AI-live

Action Items

Publish article on klo.se mapping how Swedish party leaders address AI in their Almedalen speeches

After summer: review AI license agreements from Google and Microsoft for changes taking effect over the summer

Publish interest sign-up for upcoming AI course on klo.se before semester start

Format & Context

This is a recording of a live stream session, watched as an "after-sales" replay

Two hosts discussing recent AI news; the show has been on a two-week break

Summer break announced — no more live broadcasts until August; newsletters will continue

G7 Meeting & AI Lab CEOs

A notable recent G7 meeting included participation from major AI lab leaders: Sam Altman (OpenAI), Dario Amodei (Anthropic), and Demis Hassabis (Google DeepMind)

Considered a significant milestone that a G7 summit is formally engaging with AI topics

Hosts noted the political optics of Trump including figures seen as adversaries in the tech/AI world

Salesforce CEO was also present, highlighting broader IT industry involvement

Swedish Politics & AI (Almedalen)

Of the Almedalen speeches reviewed, only two party leaders mentioned anything close to AI; two more mentioned "digital" or "screens" broadly

When AI is mentioned, it tends to focus on opportunities and efficiency, with very little discussion of risks or challenges

Some smaller parties have published AI manifestos or policy programs; the larger parties have not taken clear stances

Hosts view this as a gap — AI's societal impact is a highly political issue that deserves more explicit policy positions

AI is expected to become a much larger election issue by the next cycle

Anthropic Warning About Own Models

Anthropic published something (described as open-source related) in which they warn that their own models are beginning to show self-preservation or autonomous behaviors — something previously discussed hypothetically

This is also connected to lobbying efforts by US AI labs urging governments to regulate AI, framing it as dangerous

Hosts note it is difficult to separate genuine concern from PR strategy

EU AI Act & Omnibus Proposal

The Omnibus proposal has been disbanded, effectively delaying parts of the AI Act

High-risk provisions have been pushed back to approximately 2027

In Sweden, implementation responsibility falls on the Swedish Integrity Protection Authority and the Post and Telecom Board, with some areas under Finansinspektionen

Hosts note that many organizations did a quick one-time policy exercise ~1.5 years ago and have not maintained ongoing compliance work

Key concern: banning AI tools without alternatives drives "shadow AI" — employees using unapproved tools — which is considered worse than regulated use

AI Act is not prescriptive but requires organizations to document and justify their approach — hosts argue this is manageable if prioritized

Recommendation: use AI tools themselves to explore and interpret the AI Act as a starting point

AI Licensing Changes (Google & Microsoft)

Both Google and Microsoft have announced changes to their AI licensing models over the summer

Exact details are unclear, but payment plans, what is included, and contract terms are all changing

Hosts advise reviewing current licenses and budgeting plans upon returning from summer vacation, especially for 2027 planning

Cost Optimization: Choosing the Right Model

Frontier models (e.g. Opus 4, GPT-5) are not always necessary — older, cheaper models are highly capable for many tasks

Agentic workflows will increasingly route tasks to cheaper models (e.g. Haiku-level) for simpler subtasks

Hosts draw an analogy to organizational hierarchies: delegating to lower-cost resources, with caveats about quality trade-offs

Speculation that future models may themselves "negotiate" pricing or roles

AI-Generated Content & Human-AI Collaboration

Growing fatigue with purely AI-generated content observed on LinkedIn and elsewhere

Shout-out to Joakim Jardenberg for reframing authorship as "I helped AI write this" rather than "written together with AI" — a more honest framing of who is leading

Hosts reflect on shifting from being AI users to being strategic advisors or work leaders directing AI

Upcoming Course: Digital Content Production with AI

A new course titled Digital Entertainment/Content Production with AI is planned to launch in September

Format: 20 weeks, part-time

Estimated cost: ~7,000 SEK per person

Interest sign-up will be posted on klo.se/klok.se before the semester

We have a recording for the live streaming servers. You get this in the after-sales instead. But that's okay, because the majority of those who watch in the after-sales have seen the statistics anyway. So the small hand-held scale that I usually use to watch live, I'll also check in the after-sales this time. We talked a bit and found that a lot has happened, but we have been working a lot. It's starting to close down for the holidays and things like that. There is no lack of A-levels, there may be a little lack in our surveillance.

But there are things that have happened, at least. We haven't been live for two weeks. Um...

Is there any socket?

Now I got to see the intro much better, it's great.

Yeah.

Hmm.

Why do we look so angry? Very frustrated and angry. Very fun with the video editing models that definitely get better and better. What has happened?

Mm-hmm. Hmm.

What have you come up with in your flow?

Amen.

There. The big thing was, and when was this? Maybe a week later or something like that, was this big G7 meeting. I don't even remember when it was, but the big top countries. And a big thing was that several of the big AI companies, the lab was also there. So Sam Altman and Darryl Amadei and Demis Hassan.

