AI-live

Lyssna på bara ljudet här: https://media.trinorn.ai/player/

Inledning

Detta var en veckovis fredagsdiskussion om aktuella utvecklingar inom AI-industrin, affärsmodeller och framtida trender.

AI-industrins tillväxt och ekonomisk bubbla

Diskussion om huruvida AI-sektorn befinner sig i en ekonomisk bubbla eller inte

De flesta indikationer pekar på att det inte är en bubbla, även om inte alla företag har hållbara affärsmodeller

Många AI-företag är fortfarande högvärderade trots att de förlorar pengar

AI har utvecklats från att vara en mjukvaruteknologi till att bli en hel industri med behov av datacenter och infrastruktur

Google planerar att starta datacenter någonstans i Sverige

Förändrade prismodeller och affärsmodeller

Tropic och andra företag har börjat erbjuda betalning per token istället för fasta prenumerationer

Detta innebär att användare betalar för hur mycket AI-modellen "resonerar" och tänker

Tidigare antagande var att priset var det enda som spelade roll, vilket ledde till okritisk användning utan hänsyn till effektivitet

Jämförelse med Spotify: istället för att betala 120 kr per månad skulle man kunna betala per låt man lyssnar på

Nuvarande modell med 1000 kr per månad ger nästan obegränsad compute-kapacitet

1000-2000 kr per månad ger betydligt mer värde än 300 kr-nivån, som börjar kännas begränsad

AI-tjänster och prenumerationer

Det finns många olika AI-prenumerationer tillgängliga: Google, Claude, Codex, OpenAI, ChatGPT, Gemini och andra

Möjlighet att köra modeller lokalt i sandboxade miljöer

Det blir allt enklare att arbeta med lokala modeller

Prenumerationer kommer troligen fortsätta existera

För AI workflow agents och mer avancerad användning kommer betalning per token bli vanligare eftersom de resonerar längre

Microsoft Copilot och strategiska förändringar

Microsoft Copilot börjar bli mer intressant med tre olika modeller att välja mellan

De använder sina egna modeller, GPT-modeller från OpenAI (fram till eventuell AGI), och Cloud-modeller

Strategin att erbjuda flera modeller kan skapa mer stabilitet

Copilot har tidigare känts rörig och mindre värd för pengarna

Microsoft verkar ha avslutat sitt samarbete med Anthropic/Claude

"Frontier fine tuning" - Microsoft tränar och finjusterar äldre, billigare modeller baserat på företagsdata inom Microsoft 365

Modellerna finjusteras automatiskt i bakgrunden baserat på vad användare gör och skriver

Hårdvaruutveckling och lokala modeller

Nvidia har tillsammans med Spark byggt ny hårdvara för datorer

De nya chipsen från Nvidia gör det möjligt att köra lokala modeller effektivt

Flera stora PC-tillverkare har byggt datorer tillsammans med Nvidias chips

Ny Microsoft Surface-tablet använder denna teknologi

Batteriprestanda är en viktig faktor för att köra modeller lokalt

AI-reglering och politik

Diskussion om reglering och politisk debatt kring AI

Det finns polarisering mellan de som vill ha mycket reglering och skydd mot risker, och de som inte vill accelerera

Saknas nyanser i debatten - fler borde bygga förståelse för dessa frågor

Önskan om att politiker skulle prata mer om AI-frågor

Gap mellan vad experter skriver och vad som diskuteras politiskt

Etik och moral

Diskussion om ett dokument från Vatikanen om etiska och moraliska aspekter av AI

Effektiv användning av AI

Vikten av att kunna delegera till AI på samma sätt som man delegerar till mänskliga kollegor

Om det är svårt att prata med AI, reflektera över hur man pratar med mänskliga kollegor eller medarbetare

Tekniska färdigheter blir mindre viktiga än förmågan att delegera effektivt

Welcome! It's Friday again. Weeks go by quickly. There's no shortage of AI skills. It's a small overpowering level. It continued all summer. It did last summer. There were a lot of things during the holidays. --um So a little crazy. So is it an economic bubble or not? But most of it points to that it is not. Then maybe not all of them are in the sustainable business model, but theyThere is also a discussion about how they earn money.

Many of these companies bleed money, but they are still highly valued. And what you have been able to see a little, I think it was Tropic that started... Now you pay per token. So much that you let our AI model resonate and think, you pay for it. And that can be let... Both positive and negative, I think. But many have just had the assumption that it's just the price that matters, so now I can just keep going and put as much as possible in use. And I don't have to think so much about is this worth it, is it effective and so on.

And it will need to resonate more around now. If we just translate this step they have taken. Let's say we have an appointment on Spotify and we pay 120, I don't know exactly how much, but that's monthly costs for the different ones. Instead, you would go over to a model that you pay per song you listen to. That could be a good comparison.

um um um um There are two variants of it. But it's that one thousand-lapp per month, you get almost unlimited compute. It's not unlimited by definition, but it's hard to consume. You have to work really hard to get it out. And that's still pretty cheap.

Yes, it is. But then we also have these Heteropro+ everywhere. All these subscriptions, Heter, that you go into Google.

Appana Claude code codex open shop capitero and the gravity of the Gemini App and Alexa is a Ganska Mickey some pecan water that I love my fat on your mat bigger hope to matter and super up Just that you can I job I'm on the local a feeler program for or the in dot or something so that clean arena shut gransknipe a bit of a can't grab it um umYou can live and work locally with them and then you can sandbox them in certain areas and they will be able to enter in such a way.

