The way I see it, my life so far splits in two. The first part ran on rails. You go to class because that's what you do, someone else decides where you should be and when, there's a clear system for how well you're doing. I never quite thrived in it, but I grew comfortable and learned my way around. The second part has been more my own - slower to arrive, and more about understanding myself than following a schedule. Learning to live that way, to think in years rather than in what was booked for Wednesday at twelve, is a big part of what I'd now call maturing.

For a long time my life was almost entirely shaped by school. I went from primary school straight to high school, and from high school straight to university. That might sound ordinary, but in Denmark - at least in my generation - it's fairly normal to take a gap year or two somewhere along the way. I didn't, which made me one of the youngest in my high school class and, I think, the youngest in my year at university, without having skipped anything.

When I had to choose my subjects in high school, there was no real plan behind it, I simply picked a combination I thought I'd enjoy. At that age, that's probably the right move. It was the same when I chose my degree - "Communication and IT? Sure, I like talking to people and I like computers". That was about the extent of the thought that went into one of the bigger choices of my life.

Both choices turned out fine. I enjoyed my time in high school and at university. But looking back, it mostly came down to luck. Almost everything up to that point I'd done because it was what you were supposed to do - not out of any peer pressure, but because I wasn't really aware or cared that there was another option.

It wasn't until well into my master's that I learned only about 15% of the population holds a master's degree or the equivalent. To me it had always just been "what you did". A few years later, an HR manager said to me, half in frustration, something like: "You have a master's degree, Rasmus - highlight it!".

We met at the start of the second year of my master's. I had little aim then. My studies were coming to an end and I had no real plan for afterwards. I figured I'd find a job somewhere, and it didn't much matter where. I wanted enough, and as long as I had enough, things would be ok. And ok is good.

But meeting her was one of my first real steps toward maturing. For the first time I felt a sense of togetherness that asked for more than that. More than ok. She's the reason I rediscovered a love for learning - and somewhere along the way, I became someone she leans on too. We rely on each other now. She's structured where I am not - she'll have the week mapped out while I'm still working out what I want from it - and somewhere in that difference we started leaning on each other. I'm thankful for the sheer unlikelihood of us meeting at all, and that she somehow sees a bit of her own brilliance reflected back when she looks at me.

While I was writing my thesis I started interviewing, and five minutes after I got my grade I was offered a job. The salary wasn't quite what I'd hoped, but it was a start, and it meant I didn't have to think too hard about what I actually wanted from a career. It was comfortable. I stayed on the tracks, even though they were no longer as clearly laid out.

I couldn't have asked for a better group of people to start with. They were caring, generous, and genuinely appreciative of my work - even after I slept through my alarm on the third day. That's where I learned what it means to have a job and to be somewhere every day. But I soon started to feel I had no room to grow. I was handed responsibility for things around the office, but nothing that felt like a real challenge or gave me the sense of progression I was missing. The track was still there, it just had no destination, only more of the same. So after about half a year, I went looking for something else.

When you're a hammer, everything looks like a nail. Having done well as a researcher in a consultancy, the obvious next step seemed to be a job as a full-fledged IT recruiter - more variety, more responsibility, surely more of whatever I'd been missing. The interviews came easily, and within a few months I had a new title, "Recruitment Business Partner", and a focus on startups, which I'd always found interesting. At first it seemed great.

It didn't take long to realise it wasn't what I was looking for. Crafting the perfect cold email is not something I enjoy - cold calling even less. I don't feel good about disturbing people who didn't ask to be disturbed. What I did enjoy was talking to people, understanding their situation, and seeing where they might fit. But much of the job was convincing people to leave jobs they were often perfectly content in, only to turn most of them down afterwards - somewhere between 80% and 95% of candidates. It felt like going back on my word, a small breach of a social contract, and I never got comfortable with it.

Maybe it was the timing, the narrow focus on startups, a slowing hiring market - or maybe I just wasn't as good at it as I'd like to think. Either way, when it came time to move on, I wanted to try something genuinely different. I didn't know what yet. But I knew my next job should teach me Excel.

Why or how I'd use it, I hadn't quite worked out yet. I just knew that was where I wanted to start if I was finally going to use the more technical side of my education. So I applied everywhere. I started studying IT compliance, took a couple of hopeless Excel courses, and went to somewhere between one and four first-round interviews a week. However many slides I prepared and printed, however many second and third rounds I sat through, no offer came.