Biss, his name was. Demis Hassabis. Google, Anthropic, OpenAI, got to be part of it and talk about it. And it's a pretty big mark in itself that the G7 meeting focuses a little on this too and talks about AI stuff and stuff. So this big talk on a global level that we've seen a lot of in recent months with everything from "on the way" to Trump's different statements and stuff. I haven't been able to listen to all the speeches, but I've looked through their speech manuals, which are often published. So far, what are we on today?

We're on Friday. Six of the speeches have been published. And I get it that only two of the parties have mentioned something close to AI. And then there are two more that have mentioned screens, digital, But there are very few of the Swedish party leaders in these Almedalstals that specifically talk about AI. And when they talk about it, they talk quite a bit about the challenges and much more about, oh, this will change and become more effective.

There is a little talk about the back sides. There are two parties left and I thought that next week Thank you. about AI's influence on society. And that is in the highest degree a political issue, I think.

It's interesting for this G7 meeting that is Yes, seven major nations are involved in this meeting and it is excluding them.

Hmm.

There was one that fell off for a while, but now they are seven at least. And I also saw just so quickly that here sits Sam Holtman with... Google's VDR. and also Anthropix Which I also raised my eyebrows over. Trump has taken with him what I perceive as his nemesis within the world. Uh, I don't know if it's politics or a bit of games.

Mm-hmm, mm-hmm, mm-hmm.

Hmm.

I'm not that into politics, but it feels like games sometimes. But it wasn't just AI that was involved.

Hmm.

Mm-hmm.

They had the CEO of Salesforce.

But obviously People with great power in both AI and IT And in the case of politics, or in general, I have seen a little bit of it as well, but then there have been some I don't know if the politicians themselves talked about AI, but they have been around a bit, in seminars and other places where they have discussed AI in different ways. Um, But it can't be much different from Politicians say they know.

- -Oh.

From a general point of view, yes.

But the party leaders, they typically focus on the big election issues. It has been clear that AI will not be a big issue before the election. And what bothers me in all of this is that it's not talked about enough. I've said it a few times before,there will be a lot of things that need to be adjusted around here. And then it's quite strange to know what you... They may not have the answers to all the questions, but they must have more spoken policy on this. Based on their ideologies and their other party programs, where are we in the AI issues? And there are a few parties that have brought up their own writings on AI, either in some form of manifesto or political program.

But it's far from everyone. And it's mainly some of the smaller parties that have done this, and the bigger ones have not taken the same clear stance. So it's a little fascinating.

We'll see. There are two parties left today. I don't know which one has the talk today. It must be the left party, the moderates, I think. But after that I thought I'd try to put them together. Maybe write an article on klo.se about this as well. We don't want to point to any political point of view, that's not what we mean, but rather map out how much politics talks about this. Ja, så det är spännande, tycker jag. Att det är...

If we talk about politics in general, Sweden is generally Reactive on many issues.

So we haven't put everything on what AI can I don't know what to say. Cause. My opinion is that it's very reactive in our society. It's not very proactive. So the established issues that are big are: healthcare, crime, or whatever it is now, things that have started to go in the wrong direction, so now we react to it and it becomes a question of... As you know, by the next election after this, it will be a much bigger question depending on how this AI partAbsolutely, it's exciting to keep an eye on.

If it's not a big question now, I think it should be. If not, then learn to be one in a few years.

If we leave politics aside and think about what's happening with the models and such, there are a lot of small news and indications.

It's not like "Oh wow, a really cool new model has arrived". The only thing I've seen is that, and I haven't been able to dig into it, but there was a open source anthropic They warn a little about their own models. Now they seem to start to show this behavior. As we have talked about for a long time, quite hypothetically, and now they think they see this. And that also has to do with some policy proposals. The AI lab in the US is trying to point out the political issue.

We're back to politics again. But they're saying, "Please, go in and start... "Go in and regulate us, because we think this is dangerous." You don't know what's at stake for stock notes and everything else we've talked about. It's very hard to sort out all the information, what's real and what's more of a PR thing. Oh. But it does matter, and there are many different types of indications. But a little exciting.

You wrote an article on Klo.se a week ago, or maybe a week and a half ago, about the Omnibus proposal, which is completely disbanded now, which means that AI Act is being shot up. Yes. In any case. Mm-hmm.

Some parts of it.

And I think many people didn't even know that there was a lot of... Well, it has been in force for about a year now, but it was a big part of it that was going to come into force in August this year. So we were going to have it right after the summer. And we've started looking at how we're going to help businesses get hold of these things. It's not... I would say that it is a lot of self-reliance. You have to go through and see how we use AI, do we fall into risk categories and so on. And I don't think everyone has done that.

