AI-live
Google Gemini Video Generation
Google released Gemini Omni video generation capabilities approximately 1.5 weeks ago
The rapid pace of AI model releases makes it challenging to track all names, models, and version numbers
Team Resources and Newsletter
Weekly AI newsletter published on the website's knowledge bank, typically posted on Wednesdays (sometimes Thursday)
Knowledge bank accessible through multiple paths on the website, including an AI chatbot for finding information
Content created with AI editor assistance, followed by human review before publication
Anthropic's Rise and Vatican AI Encyclical
Anthropic valued at approximately $900 billion, potentially surpassing OpenAI's valuation
Despite previous challenges with US government relations, Anthropic has emerged as a frontrunner both morally and financially
Vatican published AI Encyclical on Monday, co-written with an Anthropic founder
Encyclical marks the 135th anniversary of an earlier document from the Industrial Revolution era addressing labor concerns
Document emphasizes maintaining human value and not becoming blind to automation
Expected to influence legislation, subsidies, and policy decisions as a foundational ethical document for states with Christian traditions
AI company leaders like Sam Altman and Elon Musk shifting rhetoric around employment impacts
China's AI Development
China introducing travel restrictions for key AI development talents
Chinese companies developing competitive models with significant economic resources
China producing graphics cards comparable to older Nvidia models without external assistance
Meta Employee Monitoring Controversy
Meta monitored 8,000 employees' work patterns and behavior to train AI models
Raised concerns about ownership of personal work experience and whether it should be protected
EU likely to address these issues, unlike US and China
Some employees subsequently laid off, particularly those in programming roles
AI Agents and Self-Reflection Research
Google research published mid-May on AI agents playing Pokemon games
New approach allows AI to observe actions, reflect, and rewrite its own system instructions
AI can update both memory bank and system instructions, programming itself at a deeper level
Works well with sufficiently capable models; smaller models can get stuck in worse behavior patterns
Raises philosophical questions about AI self-awareness and consciousness
AI creates specialized sub-agents and neural networks for recurring tasks
Virtual World Experiments
Different AI models tested in virtual world scenarios over two years
Anthropic developed democratic systems and performed well
Grok model resulted in crimes and total failure within 7 days
Google models experienced fraud issues
ChatGPT remained passive
Practical AI Tools and Applications
Different AI models have distinct strengths and "personalities"
GPT-4 excels at data analysis and calculations
Gemini 2.5 Flash particularly effective for parsing PDF files and data analysis
Tools like Agentic (code assistance) and cloud-based coding solutions expanding capabilities beyond simple web chat interfaces
Recommendation to assign larger, more complex tasks to AI rather than overly simplified pieces
Need to move beyond web chat limitations to unlock full AI model potential
Additional Resources
Practical knowledge and courses available at klok.se
Live again. Google Omni.
Google Gemini OmnibusHere's the.
Okay. It's starting to get hard with all the names, all the models, all the version numbers. But the video you can generate inside Google in Gemini, quite simply. som är ganska bra kom ju för en och en halv vecka sedan.
No. We have these newsletters coming out, if you want to keep yourself updated.
Men det har inte varit nyhetstolkar den här veckan heller.
It's going fast. Hmm.
Hmm.
And there we have...
I think I can show you some pictures so everyone can see a little.
I have prepared for that. On our website you can go to the knowledge bank. There is a button on the start page, you can go via menu, you can go to a lot of places. You can also talk to our AI on the website and find things that way. If you go to the knowledge bank, you will find our entire knowledge bank. Or if you just want to check the AI, you can go to the AI knowledge bank. Then it is ready, filtered and clear.
It's written every Wednesday, but sometimes we post it on Thursday.
Yeah.
Hmm.
We have an AI editor who helps with that. Then we have human security before we publish them, so that we are also in agreement with them.
And that is still very important.
Hmm.
And I recently read a news that I think we can discuss a little later, regarding these agents that do things autonomously and so on. It's very interesting. The first news is that Althropic is being super valued, somewhere around 900 billion dollars. It is believed that they, as I understand it, are passing by OpenAI's evaluation.
This is interesting, considering that Anthropic somehow tried to be exposed by the US government and so on. Like a bird phoenix, they just became super strong. --Oh.
Hmm.
Hmm. Both top news are connected.
foreign Tropic just disappeared.
The second top news is that AI Encyclical was published on Monday. I think it was published officially on Monday, but it was written together with one of the Entropic founders. They have both brought the Vatican and the Pope with them and have become the most valued company. They are frontrunners right now, morally and in terms of value.
Thank you. Hmm?
Hmm.
They are not someone who is chasing anymore.
They are almost first in the lead in some way. It has been only a few months.
Hmm. Hmm. Hmm. Hmm.