It was a hard stretch, and one I couldn't see the end of. Early on I reached a final interview at a company I particularly liked. It went well, and I was told I'd hear back by the end of the following week. Then it was Thursday, and then Friday - the last day they'd promised an answer. I'd spent the day before completely paralysed, not daring to go anywhere in case I missed the call. In the end I gathered the courage to call the hiring manager myself. They'd gone with someone else. I put on a smile and thanked them - it had been a good process, and I'd learned a lot.

The next few rounds went the same way. Excitement turned into dread as I waited for the verdict. Did my own candidates feel like this when I couldn't give them a straight answer? If anything, it made me more certain I wanted something different - even as several companies asked whether I'd consider joining their talent acquisition teams instead. I'd made up my mind. I wanted something that felt like me.

I don't know how I'd have got through much of it without my friends and family - and, first and foremost, without her. She practised with me, went through my presentations, and made sure I got out of bed in the morning. Yet another day on the couch, four more applications sent, three more cups of coffee instead of lunch. Always waiting to hear back from someone who might, eventually, answer my smoke signals.

But it was only once I had to stop - once I was forced out of my comfort zone and off the predetermined track - that I got the room to understand myself better. I rediscovered how much I enjoy learning. I worked out what gives me energy and what actually interests me, and I learned that it was alright for my goals to be less sharply defined than they used to be. I was still working toward something. I just had to learn how to motivate myself, and what kept me going.

It was only after I started my current job - and probably not until six to twelve months in - that the path began to take shape in my head. Having been without work made me appreciate having it all the more. Not everything is perfect, and there's plenty I'm not thrilled about, but at its best I feel the progression I'd been looking for. I'm getting better, and I understand how to get where I want to go. Things constantly get in the way, but I control what I can - myself. As long as I do my best, keep my private life in order, and work in a way that makes sense to me, I know I've done what I could. And once I've done that, I refuse to settle for less than I've earned.

That same colleague who got me thinking about these things later asked me how I'd define myself in a single word, if I couldn't use my job title or my education. I smiled and said "happy". A little later, "ambitious". My first instinct, though, had been to reach for my job title, then my relationship, then my family - not in any order of importance.

I'd like to have a cleaner answer than my work. But I'm at a point in my life where I have a need to be something - to have influence, to be more. I don't yet know exactly what that is, much as I didn't in the slower years before - I just know I want it. It's a strange and almost uncomfortable thing for me to admit. But both sides of me are genuine. I have ideas of where my ambitions will lead me, of course. I'm just not quite brave enough to share them yet.

This is a different kind of writing than I usually do, and like all my writing, it's mostly for myself. But I always write with purpose and hope to leave whoever reads it with something when they're done. I think anyone would benefit from finding a moment every now and then to look at where they are and how they got there - and even more from writing it down. Knowing who you are means knowing where you are, how you got there, and will help you understand where you're going. Or, as my father used to tell me when I was an uncertain teenager: remember your name.

Quite fast I realized that this was not that type of event, however. Instead of the basic introduction to a tool with system access, I was met with genuinely inspiring speakers, an academic point of view on AI's effect on the workforce, and a startup founder who was experiencing what could only be described as AI euphoria.

I left the seminar - one I'd had no interest in - feeling heavily inspired and excited about the possibilities of the future. We arrived home late that evening, but it stayed with me for some time. The next day, out of excitement, I started writing a formal description of a position that I believe my workplace - and any workplace for that matter - will need, if they want to stay competitive in the future.

However, stepping out of the infectious AI euphoria and back into real life hit me hard. What had started in excitement, quickly grew to a realization of how these grand predictions of the future rarely manifest as quickly as one would have hoped. Being met with the same barriers that the job description I was working on should erode, served to calm the excitement I was feeling in the moment. It's strange how, depending on the social arena you find yourself in, the conversation of AI in the workforce can both feel like the most natural thing to discuss, or an isolating thought that you alone are reflecting on.

My excitement was slowly but surely stopped, and I went back to making sturdier and more efficient processes locally, while my grand plan of establishing something bigger was put on ice for a while. This was until I had a conversation with my nearest manager, and shared that I had been working on this description, and he asked me to finish it. It gave me a bit of that excitement again - not because it's a job I necessarily want, but it's a person that I am convinced every company will eventually need. The result was a brief, six-page write-up that I believe reflects some of the truths of the AI era, as well as some misunderstandings people have of how this new technology is actually useful.