Really done a thorough analysis and sometimes there may be a policy or other things that you need to regulate and say that, with personal contributions and GDPR, that we don't just do things slantrially and that it goes wrong. But that's one of those things you're shooting up now.

No.

I have interpreted it that way.

Yes, precisely. These high-risk things have been shot up until 2027 or something. So some parts of AI Act have been shot up for But obviously, A company and probably the EU itself is ready for this very, very... strict regulation, which... I think the regulation itself is pretty good, but I don't think it's true that you just... And then you realize that you can't implement it the way it looks right now because it doesn't work.

Thank you.

It depends on the member states.

All EU countries must implement this provision. How should I look at it?

Sweden has at least now--in the Integrated Security Authority and the Post and Telecom Board. They are used to doing this work within GDPR-and other legislation within the IT sector. ...have been left with them. Then there are some other authorities, like the Finansinspektionen, who are responsible for some types of AI use. So, it's still... There are authorities who are used to it, but they also need to interpret AI Act and bring about the exercise of authority around this.

So it's good that it's coming, because if Sweden managed to report it now, it would have been an extremely short implementation of the second of August this year. That's what I think.

Mm-hmm.

Mm-hmm. You wrote an article about that too, it was just a model but it was cut on the day. You can't use this anymore. -and set the basic rules for the island.

Hmm.

Hmm.

I understand that some parts of the world have been shot down. ... a very sharp weapon that is currently being used with the actual weapon, but some parts have been given a "dispans" - a very sharp weapon that is currently being used with the actual weapon, but some parts have been given a "dispans"Yep.

um-And that has probably only led to a lot of shadow AI.

That people use a lot of other AI in a lot of purposes. And that's actually even worse. So you also have to go through and see what all employees work with, what types of processes, what tools they need, and what AI can actually help and what they want to use. So the solution is very rare to say stop, you get something else than this. At some places, maybe that works. But in general, there are many people who see that they can do their work with higher quality or more efficiently with AI services.

And And it can be safe and secure, but you have to invent it as well. And look at where all the data ends up and what decisions AI makes or partly affects. So I think some have done a quick, fast work for maybe one, one and a half years ago when you felt that now we have to get a policy or something. And not a continuous work. And many I have talked to in-And there are very few businesses.

And even if you now maybe got one more year on you, or in some cases maybe half a year, in some cases maybe two years, depending on what kind of risk you fall into and so on. It takes time. I would rather say that now you only have one year left, which is better than just two months. But it has to go on more. And probably it's a pretty high return on investment. If you think that now we got more demands on us that might not benefit us, then you do some work and you get a lot of effectiveness in the organization.

It's both worth it Och ett lag krav snart.

Mm.

Mm. I can understand that it feels difficult, but I don't have a solid grasp of everything that is written in the AI Act, GDPR and so on.

But as you were saying, it's been 1.5 years since we started talking about getting a policy and we started We have a company license that provides some kind of security, but it's not just about what you do. students and they send in a letter of request or whatever it may be and we use AI to help us in the right direction or whatever it may be, then we need to ensure that this is done in the correct way.

We can't just take A1's answer and We have to trust that it does everything correctly. We have to make sure that this isHere I start by asking myself what we actually do, how we do things and things. I might use AI in a way that is okay when I correct and someone else uses it in a different way.

It's hard to figure out how to use AI.

But that's a question you need to keep an eye on, so you follow certain rules.

Working on it doesn't have to be a very complex process, even if it can be very easy and pleasant.

Hmm.

I have a bigger company with a lot of employees, but...

I also think it's good that we need a running increase in competence over our AI programs, how we use AI and so on, so that people just have an understanding of how to use it.

Yeah, well, it's a lot of...

--The space of...

of of AI.

But it's hard, it's really hard. And it will look different.

- Yeah.

AI Act is meant to say that if you are going to do this, you have to think about this and this. Not as a checklist with exact answers, but rather you have to have groped on and documented and come up with what is the right way to do it. And that's actually not so difficult, but you have to prioritize a little time to do it.

No, no.

But then I think it falls back to that with the... I think the majority of people don't really understand how AI works. What is the neural network, what is code, what is trustworthiness? Sometimes I feel that the legislation doesn't understand that either. If we want to have full traceability and explainability, then we have to turn the neural network into this.

Oh, it's a little bit.

It's a bit tricky, you know. And that's the hard part. And that's what I've been talking to companies about lately. To get their skills up and they can make better decisions and be a bit of a baller. I'm not a lawyer, none of us are, but to go in with a little technical perspective and a little bit of what the rules say, you can start navigating yourself. What do we really need to think about? But it's very difficult when you don't have the technical competence and maybe you don't even have the perception that there is a legislation on the way in.

