in to any one of them.

If you call models through an API with the same long instructions every time, ask each provider about caching (reusing repeated input instead of processing it fresh). Several major AI providers document some form of it. No prices are stated in today's stories, so check the current price on each pricing page.

What you give up: about an hour up front to build the sheet, and it goes stale as your work changes, so refresh it quarterly.

Prompt clinic: the grader for your bake-off.

Weak: Which answer is better, A or B? It has no criteria, can be swayed by whichever answer comes first, gives no evidence, and cannot be pasted into a sheet.

Better (copy-ready, illustrative, works in any chat assistant):

You are grading two answers to the same task.
Task: [paste the task]
Must-have checks: [paste your three checks]
Answer A: [paste]
Answer B: [paste]
For each answer, mark every check PASS or FAIL, with a one-sentence quote from the answer as evidence. Then rate usefulness from 1 to 5. Do not reward length. If the answers are equal, say TIE. End with one line: WINNER: A, B, or TIE.

Why it works: the criteria come from you, evidence is required, the length instruction pushes against a known bias, and the last line is easy to paste into a sheet. Customize: swap in your own checks or add one for tone. Quality check: run it again with A and B swapped; if the winner flips, trust neither and read both yourself.

Background, not today's news. Published research on using AI models as graders has reported that AI judges can show position bias (favoring whichever answer appears first), verbosity bias (favoring longer answers) and self-preference (favoring their own style). It does not show that any grader is right on your tasks.

Decision this week: swap the answer order and run twice, use a grader from a different company than the model you are testing when you can, and send any disagreement to a human.

Agentic AI is a particularly busy area today. Agents make many calls in a row, so reliability and cost scrutiny matter more, which ties the investor worries to the drift problem.

• Gemini 4 Carbon confirmed by Google: STILL OPEN; confirmation would be an official Google post.
• Cyclone impact on India's data center corridor: REPORTED, details unverified; confirmation would be statements from the companies involved.
• Spending-return debate reaching plans or pricing: expectation only; evidence would be a published pricing change.

Models drift in production. So does my weekend plan, but only one of them has monitoring.

That's it for today. Tomorrow we're watching whether Google says anything official about Gemini 4 Carbon, and whether the cyclone's impact on India's data center corridor comes into focus. If this made you smarter, pass it to one person who would find it useful. Thank you for spending part of your day with us.

SI Agent News — a podcast newsletter.

Made by AI.