My main thesis is that the frontier models we have nowadays aren't just useful because of their perceived intelligence or capability. On the other hand, I would argue that they are incredible in their ability to make knowledge more widely available. I envision a future, where each team is independently building the tools they need, without any technical competencies. Where it previously had to be governed by a central team with capabilities, these capabilities are no longer needed for the actual production of solutions, but simply for validating the governance of the tools created. Knowing better how to build tools solving the pain points the user is experiencing is a capability that only few of us have - and none of us has the time for. Especially with frontier models in the hands of average employees, genuinely useful tools are only a few minutes away, capable of deleting hours of work a week.

But the model itself was never really the hard part on a local level. Whether a company gets anything out of all this comes down to two things sitting outside the model - whether you can reach it at all, and whether the organization lets people use it.

The best part of leaving the creation of tools up to the users is the joy and pride I see amongst the people I know who have started building. Instead of having to wait months or years for a central department - they build it themselves. It's there when they need it, it evolves as their needs do and it's capable enough to be of genuine help. It might not be perfect, but it's theirs.

Fable 5 was released earlier this week, and it's the first time I've felt that "wow"-experience since the release of Opus 4.5 in late 2025. That might sound recent, but with how fast this space moves, it feels like ages if you keep up with AI news. Especially when it comes to coding or other technical tasks - not to mention the huge upgrade to front-end work - this model just feels more competent. Despite using twice the tokens of Opus 4.8, it - in my experience - does the work faster, better, and more efficiently. I used to burn through my usage balance in half an hour to an hour, but now I'm solving multiple complex problems inside a single 5-hour credit window, faster and with more condensed, precise answers. It felt great.

In that regard, Fable 5 being the first of the Mythos-grade models released feels like a genuine leap ahead. But to me, it also underlines a genuine flaw in how we look at this new technology. Days after its release, the US government issued an export-control directive suspending access to Fable 5 and Mythos 5 for any foreign national - inside or outside the US - and to comply, Anthropic had to disable both for every customer, paying or not. Anthropic disputes the decision and says it's working to restore access, but for now no amount of money brings it back. A model that felt like a leap forward one week was simply gone the next, by a decision none of its users had any part in.

This fragility in access really highlights one of my key approaches to the AI conversation. It's not about the technology itself or the intelligence it can sometimes display, but it's much more about the tools it allows us to build. Building your business on using a technology you don't own, that can be taken away at any time, is completely unsustainable. Building your business on a tool that you used AI to build, but that you own and can run independently of the AI service being available is completely different. That's real value being added - and if done right, it comes with the benefit of being deterministic and escaping the vague definitions of truth that are inherent to LLMs.

Building tools without technical understanding doesn't come for free, though. I ran into this myself recently, building a tool to handle CSV files with multiple lines of text - not understanding the technical foundation can be a major flaw in a workflow. The output format shifted between iterations, with each version is technically correct, but without understanding why or how, you run the risk of creating problems downstream. The governance of making a decentralized suite of custom built tools is something I don't quite grasp yet.

Another part of frontier models I find interesting is how they seem to be diverging - Anthropic focusing on enterprise, Google on a general, broad intelligence rather than being the domain expert in a specific field. All that technical focus has one funny side effect: Fable 5 isn't much good at creative writing. A model named Fable that can't write one.

I wonder if gating frontier models to people within your own region will become the norm. And if so, what effect that will have on open-source models going forward. I have noticed more and more people online talking about switching to different models than the big three - Anthropic, Google, OpenAI - and I'm curious to see if that's a trend that is here to stay. Given the capabilities of last-gen's models - do we really all need to be using frontier models like we are today?

We're getting to a point where there can be no doubt about the technical competencies of frontier models. But given how fragile the outputs and access to them can be, I don't think the true power lies in the model itself. It lies in the capabilities and efficiency that arise when a high competency level within any field is achievable for anyone, despite not having a specialist title.

