AI-live 8 maj

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

Bra video kring Figure (robot): https://youtu.be/ch_UM_JJU9w?si=MGzjPaGL80xI3vKU

Introduktion och teknisk start

Livesändningen började med att välkomna tittare på olika plattformar - LinkedIn, YouTube och klog.se. Det nämndes att det varit två veckor sedan förra livesändningen och att nästa vecka är en reklamedag, så nästa sändning blir om ytterligare två veckor.

OpenAI GPT 5.5 och kodning

Den nya GPT 5.5-modellen från OpenAI har rullats ut sedan förra livesändningen och många har börjat testa den med imponerande resultat. Modellen beskrivs som extremt kraftfull tekniskt sett och briljerar på att bearbeta stora mängder data utan fel. Dock konstaterades att den inte är lika bra på copywriting och att texter inte blir särskilt mänskliga.

När det gäller programmering har GPT 5.5 visat sig vara mycket kompetent. Den kan ta emot instruktioner på naturligt språk och leverera önskade resultat, med funktioner som planeringläge där hela planen går igenom innan implementering. Många programmerare börjar nu överväga att gå från Claude till Codex för kodning.

Codex rekommenderades starkt som verktyg, inte bara för programmering utan för många andra uppgifter. När det kopplas till det lokala filsystemet får det "superkrafter" och blir mycket mer komplett än andra verktyg. Det blir snart mycket enkelt för alla att bygga egna appar mot sin egen data.

Rättstvisten mellan Elon Musk och OpenAI

En stor händelse som diskuterades var den pågående rättstvisten mellan Elon Musk och OpenAI/Sam Altman. Tvisten handlar om huruvida OpenAI fortfarande är en non-profit-organisation och följer sitt ursprungliga uppdrag. Elon Musk donerade stora summor pengar till det som skulle vara en non-profit med ideella mål, och han anser att de har avvikit från detta.

Det finns en möjlighet, även om den är liten, att OpenAI kan gå under om rättegången faller ut på ett visst sätt. Det skulle i så fall innebära att GPT 5.5 och framtida utveckling stannar av. Mest troligt är dock att det blir någon form av kompromiss, men mycket står på spel i denna rättstvist.

OpenAI:s integritetslösningar

OpenAI har varit mycket aktiva i Europa de senaste två veckorna. De har släppt något som kallas "privacy filter" - en relativt liten modell som kan köras lokalt på laptop eller i webbläsaren för att minimera data, maskera över personliga uppgifter och köra lokalt på datorn. Detta är ett försök att möta den europeiska efterfrågan på bättre dataskydd, men även amerikanska användare är oroade över bristande skydd.

Google Workspace och Microsoft Copilot

Google Workspace inkluderar AI-funktioner i basplanen utan massa extra licenser, till skillnad från Microsofts mer komplicerade licensjungel. Google har länge haft AI inbyggt i Google Sheets, Docs och Gmail där man direkt kan generera svar på mail utan att behöva kopiera och klistra in via ChatGPT.

Microsoft Copilot har dock blivit mycket bättre nyligen, särskilt när det gäller att söka och hitta information i SharePoint och OneDrive. De stora teknikföretagen börjar nu använda AI-stödd programmering för att rulla ut nya funktioner mycket snabbare än tidigare.

Bildgenerering i ChatGPT

ChatGPT:s bildgenereringsmodell (Image 2/Images 2) fungerar nu extremt bra. Det som tidigare kunde vara problematiskt, som att skapa serier av bilder för till exempel en kampanjkarusell, fungerar nu felfritt. Man får chockerande bra resultat direkt, ofta med "one shots" där första försöket redan är bra.

AI-smartphone utan appar

OpenAI skissar på en AI-smartphone utan appar, vilket väcker frågor om hur en sådan revolutionerande produkt faktiskt skulle se ut. Efter att vi gått från telefoner med fysiska knappar till iPhone:s touchskärm, är nästa steg en telefon utan appar överhuvudtaget. Idén är att göra det enklare än hur vi använder telefoner idag, men det är oklart om det blir en modul eller mikrofon man bara pratar i.