It gets easier and easier. So I also think, as you say, they can move over there more and more, but then they will need to have A very simple average.

and clearly divided what it is.

I think that the subscriptions will live on, but should you have a proper computer and all this...

I don't want to call them agents, I call them AI workflow agents, but I will start working with them. We will see more and more, you will have to pay per token, because they resonate for a long time.

Hmm? Yeah.

The real value for that 1000-billion or 2000-billion a month is so incredibly much bigger than what you get for 300-billion. I'm starting to feel that 300-billion is very limited. There's a lot I can do with it. It's a good idea, but it's a lot of manual work and I have to do things. It's more like a question mark. --It's really exciting.

We talked about the valuation of the company and it's happening a lot. It suddenly requires a data center. Google wanted to start a data center somewhere in Sweden now. There are several. But AI is not just a form of technology and software, it becomes almost a whole industry of it. Which is interesting in many aspects. I think both what it does for humanity and we talked a little loose about this The document comes from the Vatican, the ethical and moral world and so on.

How do you know that this is not just a small thing anymore, but a whole industry around AI?

-I think that it might be a little---Um, But also that they have their own models, plus that they are using GPT models that they get from OpenAI, until they reach AGI, if they ever get to transfer out of AGI. And then we also have the cloud models. So now it's all of a sudden, the Copilot Microsoft Suite is starting to get exciting, because they have three different models that you can choose from, they're trying to build more stability.

It might be a good strategy, because it's been really messy in Copilot, and you feel that it's just a little less value for money there. It might be the opposite. It might be the most valuable for money.

umI'm not going to swear on it, but I'm not going to say that they are kicking off Anthropix. They are not going to use cloud anymore, so they have screwed up that idea.

-But it's a bit of a twist, because they've been laying around for a lot longer. Their models have been interesting, but not at all close to being usable in that way.

It's a bit interesting how we talk about this, because we always ask: "Which AI services have you tested?" And there are a lot of questions. It's quite tricky to write, because we write AI services, and then we can think about what an AI service is, what an AI model is, etc.

umBusiness um The third thing I have in my head, which I think is a bit... I'm not going to blackmail Microsoft for something I haven't thought about, but what I think they seem to have said is that they had a name for this, "frontier fine tuning", something like that. Frontier Tuning And that's really about taking a pretty good model, not the latest one, but a slightly older, cheaper model, and then training it and fine-tuning it based on the company's data within Microsoft 365.

So what people do, write, it gets fine-tuned by itself in the background. umOkay.

foreignumThe new Nvidia, it's also called Spark, right? It's always called Spark, now. Which is so confusing. But it's some kind of... They have built hardware that now runs on computers, which makes it possible to run local models. They don't build their own laptops, but they do it with a new Microsoft Surface tablet. I'm so bad at the PC side that I'm a Mac user. But a lot of the bigger PC manufacturers have built computers together with chips from Nvidia, which actually makes you But I'm telling you to make them come up with the same level of so past bra compute that's a fact this can show Modeller local to ordently men develop a snack is a cream way. Yeah, I'm killing a chief for Brooklyn being on battery to the American. Yeah, they're certainly to call a modeler at the maximum robot.

They did it. So that's a battery that's your Obama. I thought he had a computer to some of soccer set them.

I'm really spitting on them-So it's not just that I've gotten an answer to something, but that it actually does something.

I usually think that you can delegate...

----um um I umumforeign You start by answering one question per dimension and then do a rough map and say that you seem to be standing here in this dimension of A. And then you can answer four questions in each dimension if you want to fine tune it into something. But it's a big challenge to take-foreignWhat I'm also fascinated by, and I'm not going to point to anything political, but I'll pick up the most general election map that exists.

It's a pretty early cluster in it, where those who want a lot of regulation and protection against risks and such, they don't want to accelerate either. --It's two opposite poles. I think we're missing some nuances in this debate. That's why more people could look at this and build their understanding and start doing it to the election question.

It's my speculation, but it could be that even those who are in thewe have those who like to feel that "oh, new things, the market regulates this" and so on. So it might just be a little bit of a Yes, and the question of knowledge. Our politicians talk about what they can.

Yes, and I feel like I wish I had read more about the social sciences and economics. There are a lot of things I don't know about. That's why it's fun to find more about it. Yeah umThere are gap analyses where you look at the differences between experts and what they write about.

And it's talked about quite a bit politically.

And several of those areas are now in design books. um Interesting and very fun initiative, I think.

Both you and I would like to have the question about AI, that people talk about AI. We take, people should lift this and it might be a little more political question about how do we hang with it and how do we regulate it and so on. So all thousands of followers who are watching this, please come.

The last 40 minutes have been a long bubble. --I'm talking about a good point there, which I also started talking more and more about.

In testing the tools, there may be some technical things, but it becomes more and more how good we are at delegating. Think for yourself if you feel that it is difficult to talk to the AI or your digital colleague. How do you talk to your human colleague at work? Or are you a deputy manager? How do you talk to your employees? How do you get things done? Translate that thinking to the AI now.

um And you answered it, I thought it was great.

I understand what he's saying, but if we do this, it works. Yes, exactly.

Now we've talked for a long time. 47 minutes today. Not really, but almost. Thank you for today, and have a nice weekend.