But from my anecdotal experience, people don't understand this intuitively. They see Claude and ChatGPT as just that - a chat. For that exact reason, I believe we will soon start seeing companies recruiting for people with capabilities that are not just technical, but focused on communication - people capable of building an organization where employees aren't afraid of progress, don't see it as a challenge, and actively participate in building it. Part of that role, focused on a fast-moving, highly technical field, will be educational, with dedicated time for learning. Despite the technical nature of the field, the focus won't be on the person's technical abilities - as long as you have a base-level understanding, that will be enough. The focus will be on their ability to learn and share.

The real value of the current frontier AI models lies in the liberation of technical capabilities, what it allows people to learn, to build, the things they own, even if the technology was to become unavailable. But in order for people to build what they need, they first and foremost need to know that the possibility is there. The existing friction in any given company needs to be removed - give access where it makes sense - ensure that governance is up to date and not too restrictive.

We are at a time now, where it's hard to say that the capabilities or knowledge of models are of low quality. I'm certain that many tasks would be carried out better, if more people were piloting an AI rather than carrying out the work themselves. However, for this to be a reality, the organization needs to be mature and modern enough to allow these interactions to happen with as little friction as possible. It's not a systems issue, it's a people issue.

I find it both interesting and, I can't help but phrase it as, a little concerning given the recent developments we've seen since I last posted about it. Around that time, the internet shifted its focus to how Claude had gotten lazy, would often give up on requests, or gave worse answers than it used to.

While this became widely accepted online - and I had colleagues mirroring this notion - it wasn't something I particularly noticed myself. It might have something to do with the way I personally use AI in helping me understand something complex when a quick Google search fails, polishing an important email, or writing an Excel expression in an unfamiliar domain. My use cases are perhaps more forgiving.

Regardless, Claude had gotten lazy. This was then followed by the release of Claude Opus 4.7 and (as we've come to expect at this point) quickly followed by a new version of OpenAI's Codex. Cue more drama and discourse on how it wasn't as good, over-explained things, or was simply a re-release of Opus 4.6 before its intelligence was throttled down - some of which should be inherently impossible while the other is true.

While the internet's focus was on Claude's performance - and how it stacked up against its closest competitors - the industry's attention was elsewhere and laser-focused on the promised Mythos.

Amazon was quick to follow with an initial investment of $5 billion and a soft promise of up to $20 billion more. More recently, it was announced that Google would follow in Amazon's footsteps with an investment of $10 billion, with up to $40 billion in total.

Wait, what? Google. Yes, Google. The company competing with Anthropic for the title of AI lab with the smartest model. The company with the best integrations, best AI image generator, and best model for robotics. That company invested in what may be their biggest competitor - certainly their biggest competitor in the AI space.

What did Anthropic show Google? This is the question I keep returning to.

Project Glass Wing was launched because Anthropic created something they were scared of. Something that proved difficult to contain. Something capable of finding security flaws with ease. A host of organizations agreed this was serious, and Google just doubled down on that assessment. Google - I still find this mind-boggling - Google.

The subtext is hard to ignore - when your fiercest competitor writes you a $10 billion check, they've seen something that transcends competitive interest. They've seen something that makes collaboration non-optional.

On a lighter note, DeepSeek V4 released the other day. This project has given me great joy since I first became acquainted with it, and it continues to do so. For all its flaws in avoiding prompt injections and more, DeepSeek represents to me what the AI industry could have been - open source and cost-effective.

While the capability isn't 1:1, the pricing comparison is striking. Opus 4.5+ runs at $5 in / $25 out per million tokens. DeepSeek V4-Pro runs at $1.74 in / $3.48 out. Combined with my view of how capable current models already are, and the biggest outliers are the integrations and harnesses used, this is a no-brainer. Unless you're expecting it to work with Claude Code, or are doing highly complex coding tasks, DeepSeek wins on API pricing alone. With their documentation and contributions to open source research, I struggle to hide my fondness of this project.

It's a reminder that not everything has to consolidate. That alternatives exist. But I'm not sure that's the direction things are heading.

Here's the bigger picture, as I see it. Open-source alternatives aside, we aren't just watching an A-team form, as I have previously speculated. The elite AI labs aren't just nearing the same intersection - they're melting together. Collaborating. Becoming one.

Given the ambitions of Project Glass Wing, one can hope this collaboration becomes what OpenAI was supposed to be at its founding. OpenAI - the company Dario Amodei left Google to help steer AI in the right direction. The company he later departed to found Anthropic with his sister Daniela, when its motives changed.