Apples nya riktning med hårdvara

Apple byter VD till John Ternes som är mer hårdvaruorienterad, vilket kan innebära att Apple fortsätter AI-racet genom god hårdvara och nya produkter snarare än egna AI-modeller. Det finns rykten om AirPods med kameror och andra wearable-enheter som kan vara mer AI-drivna utan lika många skärmar. Apple har redan ingått avtal med Google Gemini och kan välja bästa AI-modellen hela tiden medan de fokuserar på det de är bäst på - hårdvara.

Framtidens interaktion med AI

Det är osäkert hur vi kommer att interactera med AI om ett år. Just nu sitter vi vid datorer och skriver till AI, men framtiden kan innehålla wearables med AI eller små enheter man pratar i. Det är en öppen fråga var informationen ska finnas och hur vi ska få tillgång till den - kanske inte genom dashboards som vi har idag.

Amerikansk AI-politik och reglering

Under den senaste veckan har amerikansk AI-politik äntligen kommit igång. Flera aktörer har börjat göra konkreta uttalanden och åtgärder som troligen kommer att bromsa den snabba utvecklingen, vilket bedöms som något positivt. Kina verkar däremot inte göra samma sak, och kinesiska modeller ligger inte långt efter.

Figure - humanoidrobotföretaget

Figure är ett robotföretag som har blivit mycket mer transparent och låter nu folk besöka deras testanläggningar. Någon har gjort en intervju och rundvandring på hela fabriksgolvet som visar hur noggrant allt byggs och vilka tester de gör. Det som är särskilt fascinerande är priset - om uppgifterna stämmer kostar roboten cirka 300 000-500 000 kronor, ungefär som en ny bil.

För det priset får man en autonom robot med 5 timmars driftstid som kan bära 20 kg och rör sig i 1,2 m/s (långsammare än en människa). Jämfört med till exempel trappstädning i en bostadsförening som kostar 300 000-400 000 kronor per år för fem trapphus, skulle roboten kunna städa trapphusen varje dag istället för en gång i veckan och även byta glödlampor och göra andra uppgifter. Breakeven skulle vara ett halvår.

Figure har varit mycket transparenta och visar att de tränat egna neurala nätverk. All compute sker på roboten själv ("on device"), det är inte beroende av internet utan kan fungera i alla lägen lokalt. De har inte berättat exakt hur tekniken fungerar eftersom de vill behålla konkurrensfördelar, men de har visat att roboten kan göra alla typer av hemuppgifter och fabriksuppgifter och att den tränats för allt direkt.

Företagen pratar om försäljningsstart 2026 eller senare 2027 och de ligger i fas med det, kanske till och med lite före tidsplanen. Det går snabbare än förväntat - nu är det bara maj och massproduktion börjar snart.

Intuicel - AI som lär sig från erfarenhet

Ett företag från Lund som heter Intuicel arbetar med AI-teknologi som lär sig från erfarenhet, till skillnad från dagens AI-modeller som tränas färdigt och sedan släpps. De har en liten robothund som tränar och lär sig själv - med ett mål att ställa sig upp börjar den justera sig själv utifrån sensorer och rörelser. Den ramlar först men lär sig sedan ställa sig upp.

Om denna teknik utvecklas vidare öppnas möjligheter för robotar att få en sorts "personlighet" baserat på sina erfarenheter, till exempel att roboten lär sig att en viss person blev arg och anpassar sig därefter. Detta är både spännande och läskigt.

Robotdesign och mänsklig upplevelse

En reflektion var varför humanoidrobotar fortfarande ser så läskiga ut. Figure har en svart visir och ser ut som en "läskig mördarrobot". One X har däremot satsat mer på en gulligare design med två ögon istället för svart visir och verkar mer fokuserade på den psykologiska upplevelsen.

Det finns en utmaning med "uncanny valley" - människor får en konstig känsla när något ser mänskligt ut men inte är det. Figure har otroligt mjuka rörelser, nästan lite robotaktiga men också nästan mänskliga. I slutändan kommer vi hamna i en formfaktor som vi förväntar oss ska vara med oss. One X:s primära målgrupp är försäljning till privatpersoner, medan Figure är mer inriktade på produktionsgolv och kontor.

Avslutning

Livesändningen avslutades efter 40 minuter med en påminnelse om att det blir två veckors uppehåll. Tittare uppmuntrades att lämna kommentarer och att hitta inspelningen, alla länkar, transkriptioner och tidigare avsnitt på qlok.se i kunskapsbanken.

