RightTech

10/12/2026

TabsMaster Chrome extension is out on the Chrome Web Store.
- An AI summary on every tab, so you still know what it was months later
- AI groups you describe in plain English, sorted for you automatically
- One search box across your open tabs, your bookmarks and your entire history
- Parks idle tabs to win the memory back, and refuses any tab you've typed in
- Finds duplicates, dead links, and pages nobody has touched in years
Built and tested on a browser holding 1,131 tabs. More on each of those below.
SEARCH — one box, three sources
Type two characters and it searches the title, URL and AI summary of every open tab, plus your bookmarks and your browsing history, each in its own section.
SEARCH — one box, three sources
Type two characters and it searches the title, URL and AI summary of every open tab, plus your bookmarks and your browsing history, each in its own section.
History is searched over all of it, not Chrome's default 24-hour window. That's the difference between finding the article you half-remember from last year and not finding it.
Results are de-duplicated across all three. If a page is open, bookmarked and in your history, you see it once, as the open tab, instead of three times in three places.
AI SUMMARIES — what was this tab, again?
Cloudflare renders the page, Gemini condenses it to 25 words or fewer, and the result is cached permanently. It survives the tab being closed, so something you shut in March is still described in October, and still searchable.
AI GROUPS — describe it, don't configure it
Type "AI tools for legal work" and Gemini reads every tab and decides which ones belong. The group becomes a filter chip across the top, and each matching tab gets a badge. The chip tells you the group has 30 tabs; the badge on a row tells you why that tab is one of them.
New tabs sort themselves about eight seconds after they load, so groups stay current without you touching them. One call judges 25 tabs against every group at once, so a few hundred tabs cost a handful of requests rather than hundreds.
Change a rule later and the old verdicts are cleared rather than kept. They were judged against wording that no longer exists, and a stale verdict quietly filtering your inventory is worse than a tab showing up unsorted.
DUPLICATES — exact, and same-domain
Both group by the duplicate rather than by the window, because copies of a page almost always live in different windows, and grouping by window scatters the exact thing you're trying to compare. Each header tells you how many copies there are and how many you could close.
Same-domain groups deliberately don't suggest a number to close. Those are different pages, and telling you to close them would be advice to throw away real tabs.
SUSPENDING — without losing anything
The tab keeps its real title and favicon and comes back when you click it. The placeholder shows a screenshot of the page, its summary, and how long it's been sitting there.
It watches for real keystrokes rather than guessing, so a half-written comment never gets thrown away, and it tells you which tabs it refused and why. Pinned tabs and anything playing audio are left alone. Alt+Shift+S parks the one you're on; or select a batch, or let the idle timer handle it.
THE SMALLER THINGS THAT ADD UP
- Already using another suspender? It decodes their parked tabs back to the real URL and title, so your list shows actual pages instead of a wall of placeholder links.
- The dead-link scan separates genuinely dead from merely blocked. Plenty of live sites reject an automated request, and calling those dead would be wrong.
- Every bulk job says what it skipped: "Summarising 258 of 528 selected, skipped 190 already summarised, 74 not web pages, 6 repeats of the same URL."
- Real last-updated dates for GitHub repos and habr.com articles, read from the public feed with no token and no rate limit. Amber past a year, red past two.
- Sort by when you last touched a tab. No scan needed, covers everything.
- Hover any row for a screenshot of the page.
- Full session save and JSON export, so a crash doesn't cost you everything.
YOUR DATA STAYS YOURS
The AI features run on your own Gemini and Cloudflare keys. No account, no analytics, no server belonging to anyone else. Without keys, everything except the summaries still works.

https://chromewebstore.google.com/detail/tabsmaster/gmcdkdgcepbkobcciimijknaiaomhlmh

09/07/2026

Right Tech Monthly Digest -

Issue two of the Right Tech Monthly Digest. What's new at Right Tech, plus the AI and IT news worth your attention this month.

09/05/2026

Seriously, it's only $1. If you still haven't learned how to vibe code and want to know, this is for you.

Comment "vibe" below and I'll send you the details.

08/29/2026

Guilty as charged

AI hasn't replaced us, it's just given us different job titles.

08/28/2026

Quick breakdown of the journey so far to being able to generate hour-long AI videos.

08/16/2026

Your SaaS subscription has another cost: your data

There is another reason businesses are going to start cutting SaaS products from their stack.

It is not the subscription price.

It is the data.