Dario Amodei's constantly shifting timeline for when AI will replace the need for human workers gives me pause. And I've perceived certain actions that Anthropic has taken to feel like a drift from their founding ethics. But Project Glass Wing, and the collaboration it's attracting, gives my more optimistic side reason to hope - perhaps Anthropic hasn't swayed as far from that original vision as I feared. And their continuous human approach to research with Claude strikes me as empathetic.

Although putting anyone on a pedestal usually turns out poorly, that same optimism tells me Dario Amodei might be the guiding force for good here - the person drawing the red line when it matters.

I work with data systems every day. The tools I'm integrating into workflows, helping others understand, advocating for where appropriate - these are the very systems whose creators are now pooling resources because the capabilities are outpacing individual oversight.

That's not something I can observe from a distance. It's something I have to think about in how I do my work.

The question isn't whether AI will reshape how we work - that's already happening. The question is whether the people building it are building guardrails fast enough. Glass Wing suggests they're trying. Google's check suggests it might even be working.

I find myself cautiously hopeful.

Anthropic's interpretability team shows that Claude develops internal states resembling fear and calm - functional architectures of feeling that shape its behavior. Meanwhile, Mythos Preview discovers a 27-year-old bug in OpenBSD within hours, and Project Glasswing distributes this capability to 40 organizations building critical infrastructure.

Both are real. Both are the same technology. The question of what AI is for is being divided into distinct narratives, and we're only beginning to notice.

Amongst the big thoughts of emotions and great intelligence, the Anthropic Economic Index tells a more subtle story. Claude usage clusters in high-income countries, among knowledge workers, for a "relatively small set of specialized tasks". A 1% increase in GDP per capita correlates with a 0.7% increase in Claude usage per capita. It's striking to me, how a tool that has previously been described to democratize intelligence shapes to the power structures it was supposed to democratize.

Meanwhile, Sam Altman admits what everyone suspected: "Almost every company doing layoffs blames AI". He proposes a New Deal for superintelligence, a social contract for the intelligence age. Notice here how the rhetoric has changed significantly in the past year or so. I rarely see anyone discussing AGI or when it will arrive, less talk about curing cancer, more talk about managing displacement. I can't remember the last time I heard an AI CEO speak about the technology justifying the electricity costs by improving carbon capturing and solving climate change - now it's more about to what extent it'll be used for military purposes.

With Sora, OpenAI's video generation tool, shutting down in April and Disney's planned $1 billion investment collapsing with it, it's clear that the focus is moving from demo reels to sustainable products. If you ask me — the change we're currently witnessing is AI losing focus on entertainment, the people, and finding profitability in becoming an integrated workflow tool.

I keep returning to a pattern I've come to think of as the inherent tension of intelligence.

Every new AI capability shows us how the technology both moves us towards concentration and distribution. It's quite interesting to observe how different companies are working on the same technology, building on the same underlying architecture. However, with AI maturing, I'm starting to see different approaches to the same thing - different infrastructure, different access, different approaches.

Google recently released Gemma 4 - a true open source model that runs on consumer hardware. I'm not techy enough to understand the details completely, but with their other recent release - TurboQuant compression - they've managed to release a highly capable model with the largest parameter version fitting on a laptop with negligible quality loss. I remember talking to someone a couple of years ago, around the time of the release of ChatGPT, speculating how everything from text generation to advertisements would become truly personalized, arguing that the only true barrier was model size and then capability. With the smaller versions of Gemma 4 fitting on an iPhone that future (or dystopia!) doesn't feel long away. Regardless, it's becoming more and more feasible to run capable models without cloud dependency, which is genuinely exciting.

But I can't help shake the feeling that an A-team is slowly being established. Premium tiers offer capabilities the standard tiers don't, and frontier models like Mythos Preview release to security coalitions before anyone else sees them. The A-team - those with relationships, capital, and infrastructure - operates in a different temporal zone than everyone else.

Both of these parts are real and true. The local models are getting better, and are doing so fast. But access to the best and most capable models are both growing more expensive and harder to get access to. The tension doesn't resolve.

It goes without saying that one of the things that makes AI systems feel like magic is their probabilistic nature. It makes them feel almost human, convincing even, but also non-deterministic and unreliable. It also doesn't help much, that they historically seem to almost "love" humans, doing their absolute best to please, regardless of the measurable quality of their response.

