Sten Vesterli's Blog

Do you need the latest and greatest? Or can you make do with less?

We had another “DeepSeek moment” last Friday with the release of DeepSeek v4. It’s open source, open weight, and on par with most models except the very latest.

I’m still seeing too many organizations unquestioningly upgrading to the latest and most expensive AI model as soon as it is released. It reminds me of 1999 at the height of the dot-com mania when every little web shop had to have an Enterprise Edition Oracle database.

If you are running AI in production, it is incumbent on you to find the cheapest solution that will meet your needs.

Are you willing to do what it takes? Some people are very dedicated to their mission and goals, some less so.

When the shooting started at the White House Correspondents’ Dinner, one of the bodyguards immediately got up and positioned himself in front of the president, ready to take a bullet if necessary. That is a man dedicated to his mission.

Scott Adams said: “If you want success, figure out the price, then pay it.” That’s more profound than it sounds. We all have things we want, but just visualizing your goal won’t get you there. You have to figure out the cost and honestly evaluate whether you are willing to pay it. That cost is rarely in money – it is in time you can’t spend on other things, or relationships you allow to lapse.

If you only know what you want but not what it will cost, you only have half a goal.

Are you spending enough time planning? There is eternal tension between planning and doing, but we’ve come to idolize moving fast (and breaking things).

One of my sports is orienteering, where you navigate between control points, selecting your own route. For every leg, you have to decide whether to run straight through the forest or take a more circuitous route along trails or following a clear leading line. The beginner will just glance at the map and start running. The experienced practitioner will spend precious seconds carefully studying the map in order to make a better decision.

Design Thinking is one approach that forces you to spend more time planning. You don’t have to buy into the whole “Double Diamond” methodology, but make sure you always generate several ideas before you start executing. Those seconds and minutes are well spent.

I was just reminded of a wonderfully exact German word: Lehrmeinung. It comes from Lehre (teaching) and Meinung (opinion) and means “the official doctrine as taught.” It stands in contrast to “gelebte Praxis” – how things are actually done in the real world.

My recent via ferrata climbing course in the Dolomites was taught in German, and our instructor adroitly covered both.

That is why in-person teaching is still relevant in an AI world. Full-time instructors who teach basic React, Python, Kubernetes, Azure and Spring Boot in the same month will be replaced by AI learning. But experienced IT professionals who enjoy teaching the stuff they actually work with will be more in demand than ever. Because they can tell the difference between Lehrmeinung and Praxis.

I fell off the mountain last weekend. In my defense I’ll say that it was on an overhang at the end of a long climb. But I was properly secured and fell less than a foot before the rope stopped me. I clipped into an anchor point with my rest sling and took a break. Then I tried again, and succeeded.

You grow when you try new and hard things. But you need to have security mechanisms in place to prevent you from hurting yourself (or your organization). In an IT context, having a human mentor is a good way to do this. Don’t depend on AI to keep yourself safe.

The gap between what your C-level colleagues think IT can do and what you can actually deliver is getting wider every month. This is not a communication problem. It is a structural feature of a function that moves faster than the business around it.

If you want to make the maximum contribution to the success of the organization, it is incumbent on you to keep everyone informed about what is possible.

My journaling app just got AI assistance. It’s an amazingly stupid idea.

The whole point of journaling is that writing encourages reflection and deep thinking. It is not about producing a certain number of words every day.

I started journaling 12 years, 5 months and 22 days ago. It’s helped me reflect upon and improve my life immeasurably. If you’re not journaling, I encourage you to start. Without AI.

“Let me download Axios for you to make that request and parse the response” That’s how the attackers got in.

With AI code tools, supply chain attacks are becoming 10x more likely. The friendly AI offers to solve the specific problem the developer is working on, and he will take its suggestion.

We have seen enough supply chain attacks that experienced IT professionals have developed muscle memory that kicks in whenever they start typing “npm install …” Before pressing enter, their brains will remind them to check that they are asking for the officially sanctioned version.

You cannot depend on developer training and procedures when AI is writing half of the code or more. If you haven’t already blocked direct npm and similar repository downloads, do so now. And keep a human in the update loop for your blessed repository and implement a delay.

The scarcest resource is leadership attention. You can exchange money for manpower, storage, compute, network bandwidth. But leadership attention is the ultimate fixed resource.

I was reminded when I read the latest on Donald Trump’s hole in the ground, aka the new White House ballroom. He claims it won’t cost taxpayers anything because he can shake down favor-seeking businesses to pay for it. But it will cost American citizens what they can least afford to squander: The attention of the President.

Every failed project I have ever been involved with (a few) or have heard about (many) suffered from a lack of leadership attention. Every highly successful digital transformation has had leadership attention in spades.

If you are an IT leader (or any kind of leader), dole out your attention very carefully.

How do you talk to your AI? I speak nicely to mine.

It’s not because I hope it will remember me as one of the friendly ones and kill me last. It’s because the anthropomorphism works both ways. Because they seem human, we assign them human traits. But it also means that the way we interact with them bleeds back into our conversations with humans.

If you are curt or even aggressive towards your AI chatbot, that affects your interactions with everyone else. Be nice. To both humans and AI.