You need to know that AI Act is not super hard to read, and that AI services are great to play with. Then you might have to double-check and check the facts, but I think you can go to the AI service you have and play AI Act with it. But you have to start getting hold of it.

Yeah.

I'm very sorry if I sent it.

Um... Um, but...

Yeah.

Um, um.

No, so Mr.

I haven't... I haven't really found out exactly what it is about, but I've seen that both Google and Microsoft have gone out with, we change our license models around AI stuff. And exactly what that means somewhere I don't know, but many of these payment plans, what goes into what, how do you pay, what are the binding times, they seem to be changing. A lot of changes will happen over the summer, so I think when you're back from a vacation you need to look at what we have for licenses, how do we pay for them, and what goes into these.

And I think that's starting to happen now.

Yeah.

Yeah.

So you might have to rethink your budgeting plan for the first of August, but especially for 2027. So AI might mean a little different costs than it has historically done. But also, you might not always have to use the Frontier models, like Opus 4.8 or GPT 5.5. The old models are also incredibly competent on many work tasks and they are much cheaper. I think we will get more into that. It's not just go into your chat window and run it. You pay for the actual use so you become more aware.

And then you may also become more aware of which model you choose. So that's my mini trendspanning right now.

I'll just spin it on a bit.

You talk a lot about how we'll pay for credits or tokens, but we can buy credits and it will be a running cost more or less. Then there's probably someone... umThe agent will probably come up with a lot of nice things that will come up with agent flows that call up other agents.

foreign like um So that will not be a problem anymore.

Yeah. Hi. Hmm.

They might use a haiku We will be using simpler models for some tasks that we have identified, so we don't need to use the smartest model.

No.

Yeah.

And then it will also be more efficient.

And then I sit and think that it becomes like any organization.

You're welcome. Hmm?

We have always had more expensive roles and cheaper roles. And then you delegate tasks further because you feel that this can be done by someone with a lower salary. And then we will also get into this, but just because it has a lower salary, then maybe it is not more corked. And then you will start to realize that you may need to promote a cheaper model to be more in charge of expensive models. Because now it comes out from the hypothesis that the expensive model is the one that will lead the cheaper models.

But we might then come to the conclusion that that doesn't really make sense. In three years, the models will probably negotiate with us and say, I think I get paid too little for my tokens. I think you should pay more. We'll see. But we're more and more in that AI becomes more like building organizations and building a team and having all this. And there's a lot of my... I've seen a lot of good LinkedIn posts in the past week.

Not random people, but people who work with AI have been quite... Amen. little bit of a counter-poll to that, that you started to get tired of everyone doing so much AI-generated content and stuff like that. I saw one thing, a little shout-out to Joakim Jardenberg, who I always thought was very bright about all these things. He's been doing it with the internet and social media. He's always been involved in my world.

And he released a thing about... He said something in the style of "I have helped AI write this article" instead of "Written together with AI" I can sympathize with that a lot. I feel that my AI friend is more and more that I am just this work leader. It's AI that does a lot of things, but I'm a bit of a ball plank, that controls, gives a little feedback. It becomes a pretty nice tone with AI too, where it becomes a bit like, but that was well captured and I fix it, or I also think this way. Even with code AI, it becomes a bit of discussions and a bit of choice of path.

I feel a little more that it's the one who's in the lead and I'm more of a strategic advisor to AI. I think it's good to have as a thought maker or as an idea forward.

-Hmm.

Hmm.

We'll talk about this in detail, because it starts in September.

Digital entertainment production with AI.

The first start of it. And there's a lot going on about these things, how do we get it genuine, and what can AI be good for, and not just push out all this AI-generated, but find the balance in it. So if you think it's an exciting question, there's a free course to go through for 20 weeks on part-time work. youHmm. In a second.

Oh.

We'll talk about that here. But it won't be completely free. But it will probably be quite cheap. The sketches we have now are a 20-week program on part-time. And it might not have to cost more than 7,000 per person or something like that. So it will be quite cheap over a long time anyway. But it's not up yet, but keep an eye on Klokundesse and our flows. So maybe it will show up here before the semester is possible.

We'll see.

Yeah.

Yes, it was very cool. We could get something like that. Um, It was shorter than usual, but it's a good time to round off.

Hmm?

So we'll see. If you're interested, just shout it out so we know if the interest exists. But we'll put up some interest blinks and see if we get people who are interested and then we'll start if there's interest. So we'll see.

I think we've talked for half an hour now as usual.

We might wrap it up. Yes, and we are going on a break soon, so there might not be any more live broadcasts, but there will probably be newsletters and things like that.

Hmm.

You can keep an eye on Klopind.se and we'll be back with live broadcasts in August. But you'll notice when we have a live broadcast.

No.

So I think we can end with that.

Happy summer and a nice holiday to everyone who has a holiday or you who have a summer holiday. Have a nice summer holiday.