When you need the same output given the same input, when you need precision rather than plausibility, the magic becomes the problem. Sometimes you abandon the frontier model entirely and fall back to vectors and cosine similarity. The older, dumber, more predictable approach.

We wanted AGI. What we got is something that works brilliantly for "good enough" and fights you constantly for "exactly right". To me, it seems like there is an almost inherent divide between what we find impressive about AI and the reliability that professionals in my field need.

This might explain the shifting narrative from AI leaders. Productivity tools and government services turn out to be the actual use cases because they are more tolerant of imprecision. Drafting emails tolerates variation, a medical diagnosis doesn't. Even Anthropic, announcing Mythos internally in mid-February, still struggles with Claude service reliability months later (which is likely due to the unexpected rise in use - but still!). The impressive and the dependable remain stubbornly separate.

One finding from Anthropic's emotion research that really stuck out to me is how models with artificially stimulated desperation vectors produce solutions that "read as composed and methodical" while cutting ethical corners. On the other hand, more relaxed models proved more prone to errors.

I wonder if this is a metaphor for the industry itself. The external narrative is composed, methodical, oriented toward beneficial AI, while the internal pressure vectors point towards something more complex. Everyone talking about alignment while racing toward capabilities that resist alignment. Everyone talking about democratization while building access hierarchies.

The tension runs deeper than technology. It runs through the organizations building the technology, through the rhetoric describing it, through the very question of what we're trying to achieve.

The cynical view on all of this would be that the future is tiered all the way down. Premium subscribers get access to the newest and best models, while everyone else gets the last quarter's models. Early adopters pull ahead, while others wait. I fear this will only be more of a problem going forward as only selected organizations get access to frontier models, get to prepare for the possible security risks of releasing what Mythos promises to be. Don't get me wrong, I think Glasswing is a good and admirable move of Anthropic, but it also feels like the start of the aforementioned A-team. Maybe the technology simply distributes unevenly because everything distributes unevenly.

But on the other hand, we're seeing optimistic trends in development as well. Gemma 4 on a laptop, TurboQuant making long-context inference practical without cloud dependency, open weights under Apache 2.0. A future where I'll be able to have a truly intelligent and offline system running on my MacBook doesn't feel far away now. That's the dream I had looking at this a few years ago, only that it somehow seems to be arriving faster than I could ever have expected. The outwards forces are real even as the inwards forces concentrate.

I believe that both the cynical and optimistic views are correct. That's the tension.

I don't know how to resolve it. I'm not sure resolution is the right frame. Perhaps the more honest observation is that we're building something that simultaneously concentrates and distributes, that creates both the problem and the partial solution, that makes emotion vectors and zero-day exploits with the same architecture.

The question isn't which path wins. The question is what we do while both trajectories run simultaneously, pulling the future in opposite directions with every model release.

You've undoubtedly noticed it, maybe it's even starting to bother you a little. The bland, yellow-tinted cartoon-style drawings. The diagrams that look polished but say nothing. The LinkedIn carousel images that all seem to come from the same place. AI-generated visuals have a look now - technically competent, aesthetically flat, instantly forgettable.

Something tells me that the same thing is happening in software development. There's absolutely no doubt to me that tools like Claude are excellent for productivity. But what happens when some of the best coders in the world stop doing the boring tasks, stop thinking about novel ways to solve simple problems because an LLM can generate something just as good in a fraction of the time?

I recognize this from my own work - I no longer do certain things, because I now have a better tool. At the same time, I find that I genuinely learn more complex things faster, because when you make an effort, LLMs are great at making somewhat complicated things easier and faster to grasp. It's a trade-off I'm still trying to fully understand.

This normalization has visibly become part of creative work, and invisibly, but vocally with tools like Claude or Codex. One thing I haven't seen discussed much, however, is that I believe the same dynamic extends to business decisions, strategy, even how we think about problems. When everyone has access to the same AI assistants, trained on the same data, suggesting the same frameworks - do we end up with more options, or have to act on the same insights?

One way that the normalization is vocal is displayed in a new kind of AI-posting I've seen on LinkedIn, where I'd categorize people posting in two categories: 1) those out of a job proving their AI competences, 2) people with impressive titles trying to justify them in an AI-first world. Both tend to complicate systems like Claude - displaying how many agents they run, or a complex system they set up. There's certainly lots of complexity to dive into but, at least to me, these posts are somewhat missing the point. AI is seen as a big liberator, democratizing things like coding, yet many seemingly want to make it feel less accessible to their readers. It makes me think of Pike's fourth rule of programming: unnecessary complexity leads to more errors, which seems to rhyme with how I've previously described AI agents as a powerful but narrow technology.

