Crank.report Survey

…a German team built [Kolibri] with the EU AI Act in mind “from the ground up”, and in Aleph Alpha’s own evaluation it scores above every compared model of its size in both languages.

Aleph Alpha Kolibri: How the Sovereign German LLM Works

I’m interested in doing a version of what we did with the Civil Society Offer Center with Eau friendly technologies. We’re wrestling with an architectural problem as a part of that but we’ll get there.

Large language models were built to talk to people. Jev is built to be called by machines. You give it some context and the exact questions you need answered, and it replies with structured values instead of prose. Because the shape of the answer is fixed by your request, Jev cannot hallucinate a format or emit an invalid type — it can only return one of the options you defined, a score on the scale you gave it, or a probability between 0 and 1.

What is Jev? TypeSafe’s System One model explained · Jev by TypeSafe AI

Could be interesting to experiment with this in any of our rubrics but particularly thinking of rubrics for scoring bits of code that could be shared in a marketplace setting.

From a technical point of view, the use of NixOS was a key one. Its use of the Nix package manager means that each package is installed in its own directory with immutable and signed contents. That alone means that the setup is incredibly easy to reproduce across multiple machines, something that is an obvious benefit when dealing with different facets of a government. It’s been suggested that up to 90% of a configuration can be carried over from one deployment to another. And as a bonus, NixOS can also run on computers that Windows 11 would usually balk at.

US sanctions force The Netherlands off Microsoft and toward alternative NixOS-based software ecosystem — trial programs running now, first release expected at end of 2027 | Tom’s Hardware

Could this be a model to work with small organizations?

Today, a designer can generate polished images, presentations, variations, and even entire visual directions in minutes. The ability to produce is becoming abundant—and when production becomes abundant, something else becomes scarce.

Judgment.

AI Has Changed What Makes a Great Designer - Core77

Applicable to many different kinds of talent.

People make pages, prototypes, reports, visualizations, and small apps with agents every day. Too often the result stays trapped in a local folder or a conversation. Spacefast closes that last mile: publish the files, get a live URL, and send it to the person who needs it.

About Spacefast | Spacefast, via

I’ve been thinking a lot about how you construct an experience for people that uses saved artifacts like way stations. This kind of of tooling could be used to support that.

The relevant question is what the load-bearing seams reveal about the assumptions underneath the conclusion, and whether those assumptions remain intact once the smoking gun is incorporated into the model.

Claude’s load-bearing seams

Oh so good.

Every engagement starts with a two-day visit. The team asks business leaders to ignore AI and name the biggest levers in their business. It then works on whichever that turns out to be.

OpenAI’s Colin Jarvis says enterprise AI is stuck on deployment, not models

Small nonprofits simply don’t have the money to do this. I keep thinking about how we bring the methodology we pioneered at Caravan Studios to this particular problem – identify levers by cohorts and then using them to help move a larger group forward.

It’s not about protecting us from the unlikely sexy apocalypse. It’s about protecting us from the far more likely, and far more numerous, stupid apocalypses along the way.

The Apocalypse will not be Sexy

A frontier AI lab showed up with a check large enough to make the question “should we sell our users' data?” feel rhetorical. Not “maybe we can anonymize and aggregate.” Not “only if users opt in.” Just: here is the number. The number at which the internal debate ends and the legal team starts drafting the “continued use constitutes acceptance” language. That number is the new definition of fuck-you money. It is the amount required for a company to accept the lab’s terms and then turn around and tell you that if you object, you can fuck off.

Dear Customer, Fuck You

I’m just going to leave this right here.

The playlist features tracks by Ornette Coleman, Coltrane, Dizzy, the Bill Evans Trio, Monk, Mingus, and Miles Davis.

Two hour mix of the jazz in Haruki Murakami novels

Oh this is wonderful.

Next boring tiny tool

I’ve been using Epilogue for a while to keep track of the books I read and finish, and the big list of books I want to read. The problem is I never open the too read books when I’m on Libby to check out my next batch of books.

I did a little bit of reasearch with Gemini and learned I can make a script that can check for availability on Libby or at my branch library and send me an email or drop a note in my Obsidian vault when a book on my list is available. I think I can also add a badge to the page to show available books so I can use the published page itself as a reference.

Now I’ve got my next boring tiny tool. I will report back when it is done.

The problem from skipping laying out the constraints is that you get a disjoint patchwork that randomly prioritizes some interactions over others. AI exacerbates the draw to wackamole design. It encourages prompting “Make X more prominent” or “Add an affordance to do Y”. In the end, more users are confused.

Ref - How I Design with AI.

This is true for so much editing and change.

So the collateral chain is short. A European lender is ultimately underwriting one American company’s continued appetite for compute it does not own. If the lease goes, the chips stay.

Lambda inks $1B private debt for Nvidia chip deal

This looks like large corps moving the risk.

Modern coding agents have become so effective at finding flaws that the slightest hint at a new bug can be enough information for them to find it, something Anil has been able to demonstrate using his own agents, switching to DeepSeek V4 Pro⁠ when Claude Fable refused the task.

Just a rumor of a bug is enough

If you expect your students to choose books, ideas and genuine learning over shortcuts and distraction, you must be a bit courageous, too. Return to the basics. Cut through the haze of abstraction, the webs of scholarly terminology, the pretense of professionalization and attempt to recover the basic human experiences that can move us. Remain a student in earnest.

I let my students watch me struggle with a masterpiece. It changed everything.

This kind of engagement is also required of managers.

Good leaders do more than issue instructions. They share context, explain the desired outcome, set boundaries, and respond to what comes back.

The same habits improve my work with AI. A good prompt helps, but a shared working context helps more. Examples, corrections, and reusable instructions reduce misunderstandings. Over time, the system becomes better aligned with how I think and what I need from it.

The investment is not in pretending that AI is human. It is in becoming better at expressing intent.

Working With AI Feels More Like Leadership Than Coding — Allen Bargi

This.

There are legitimate conversations happening now to explore open source harnesses (opencode/pi) and open weight models at the company in order offset the costs associated with going through a standard provider.

Is the industry ready for tokens-constrained work? | Hacker News

From the comments on an article. I’m interested in how we set these harnesses up for small civil society organizations.

Being efficient with tokens doesn’t mean using fewer of them overall. It means making sure the ones you do use go towards the thing you actually asked for.

So let’s look at what decides the price of a token, then what decides how many of them a session sends, and along the way, what that means for how you run a session.

Maximizing the value of your Claude Code sessions | Claude by Anthropic

Good tips for Frugal AI usage. Written for developers. But easy to understand the principles and apply them to a variety of AI work.

The main object of AI security policy should be to understand this picture and figure out how to minimize overall net harms given AI’s affordances for both attackers, defenders, and victims given surveillance information about both groups.

We urgently need a coherent national AI cybersecurity policy

We need to think about this for civil society. How are investing in the ability of organizations of all sizes to operate in the changing cyber security landscape.

Why not one agent for everything?

This might be pointless, but the main reason is least privilege. Some agents don’t need access to GitHub, deployments, or my calendar. Why increase the blast radius and potential for mistakes if I can avoid it? Just like humans…?

Chad Arimura | My agent setup

How to decompose a task into how many agents is something I think about all the time.