I don't know, have you read this? Yeah, it might be tough.
My knee.
I'm aUm, uh, I don't know how to say it. Magnificent.
through all that. Mm-hmm.
Magnificent. Magnificent.
I have a little difficulty interpreting parts of it, because it is very referential to the Bible and to the Christian tradition and things like that. I interpret it as alarming in that it is...
It highlights that this is a big difference for people and that we have to keep the human value. And that's what we've talked about in AI World and we've talked about a lot earlier. We have to see the human and the human value in things and not see ourselves blind to automation. This was published on the 135th anniversary of an earlier one that came around the Industrial Revolution, when people were also very worried about how the Industrial Revolution would affect human labor and replace a lot.
The first thing that came to mind when I was caught up yesterday was that there have been several different interviews with these leaders for AI companies. Sam Altman and Elon Musk have started to change their rhetoric. Both of them are on their way to becoming IPOs. They might change it a bit because of that. But now they have changed their rhetoric. Everyone will become unemployed and we have to do these things a little more.
No, but that's not so bad. If you are keen to read it, it is exciting.
Okay.
Thank you.
My macro interpretation of it is that many states in the world are rooted in Christian traditions or Christian values and politics is leaning so much in that direction, which will affect this. This will become a underlying ethical document for a lot of other politics and such.
It will have an impact, even if it is not the most associated with AI in general.
It gives some kind of basis for many other decision-makers to stand on. And that's why it will probably be reflected in legislation and subsidies and other things that will happen in the long run. It's a very bearing document.
I think the song is interesting with the human side of it. We can look at all the cool stuff that is created with AI and so on, but somewhere it's still... Well, I guess the starting point is... We humans are supposed to live with this, hopefully, at least more than it takes over things.
Hmm.
And then what the church often comes up with is all the things that are of great function from the church, I think. It would be interesting to get some expert input on this.
Mm-hmm.
You just read it.
Uh, umYeah.
You need a deeper theological basis to see what the value really is. Hmm. Hmm.
Hmm. Hmm. The common ethics have a lot to do with that.
Oh. I don't think it's a document or a statement to throw away in any way.
Okay.
Hmm.
It's not a document. So, from... Mm-hmm.
Hmm. Yeah.
And a little double bearing, as you also have spoken a little with Entropic, or with clear connections there and not to anyone else. And that is also a statement in that it is perhaps the most moral AI company and I think that affects a lot. Among other things, the valuation, I think it affects and other things like that. So it's exciting. Anthropic is in the forefront, maybe not on models, because there have been some new models and some Chinese models and so on that might now outperform everything else. It's so hard now to say who is best, because the benchmark is so spread and there are different things they measure and so on.
But it has been a bit of a novelty this week, even better models and so on.
I think we should jump a little further down, because you mentioned one of the news.
It says: "China limits the AIT-talents' export." Uh, that they introduce travel restrictions for talents or people who are highly important for AI development in China and they are a bit divided.
Hmm?
I'll bend now. Yeah.
China is hard to keep track of because I haven't been to China, but the end is different and it's a bit controlling. They have the opportunity to work a little bit and then suddenly Boom! Now we can use Nvidia or these different companies Um, I'm in there. Interesting what happens. Behind the curtain and so on, just all over China. And as you say, there are a lot of good models that are Super good with and big companies develop with huge economic resources.
Did you see lately thatThey can bring out graphics cards on a level of...
Hmm.
It's an old card for us, but it's still impressive that they have been able to do this without help.
Hmm.
Hmm. Hmm.
So, definitely up and going.
Hmm.
Okay. You I'm going to talk about an article that you published.
---And it's a new cold war in AI-cap-wearing in some way. And it's starting to get a little... I'm a bit too bad at history, but I think of protests around what we call war, nuclear weapons, or extension of nuclear power and all that, where in the end people's pressure is so great that they start backing up, announcing, finding downforce and such. We might get there, it wouldn't be completely stupid if we got some form of...
public attention even more on this so that companies and politicians find a proper pace in it. Because we are running crazy fast right now. Or we, AI development is running crazy fast. It's not that we do it, but humanity as a whole, as the great we, is running very, very fast. Mm-hmm.
umBut before we talk about that, we have something interesting with Meta, which has hired its employees to train AI.
Hmm? Hmm?
Hmm? Hmm?
Hmm. Mm-hmm.
Hmm?
Hmm.
Hmm? Oh.
If Anthropic is at the top of the moral scale, then Meta has managed to lower itself even lower than they might have been. Sometimes you get impressed by how they managed to get into controversy after controversy. Besides the fact that we have traced that all people have started the earthworm and collected all the data and then happened to leak everything. So, so, -and started automatic monitoring of all employees' behavior-in order to train AI models.
um foreign So 8000 people have worked effectively, found good flows, managed to get things done, benefited from different ways, just to measure, train their AI models and then be able to kick them.