I know.

Why can't I hear anything? I can't hear anything. Now I can! Yes! -Hey.

No, no.

Welcome to the AI Live broadcast. Two weeks ago, it was a regular Friday. Next week it's a reclame day. And what I said was that we are live on a lot of different places. Follow us on LinkedIn so you can write comments and questions there.

So we try to see them.

And if you are on YouTube, you can write there too. And it also goes via klok.se to find this live broadcast if you want to watch there. So I think we are live everywhere. It was just me who didn't manage to mute myself twice. So I forgot to unmute myself.

Yes, in addition, if you just want to listen and turn off the phone and do something else, I will post the link in the description.

Yeah.

Super.

in the comments below.

Uh... Yes, it has happened.

It was two weeks ago that we talked about it. It has not been news-dry, it is never on the AI front. We try to publish these news letters every Wednesday. There are two of them on Klog.se with a little bit of what has happened. I think that if we start a little further back in this, then maybe we should also mention that the new chat GPT model has rolled out, or OpenAI GPT 5.5. It happened last time we sent it, but I have seen that many have started to test it and are very impressed with the level that it is at.

I It's a really powerful 5.5 model. It's worthless on copywriting though. It's not good to write text with. It's not so human, but it's a fucking technician who just burns through enormous amounts of data and doesn't do anything wrong, which I can find at least so far. Um...

My spontaneous reflection is that I have tried to code with it.

Claude, and it has been this big model that everyone is asking to program with. Uh... But no-You can write to it as a human and you get what you want. They... It can be a plan mode, so you go through the whole plan before you implement it and so on. I think it's a very good user interface and it does a great job.

Hmm.

Hmm. Hmm.

And I've also seen that those who are programmers start to feel like this: It might be worth going to Codex instead of Code.

I've heard a lot about that, from all the sources I follow.

And it can change anytime, but right now it's very, very good.

Hmm.

Another big talk that we haven't had for so long is the right between Elon Musk and Sam Altman. It's between Elon Musk and OpenAI, but Sam Oldman has to represent that. It's about whether they are still non-profit and have their mission. It has kicked off and is worth following. It is exciting in so many ways, from how non-profit teams work in the US and so on, but within the AI area,It's quite exciting to see.

If it falls out in one direction, OpenAI can go under.

Oh.

So, and then there is no... GPT 5.5, or somewhere this will fall further.

Hmm.

I think it's a bit of a...

It's not so incredible that it falls out that way, but there is still a possibility that this time it will end up in Elon Musk being right, and then there is no open AI left. Worst case.

I don't know how true that is, but there have been some screenshots of messages they have sent to each other.

Huh?

Should we share them? Can't we both drop everything?

Like Elon Musk saying "You're going down" or something like that. Yeah. It's boring if it's only about That they are bitter with each other and just try to shut each other down. Sure, Elon Musk also had a great model in Grok, but I mean... It would be a shame if chat GPT would go under. But then you can think what you want about it. But from the perspective that we will get as good AI as possible, It's worth it that everyone is still here.

Oh.

Absolutely.

The exciting part is that there are many parts in this atmosphere. But what's left is that Elon Musk donated a lot of money to a non-profit that had an ideal mission and he thinks they have deviated from it.

Hmm.

And that's what they're trying to decide. And if they have deviated a lot from it, they can be both responsible for repayment and a lot of other things can happen. I've listened to some American lawyers who have reviewed this. American law is also quite different from Swedish. It's a pity that the court in such cases doesn't make such very hard decisions in one way or another. Because it becomes so very hard to prioritize everything else.

The most likely thing is that it slides out in some kind of middle-of-the-line thing and someone has to do something and someone has to do something else. It's not very incredible, but it's a pretty big right and it's pretty much at stake, you could say.

Hello.

So it's a micro-new that we also have with OpenAI. OpenAI has been very... Very much in Europe lately. It was Entropic two months ago, but now it's been Open AI for two weeks.

So it changes a little who you talk about most. So I think we'll soon get Google to turn it around. Open Source had something called...

Okay.

What was it called? I can't remember the name.

privacy filter, I think they call it, which basically means that you can run a fairly small, even locally, on your laptop or in your web browser, or something like that.

minimize data, mask over personal requests, run time locally on the computer, etc.

um Thank you.