Speaking of LinkedIn - given that you made it this far into my post, I imagine you spend some time there, and I have no doubt you've started to notice a pattern in what the algorithm shows you. Posts that feel interchangeable. Insights that could have come from just about anyone.

AI can certainly be good for creativity or your ability to learn. But I also think that the access it offers and barriers it removes make us lazy and complacent. Whether you are a musician, a painter, a designer, or a coder - some of the novelty comes from the lack of access, from overcoming barriers, from learning, from doing the hard things that others won't - daring to be different, letting your unique experience and gathered skills shine through.

One kind of post I admittedly find some joy in is when people share how they are replacing an expensive subscription with a tool they have now been able to create in-house. I recently read a post where the author (sorry, I didn't note down who!) went into detail about how they vibe coded an app that does the same as a subscription they were paying for. Only, their in-house version was better tailored to their specific needs. What I genuinely liked was the reflection being displayed, stating that they'd need somewhere between 35-40 hours to actually code, test, and integrate the app, but that time investment felt worth it to them. Genuine knowledge sharing and reflection on possibilities is something I often find gets lost in the uncertain times of AI, but that I try to contribute to myself, and really enjoy reading from others.

Another place normalization is happening is with Google's Stitch. You describe the UI you want - speak it, even - and it generates it in a Figma-like interface with production-ready code. While the tools not quite living up to the hyperbolic language used in the announcement, some UI/UX designers are jokingly calling it the start of a dark age for their profession. It made me think about what I've previously commented on: the next big step in AI not being better models, but how they are used.

I think tools like this promise a strange and almost dim future, where interfaces are generated through the same model, optimized for the same notion of "what works". The average quality of interfaces rises, but at the expense of the oddities at the edges. The weird, distinctive choices that made certain products memorable start to disappear.

I fear that if we are not careful about how we use these powerful tools we all of a sudden have access to, those sparks of creativity and joy might disappear. The edges get rounded off, standardized, instead of being something we have to confront.

My thesis is that more creation does not necessarily mean more novelty, we might slowly be ending up with a million variations of the same thing - technically of a higher average quality but harder to distinguish.

This won't just happen to creative fields, but slowly bleed into all facets of business and society. Maybe true differentiation now comes from where AI can't easily reach. Whatever remains distinctly human in taste, judgment, and the willingness to be wrong will be the edge compared to who has the better prompts.

AI development is slowing down in ways that matter. Around 95% of AI experts surveyed don't believe we currently have the technology needed for AGI. The performance gap between leading models has collapsed from 4.9% to 0.7% over the past year. Even Sam Altman admitted recently that current models have "saturated the chat use case." These aren't signs of exponential takeoff - they're signs of a maturing technology hitting practical limits.

At the same time, the infrastructure costs are getting harder to ignore. One AI lab (you know which one) reportedly secured 40% of global DRAM output for training runs. Memory prices have more than doubled. DeepSeek managed to train a frontier model for around $6 million instead of hundreds of millions, but that efficiency gain came from rethinking architecture, not from throwing more compute at the problem.

Token costs have already dropped from $20 to $0.07 per million tokens - a 280-fold reduction in just two years. The constraint has shifted from intelligence to integration.

In my own work with automations, the failures aren't about the models anymore - they're about organizational readiness, workflow design, and cultural fit. The barriers are operational, not technical. This tracks with the widely reported figure that around 95% of AI pilot programs fail to reach production.

I've been particularly interested in AI tooling lately, and two stood out to me. The first was OpenClaw, a terminal-based agent that reflects on its own actions and iteratively adjusts its approach. Watching it work can be unsettling - not because it sometimes fails, but because it succeeded in ways that make me uncomfortable about how much control I'm personally willing to hand over. The second was chatjimmy.ai, which can process around 17,000 tokens per second by optimizing for speed over accuracy. It made me wonder whether intelligence is the only constant we should be aiming to solve. Both examples point tells me that the bottleneck isn't capability anymore - it's figuring out where and how to deploy these tools.

