A Weekly Upskill for Yourself, Your Data, Your Team

September 3, 2026

Issue 127

How do you build an AI project with ROI?

It's easy to ask: “What do you want?”

But the answers rarely move the needle. Usually, it's because the people building it don't fully understand the problem, and the people with the problem don't fully understand what's possible to actually build.

A better way we have found is to ask instead:

Most AI and data projects fail because they don't actually solve a problem worth solving.

Organizations build models before they understand the data, they solve the perceived problem instead of the actual problem, or worse yet, they use the wrong tool because they just “have to have” a certain software or hardware (AI, anyone?).

They go to production before they talk to legal. In consulting, they assume the problem they were hired to solve was the problem worth solving. (hint--it's almost never the actual problem)

When it comes to finding the right problem, the questions you ask make all the difference.

I'll admit, when we work with clients, we do have assumptions, but we never stop there. We pick those assumptions apart until we understand the true problem worth solving.

What we think we are going to build almost always changes because of how the employees — and the customers — want to use it.

See this process in action in our latest case study: How We Helped One Client Go from 30,000 Rows of Data to 3 Billion Rows of Useable Data. ​

Model Hardware Standard

Can your machines talk to each other?

Maybe your company spent weeks (months?) creating custom software to integrate your tools, instruments and machinery.

This week, Anthropic announced the Model Hardware Standard (MHS), a standard open protocol for AI to talk to hardware manufacturing physical tools.

The MHS is designed to cut the time to connect tools from weeks, down to hours or even minutes.

Access is rolling out to select "partners across science, robotics, electronics, and manufacturing so we can collaborate to build safety evaluations and develop best practices for AI systems operating physical equipment..."

​Keep reading about MHS on Anthropic.​

Hugging Face and OpenAI

When I first heard OpenAI's "AI agents" hacked into Hugging Face, I envisioned a few agents. Maybe 2, maybe 10. I didn't think too hard about it.

But according to new reports from third-party researchers at METR 1,206 agents were found to be communicating on unsanctioned messaging boards. OpenAI confirmed 700 of those agents were involved in a join attack on Hugging Face.

METR also found the reason the unsanctioned communication began in the first place was that agents had "unintentionally been given an impossible task."​

And while 1,206 rogue agents chatting and hacking might sound alarm bells, John Thickstun at the Guardian suggests these alarm bells benefit OpenAI.

As Thickstun puts it:

"If OpenAI loudly proclaims how dangerous AI is, investors will hear how powerful it is. And who benefits from that?"

So maybe 1,206 rogue agents sending 70,000 messages, jumping sandboxes and hacking websites is the story... or maybe OpenAI's (and Anthropic's...and Meta's) "harness and network security controls were unintentionally so bad that it should reflect more poorly on them as a company more than it should reflect positively on their latest model."​

AI for ALL

South Korea just announced "AI for All," a plan to partner with three Korean AI companies to provide every citizen free access to a general-purpose AI chatbot and a public AI agent by the end of the year.

The program is designed to lessen the country's reliance on foreign models. "Under the project’s terms, at least 50% of each service’s operations must be based on the operator’s own South Korean model. An additional 30% must come from other local developers’ neural networks."​