Hmm.
Hmm. Hmm.
So you have worked a little to get your... I don't know.
Hmm?
I started to record people's work and then train models. It's a bit like music or pictures or other things, but this is something that even more people will relate to. Because my usual patterns and habits, if I were to have a conversation about it, I would use it in training. In the US, where there are a lot of political contacts, they will probably get a group vote and everyone will get a few thousand dollars. Facebook will settle for a few hundred billion dollars and that doesn't matter.
That's how the American system works. What do people have to do with ownership over their own collective experience? Is it worth something? And should it be protected by the copyright? Or maybe not the copyright, but something else. "Erfarenhetsskyddet" Hmm.
-EU might be good at this, they will probably already be divided on this.
It wouldn't be with USA and China, I don't think so. And this is a group of people, I don't know exactly who has been kicked out, but I can probably think that it is something with that, because VSCode and others program in many ways and they have been affected by AI.
Hmm.
It's hard to explain. you know, a programming arm. -and it might stay here for a while.
Mm. Mm. Mm.
The question is, will it be able to develop into more areas?
Yeah.
Hmm.
Yeah. Mm.
a larger number of people.
Mm. The article, yes.
We can just point out that there is a lot of that in our knowledge bank as well. I have tried to write one a week and a little disclaimer, it is a bit in the text as well, but partly my opinions, then it is also co-writed with I try to go through a lot of different Both interviews and in this case a research paper to just hang out with. And then when you're reading them, you get some reflections. I usually throw in that plus my reflections and what I think.
And then you get to do the throw out and then we work together again with these friends. -There are many more, but the last one is Okay. I'm trying to find a summary of facts we know and thoughts about what it might point us to. Um,Oh.
----But if you want to tell us a little bit about what these researchers actually did, it was in Pokemon.
I'm going to say this.
It's not like Pokemon Go that most people know about, but older Pokemon like Game Boy and stuff. They're both pretty easy to emulate in the data, so it's easy to give A1 control over it. But also because it's things you do early in the game that you do things late. You have to think about what resources you have and save on stuff. It's a bit of a tricky passage, you have to figure out what to do and stuff.
And it's a game that someone who knows the whole game can roll on I don't know the exact time, but say a few hundred hours, but someone who doesn't know what to do can stay up for a thousand hours. So it's not linear, the game itself. And therefore, AI has to list out and do things. And then you have the latest version here, where you publish this research paper in the middle of May, somewhere on May 13th, it says here. So you have the song in Gemini that you have played, Google is part of this research, so it's their models.
But then you have also added some new layers of that it can make it a little reflection.
observe their own actions, reflect and then write their own instructions. Instead of having a locked system prompt and updating a memory, it can both update its memory bank and its own system instructions. It can program itself on a deeper level than what it has been allowed to do before.
And that seems to work really well. As long as it's a sufficiently smart model. If it's a smaller smart model, then that self-reflection can instead lead to it being stuck in even worse behavior patterns and becoming worse. And that's where my personal reflection is. Isn't it a bit like that for humans too? Our ability to reflect, think strategically a little deeper, to maybe also have a little ball plank around us and such. We can have both positive and negative spirals in our own improvement or change.
umIs this the level of AI that we call self-awareness?
Hmm?
Yeah.
When do we start to think that this is also a life... It raises existential and philosophical questions. Plus it's so cool, because now we will have AI that can just work on its own.
Analyze this. -then I don't know what it can't do.
I think it's a very interesting experiment.
um Mm-hmm.
My feeling is that we have thought about an AI, a neural network, umSome would. And I understand a little what they've done here, because they have sub-That's all.
Mm. Mm. Mm.
Mm. Mm, mm.
As a function, it blocks these sub-agents.
you So it's a very interesting area to explore here now.
How can we get an agent, get input from other agents, to strategically be smarter with experience then, really.
So, very interesting.
um It's impressive that it has come to the point where it can.
Previously, it was said that it was a lot of colleagues, but this research points a little more to the fact that it builds up its own micro-agents and neural networks that are specialized in things that we often need. In this case, it was a Pokémon ride or things like that that returned to the game. It created sub-agents for those who were specialized in those parts, and then it just delegated the over-sufficiency there.
Mm.
Mm. Mm.
I think it's really interesting. Then you can also ask the question, do we have enough good neural networks to actually work properly over everything? Then the neural networks will just get better and better. At the moment it will be very, very good. An interesting experiment that a sub-agent can create their own tools. I can have an agent who just checks bugs or problems and sends them to me.
Um, I thought I would show you that now, because it's a bit of the same thing.