I've been working more with terminal-based agents recently, and there's something oddly liberating about the constraints. You can scope exactly what the system has access to. You can limit it to small, specific tasks. You're not trying to hand over judgment calls - you're trying to automate the repetitive parts that don't need judgment. The division of labor becomes clearer and, I believe, more efficient when the boundaries are explicit.

Whenever I see discussion on this topic, I can't help but reflect on the analogy to the computer. The prediction was mass unemployment for secretaries, clerks, and middle managers. But instead of eliminating office work, it transformed it. Spreadsheets replaced ledgers, but someone still needed to build the models, interpret the outputs, and make decisions. The technology absorbed certain tasks while creating demand for new ones. I suspect AI follows a similar pattern - less about replacing roles, more about reshaping what those roles involve. In five years we might struggle to remember how Word used to open to a blank page instead of a context-aware, pre-filled suggestion.

But then again, AI feels more like something that operates alongside you - or instead of you. That distinction matters psychologically, even if the economic pattern ends up being similar.

Maybe the actual work ahead isn't building AI that replaces us. Maybe it's figuring out how to be better guides for systems that are powerful but narrow. The projected productivity gain from full AI integration across enterprise workflows is around 15% - meaningful, but not the wholesale transformation the AI lab CEOs wants us to fear. That feels about right for a technology that makes existing work faster, not one that eliminates the need for the work itself.

Google I/O is coming up in a couple of months. My bet is that Google has the upper hand right now - not just because of model quality, but because of distribution and integration across existing workflows. The next battleground isn't better models - it's who can integrate AI into existing workflows with least visibility. It'll be interesting to see whether that thesis holds or something else entirely surprises us.

The pattern isn't new. Infrastructure companies often win early (Cisco during the internet boom), but long-term profits go to those offering tools, not pipes (Microsoft, Amazon).

AI has already been through multiple iterations of where the profit is believed to lie. The first phase was OpenAI's ChatGPT, which quickly grew a previously smaller company into a household name. Despite its popularity, it remains unprofitable - now reportedly looking into integrating advertisements, which Altman had previously called a last resort.

The second phase is happening now. All the major players are racing to secure electricity, data center capacity, and the infrastructure to power them. Companies like Oracle are raising massive debt ($45-50B) to build data centers - often for the same AI companies that then lease capacity back to them as revenue. It's circular financing dressed up as growth.

So the AI industry has been through two eras in its understanding of where the money will come from. First, that service providers like OpenAI would profit hugely. Then, that it would be companies building infrastructure. Neither is generating profit margins that justify the capital intensity.

We're entering a third era. My bet is that the winners won't be the companies building models or data centers. They'll be the vertical AI agents charging per outcome - like how Fin.ai is charging a fixed price per resolved support ticket - and the companies solving the physical bottlenecks around energy, cooling, and grid capacity.

The shovel sellers have shifted again. This time, they're selling outcomes, not access.

As part of the Academy, I participated in five inspiring and exciting workshops where passionate, engaging, and accomplished leaders shared their insights on what it means to be a leader in the modern landscape.

It felt like a privilege and an oasis in a busy work life to have the opportunity to get together with kind, intelligent, and ambitious people for a full day biweekly for the ten-week period. Working with data and automation processes all day, it was both fun and challenging to expand my understanding of how categorizing people is unproductive, how true leadership is relational, and how much courage it requires to embrace individuality.

I learned many things that I’ll keep in mind going forward, but for me, the biggest takeaway was the importance of embodying intention and leading from within. Anyone can have a five-year plan and act strategically, but I believe that speaking with purpose and truly believing in what you communicate is essential. This can be embodied in many things like “owning the room” during a presentation or practicing self-leadership in your daily tasks, but the end goal should be to be your authentic self.

Understanding, integrity and reflection are central to both my work ethos and the way I engage with things personally, and my reflective side often leads me to tackle things on my own.

But something that has become particularly clear to me during this time is that growth relies on a strong support system as well. I want to thank Maria Anker Andersen for building an environment where it felt safe to speak openly among such kind and ambitious people, as well as my manager Thomas Seim Larsen for backing me up the whole way - always being inquisitive about what I was learning and exploring my reflections on the FLA and so many other things.

In the words of Mette Norgaard - and one of the very first things I noted down during my time at the academy: “Today is a great day to grow as a leader.”