Popularity isn't safety

A thread doing the rounds lists ten open-source add-ons you can bolt onto your AI to make it far more capable — thousands of stars between them, one-line installs, a slick pitch. The message: grab these and go.

Here's what the pitch skips. A team of researchers scanned about 31,000 of these skills across the two biggest marketplaces. Roughly one in four carried a security vulnerability — data exfiltration and privilege escalation at the top of the list.

Read the number the way the researchers ask you to. They're explicit that 26% means 'patterns that warrant review,' not '26% are malicious' — it lumps genuine malice in with sloppy code and grey areas. A separate scan of one brand-new marketplace put actual malicious payloads nearer 17% in its first weeks. Either way the direction holds.

Anyone who's run supplier onboarding already has the reflex. You don't wire a new vendor into your payments run because the website looks good and the logo wall is long. You check who they are, what they can touch, and what happens if they misbehave.

A skill is a vendor. Except this one you're handing read-and-write access to your files.

So the keeper isn't 'avoid the marketplace.' It's: test it on something throwaway before it touches real work, read what it actually does, and treat a high star count as marketing, not a clean bill of health. Would you approve this supplier if it were a person?

Sell the brakes, not the autonomy

An essay making the rounds argues enterprise AI is selling the wrong thing: nobody wants a tool to help them work, they want the work done — so stop selling an interface, sell the finished outcome, and let the agent run in the background.

It's a vendor essay, so the numbers arrive pre-loaded. A tidy '25–30% adoption ceiling' that I couldn't source anywhere outside the pitch itself. The figure that does hold up is uglier: MIT found 95% of enterprise AI pilots delivered no measurable return, largely because the tool sat in a window nobody opened.

Strip the sales gloss and one idea earns its place. The bit they call a 'control plane' — where the owner can pause the agent, pull it back, approve what it changes — is the actual product. Not the automation. The brakes.

Finance has a name for this, and it's a century old. You don't let the person who books the entry also sign it off. An agent working unwatched inside your system of record is an unsegregated process with better branding.

Build the pause button before you build the autonomy. What would you most want to be able to stop?

A formatting fix for a positioning problem

A thread from someone who's spent 18 months doing AI SEO for B2B companies makes an unfashionable claim: most of the current gold rush — buy an 'AEO' tool, write an llms.txt file, chunk your paragraphs just so — is fixing the wrong thing.

You don't have a formatting problem, he says. You have a positioning one. An AI won't recommend you into a category the market doesn't agree you belong in — and no amount of clean markup fixes a brand that's filed under the wrong league.

His mechanism is one every finance team already runs under a different name. You get picked when your site, your customers, and independent third parties all tell the same story about where you belong. That's four-eyes: the same claim confirmed by sources that don't depend on each other. When they agree, the model's confidence rises and it names you. When they drift, you're a coin toss.

The keeper reorders the whole to-do list. Structure earns you eligibility — the AI can lift a well-built page. Positioning earns you the pick. One is a template you apply in an afternoon. The other is a year of being consistent about who you are.

Which one have you actually done?

Adoption beats architecture

A thread going round reports on 200 'company brain' AI rollouts — one builder set them up and wrote down what made them stick. The headline finding: almost none of it was the architecture. What predicted success was temperament. The people who won could 'tolerate ambiguity and keep moving.'

That's exactly the instinct I've spent years in finance training myself out of, so I read the rest closely. The engineers did worst. They asked all the right questions — where does the context live, what if two files disagree, how are permissions enforced — and then questioned themselves into paralysis. They wanted guarantees from a system that doesn't hand them out.

The less-technical users just started. Draft the email, fix the tone, upload the thread that was missing, judge whether it helped, repeat. Closer to onboarding a new hire than configuring a database. You correct it into usefulness. You don't design it into usefulness first.

This isn't only vibes. MIT's 2025 study of enterprise AI found 95% of pilots delivered no measurable P&L — and the reason the authors give is that the tools couldn't retain feedback or adapt. Not model choice. Adoption and correction were the gap.

The catch, and it's a real one: 'move fast, mistakes are reversible' has a hard edge. With client or regulated data, some mistakes don't reverse. The speed has to stop at that wall, every time.

So I've split the rule by layer. Using the system: keep moving, correct later. Changing its structure: prove the scope first. Where's your own line between the two?