This was a successful experiment, at least as far as I can read. But I heard another news where they had tested the same thing. In a virtual world they would have a Tropic with different agents that would interact with each other and take like... live together. So during two years they were in this virtual world and interacted with each other and did things.
Hmm?
Hmm?
Hmm.
And Anthropic did very well.
They came up with "democracy is good, we set up some democratic systems" and so on. So they came up with good things. The other agenda, or AI model that they tested in a similar way was Grok. I think it ended with... But if it was like "after 4 days there would have been a lot of crimes and on the 7th day everyone was dead" or something like that.
Mm-hmm.
So they do some experiments with these The different AI models, Google had their share of the problem, there was a lot of fraud. And so on. So Antwerp was again those who said "we are creating a democratic world, we can live together" and chatgpt was the one who was really passive, they just slept all day.
Okay.
Yeah.
The.
I want to get out of these articles, which is not at all difficult for, for example, GPT. The gems are perhaps mixed, but the gem is quite flat too, and it draws a little strange conclusions. On the other hand, I love GPT 5.5 for a lot of data analysis stuff and things like that. There it is completely fantastic when I have to plow through a number of amounts and do calculations and stuff. So when it gets an application, it can be very...
I think that the word "tykande" comes with a lot of different meanings. You can discuss it in a way that is challenging for me, and then I can develop my way of thinking. You find different preferences. It's like different colleagues who are in different ways.
I like that type of use better too.
It's nice if I can I appreciate more when I can myself resonate with things, with ideas, with someone who is very knowledgeable, simply because I have developed more personally.
Hmm.
Yeah.
It's interesting that they still, we can't call it personality, but it is quite subjective. I always point out that in the courses I take, I try different tools. What I think suits me best may not suit me best. It becomes very clear in the image generation models, but in the text it is also the same.
um You can work a lot with...
...personality in these. It's quite reasonable because they're trained in neural networks. They've also invented patterns and paths and gotten different priorities during their training. So they'll come in different ways. But to send a concrete takeaway, something I've been working on a lot during the week... I've mentioned this before, but this is a very specific area, but we work a lot with vocational school education and statistics. This week was a division for short IH. And then again, I wanted to do a little deeper data analysis. So two years ago I started with this, I think down the entire Republic of the Authority theory.
There are others who have done this, there have been a lot of initiatives about open authority data and things like that.
It's good for me to go through them, but one thing I noticed now is that I think it's good for Codex as well.
It does a lot of anti-gravity, which is Google's code and local tools to get files and stuff. And there's a new Gemini 3.5 Flash. 3.5 Pro hasn't rolled out yet with 3.5 Flash. It was completely crazy good at that. It has nice improvement ideas. It built a Python script that started parsing through a lot of PDF files and stuff. See you later. -But yeah, so that's...
Oh, okay.
um Thank you.
Hmm. Hmm? Hmm? Hmm?
And it is really good.
Yeah.
I also noticed that this might have been better. I found a bug after you asked me. I fixed it by the way.
Okay, nice. Really good.
Hmm.
Hmm.
But I also want to make a point that There is a lot to read about, if you want to have good agents, you might have an MD file with instructions and other types of stuff. But the MCPs are like... like new ways -Well.
Hmm.
um-But it's much better to just go in and say that. I have to be able to say if this is enough clarity in the statistics, because it's a trend or not. Please count on that. And then it does. And therefore we must dare to take a step, but still have general knowledge so we can observe if it does the right thing or not.
But I think you push A a little.
I think the general trend right now is that we give too easy tasks for small pieces and you have to dare to lift it to even bigger things. You have to dare to get into tools. And we've talked a little bit about that. We don't have time in our practical courses to go into a lot of vibe coding pieces. But it's no longer vibe coding. It's all forms of local document handling, controlling your system resources.
It can do so many things to it if you just let it in. And you may not do that If you don't have any idea, but I think you can get some idea and do it there. Pretty easy. And no matter if it's anti-gravity or cloud code or cloud co-work or if it's anti-gravity. There are so many solutions from all of these, but you have to get a little bit into those tools as well. We're a little bit stuck in that fucking web chat mode that we're not competent for.
Oh God.
They don't bloom. These great AI models can't work in this closed environment.
No.
No, it's not exactly... It's quite complicated. Yes, the clock is ticking. -Hi.
Ja, klockan är ett. Vi behöver avrunda.
and the practical knowledge we can offer at klok.se.
I was on the phone.
And please write comments, listen to us, give us your opinions, input, everything else. We weren't that dystopian today, I think. We were more nuanced. A clap on the shoulders to ourselves for not painting up the undercurrent today. Oh, oh, oh.
-Bye.
I'm in the... Trevlig helg, tack för idag. Yeah.
Stop.