For years, the trade was pretty straightforward. You gave a software company your customer records, documents, conversations, analytics, sales pipeline, support tickets, internal processes—whatever the product needed—on the assumption that it would be stored safely, locked down, and accessible only to you and the people you explicitly authorised. In return, you got software that would have been far too expensive to build yourself, and you trusted that the provider would keep it secure, private, and properly isolated from everyone else.

That was a reasonable deal.

It isn’t anymore.

A good example happened with HubSpot just a few weeks ago.

HubSpot announced changes to its terms around a new Contact Discovery product and its enrichment dataset. The idea was that customers participating in enrichment could contribute certain professional contact information and email engagement signals to help maintain a shared commercial dataset.

Customers were not impressed.

Four days later, HubSpot reversed the changes.

Their response was unusually direct:

"We made a mistake."

They acknowledged that the rollout made customers feel like the relationship with their CRM, and therefore their data, was changing underneath them. HubSpot cancelled the terms changes and said future enrichment capabilities using customer data would be clearly opt-in.

Credit where it is due. They listened and fixed it.

But the interesting part is that they could propose it in the first place.

Think about what sits inside a CRM.

Your customers.

Your prospects.

Who buys from you.

Who nearly bought from you.

Your notes.

Your sales process.

Years of accumulated commercial intelligence that your team has paid real money to build.

And suddenly you are reading a terms update trying to work out whether some part of that information can contribute to somebody else's commercial dataset.

That would make me uncomfortable too.

One HubSpot customer commenting on the reversal said the software had now been put on their internal "at risk" list. Another said they were initiating an RFP because they no longer felt safe adding commercially valuable data to the platform.

That reaction makes sense to me.

Because once your business runs on somebody else's platform, you do not just depend on their software.

You depend on their future business model.

And AI has made data far more valuable to those business models.

Look at Slack.

Slack says it does **not** use customer data to train generative AI models unless a customer explicitly opts in.

Good.

But its predictive machine-learning systems can analyse customer data such as messages, content and files to improve global models. If an organisation does not want its data contributing to those models, it has to opt out.

Again, I am not saying Slack is secretly stealing your messages.

That is not the point.

The point is that the data you put into a piece of software can have value to the company running that software beyond simply providing the service you paid for.

And the rules around that can be more complicated than most customers realise.

Google makes the problem even more interesting.

There was a strange Reddit post recently from an indie game developer.

A player was asking Google's Search AI questions about his game. At one point it returned the exact name of an unreleased character: "Vantage Tripod."

The developer said he had never published the name and, as far as he knew, the only digital copy existed inside one of his private Google Docs.

Now, before everybody reaches for the pitchforks, there is no proof that Google leaked his private document.

The name could have come from somewhere else. Metadata. An old build. A forgotten file. Some obscure public source. Nobody has demonstrated what actually happened.

But the story is interesting because Google's own documentation shows how blurry these boundaries are becoming.

For eligible personal Gemini accounts using Connected Apps, Google says data from connected services can be used to improve Google services, including training generative AI models when the relevant activity setting is enabled.

Google says it does not simply dump your entire Gmail inbox or Drive into model training. But summaries, excerpts and inferences from relevant emails and files can be used, and Google explicitly notes that if a file is short or particularly relevant, the "summary" may effectively be the file itself.

Their own documentation tells users not to connect apps containing personal or confidential information they would not want used for generative AI training.

Read that sentence again.

We spent twenty years teaching businesses:

"Put everything in the cloud."

Now we are adding:

"...but make sure you understand which AI settings might cause parts of it to be used to improve somebody else's model."

That is quite a change.

And this is before we get to actual security incidents.

In 2024, attackers compromised Snowflake customer environments using stolen credentials. Mandiant and Snowflake notified around 165 potentially exposed organisations.

Importantly, Mandiant found no evidence that Snowflake's own enterprise environment had been breached. Attackers were getting into customer accounts using credentials stolen elsewhere, helped by things like missing MFA and weak network restrictions.

Technically, that distinction matters.

From the point of view of the company whose data has just been stolen, it probably feels slightly less important.

AT&T disclosed that attackers accessed one of its workspaces on a third-party cloud platform and copied records covering calls and texts for nearly all of its wireless customers during certain periods.

The contents of the calls and texts were not included, nor were things like Social Security numbers or dates of birth.

Still, "nearly all of our wireless customers" is not a sentence anybody wants to put into an SEC filing.

None of this means self-hosting everything is automatically safer.

It is not.

I have seen enough badly configured servers, forgotten admin accounts and applications held together with hope to know that "we built it ourselves" is not a security strategy.

A good SaaS provider can absolutely have better security than a small business.

Often they do.

But there is another side to that equation.

If you use twelve SaaS products, your data now exists across twelve vendors.

Twelve authentication systems.

Twelve sets of employees and contractors.

Twelve lists of subprocessors.

Twelve terms of service.

Twelve companies making their own decisions about AI over the next five years.

And probably a collection of integrations copying data between all of them.

Every new system becomes another place something can go wrong.

That used to be a risk businesses mostly had to accept, because the alternative was spending $100,000 building internal software.

That is the bit AI is changing.

If you are paying $800 a month for a system because you need a database, four screens, a few automations and an email notification, building your own version of those four screens is no longer necessarily insane.

You do not need to recreate Salesforce.

You need to recreate the tiny bit of Salesforce your company actually uses.

You do not need to build Google Workspace.

You might need a small internal application that stores one category of commercially sensitive information somewhere you control.

You do not even necessarily have to "leave the cloud."

There is a huge difference between renting infrastructure from a cloud provider and handing your entire business process to an off-the-shelf SaaS product.

With a purpose-built application, you decide what gets stored.

You decide which external services see it.

You decide whether it is sent to an AI model.

You decide how long it is retained.

And if you want to change those rules, you do not have to wait for the vendor to update its privacy policy.

This is why I think the SaaS argument is becoming much bigger than price.

The monthly saving is nice.

Getting rid of ten features nobody uses is nice.

Having software that actually matches the business is nice.

But owning the rules around your own data might end up being the bigger advantage.

The old calculation was:

**Why would we build this when we can rent it for $500 a month?**

The new calculation is starting to look more like:

**Why are we paying $500 a month to give another company our data, accept their roadmap, accept their security model, accept their future AI policy and use 15% of their product... when building the 15% we actually need has become dramatically easier?**

That does not mean every SaaS product should disappear.

I am still not building my own Stripe.

But there are a lot of other subscriptions I would be looking at very closely right now.

And not just because of the invoice.

08/11/2026

Vibe code life. If you know, you know.

08/01/2026

Comment “beta” to get 500,000 free tokens. More than enough to build your first app!

07/25/2026

AI is useful right up to the moment you let it make a serious decision for you.

That’s the line I try not to cross.

If the task touches money, contracts, health, or technical specs, I use AI to help me think and draft.

Not to approve.

My rule is simple:

1. Let AI do the first pass.

I’ll use a standalone tool like ChatGPT, Claude, Gemini, or Perplexity to organise notes, draft questions, summarise options, or turn messy thinking into a cleaner starting point.

2. Move the work into the right place.

If I’m already in Google Docs + Gemini, Google Slides + Gemini, or Microsoft Copilot inside Office, I’ll often stay there so the review happens inside the real workflow, not across five tabs.

3. Check the risky parts manually.

AI has three big failure points in high-stakes work:
Bias, which means skewed patterns from training data.
Knowledge limitations, which means it may miss recent or specialist information.
Hallucinations, which means it can state something false as if it were fact.

4. Verify against trusted sources.

For legal, financial, medical, or technical work, I check the output against current documents, official guidance, qualified experts, or source material.

5. Treat it like an assistant, not an authority.

Helpful for drafting.

Never the final sign-off.

That one habit avoids a lot of expensive confidence.

How do you verify AI output when the stakes are real?

07/25/2026

The idea that AI is neutral lasts right up until you ask it the same thing three different ways.

Then the pattern starts waving at you.

A model might prefer one kind of language.

One kind of candidate.

One kind of solution.

Not because it has opinions in the human sense...

But because bias gets baked in through training data.

That just means the examples and information it learned from were uneven, incomplete, or skewed in a certain direction.

So when people talk about ChatGPT, Claude, Gemini, or Microsoft Copilot inside Office like they’re perfectly objective machines, I get a bit suspicious.

That’s like calling a mirror neutral when someone has already drawn on it.

This matters more than people think.

Bias is not some side note for technical teams.

It shows up in everyday output.

Hiring drafts.

Performance feedback.

Customer prioritisation.

Even internal business assistants or industry-specific AI agents built for one company can repeat the same blind spots if nobody checks them.

The practical fix is not panic.

It’s review.

Compare outputs.

Question patterns.

Add more relevant context.

Treat AI as an assistant, not an authority.

And if the result affects anything important, verify it before you act on it.

Have you noticed certain patterns showing up in AI responses more than once?