...an American company, and you ask if they're trying to appeal to... You. Yes, please.

to get a little more protection. It's a bit exciting.

It was a micro release, but still a good use case of AI, I think, if people hear it themselves and get it.

Most likely.

But in the American private debate, everyone is very concerned about the lack of protection. So it's also a matter for people to have their own protection in some way. So I can appeal to both of them.

-So it seems exciting, but I haven't had time to work with it.

Hmm?

I don't have anything new to say, but just this, I'm not going to answer this, but If you have a Google Workspace ingår de här sakerna? Det är inte en massa licenser och annat som är i Copilot va?

Yeah.

I think the base package is important.

You can still buy Gemini, but I haven't talked about the new features in the payment plan. I think it's quite free.

And it's nice because you don't have to worry about the whole complicated jungle that Microsoft has set up.

I think Google is really running on this more in-built Hmm. Hmm.

And it's been around for quite a while that it lives in Google Sheets, it lives in... What are they called now? Their counterpart to Word. Or in the mail for that part, that you can directly generate a reply to an email and ask it and then send it instead of having to copy and paste it in chat GPT and so on. So that's nice. I think it's been a long time since I've used it. When you're looking for one drive, they've been pretty bad so far.

It works really well now. I have noticed in the co-pilot how much better it has actually become. If you sit in the co-pilot chat and ask about things that are somewhere in a SharePoint or something like that, you will often find that things are much better. They have done something to make it more effective. Um, So a lot has happened in Both Microsoft Copilot and GoogleThank you.

I think we are a little early in this era, now the big tech companies are starting to work with much more AI support programming, so they are also rolling out features much faster.

Uh.

So the speed.

Oh.

And that's my thought, if I have a concrete tip for everyone, it's not as cool as it is, because we have this image generation model and we could share a lot of lives on it, and it's very easy to just go into ChatGPT and use their image model. I've also used it a lot in the past weeks because it's incredibly good.

You get shocked every time. It's like one shots on everything. It's always good at first.

But I've started using code more for things that aren't programming. And for those who are in chat, it can be worth loading code. Because I think thatI'm not sure. It's not only the chat GPT, but it also gets superpowers when it's connected to the local file system.

Oh, Thank you.

um you It's more complete than the others.

So if you haven't downloaded Codex, do it. It looks like a chat GPT. It's not a big deal to start there.

Mm-hmm.

You have Cloud Code, but you need something else. umI don't know.

Hmm.

um-umyou build your own boundary with it and get what you want.

I Very impressive.

We've talked about this before, that everyone will be able to build their own apps against the pure data. Codex is well close to just working out of the box. It's not far until it's super easy.

But it also follows us a little bit into one of the points that was in our main news. What it was about, let's see here. OpenAI sketches on AI smartphones without apps. And that is interesting, first and foremost, that in my head I see How does this phone look like?

It doesn't have any apps.

We've been told that the phone has physical buttons that you press and then we get the first iPhone where there was nothing and everyone was like wow, completely new idea. No one had thought of this before, except Steve Jobs. And now we come to a phone that... should not have the apps. So, is it a dictaphone?

It's interesting how this revolutionary new product, in that case, actually comes.

Hmm.

look like and The idea is that it should do The way we used the phones before, hopefully it will make it easier. The question is if it's just a module, a grunt that you go around and hold in and talk about, but just a few things. I haven't read much about it, but I think it's very interesting howand how we interact with them.

Mm.

Mm. Mm.

Hmm. Hmm.

Yes, we talked about it two weeks ago. Apple is changing its CEO. The CEO, John Ternes, Uh... AI is also a more hardware person, and they might continue to run the AI race through good hardware and new products. There is a lot of rumour that they might get even more involved. AirPods is perhaps talked about with a camera in it, and that it becomes wearable devices that may not have as many screens but are more AI driven.

Even if they don't have AI themselves, they've ended a deal with, for example, Gemini and Google.

They can pick the best model all the time and just build really good hardware.

That's what they're best at. So maybe the new CEO who goes that way and that it's like...

And one of the things is that Amen.

-that will make the focus more on new ways of interacting with technology-which might not be as easy. Smartphones. So there are many indications.

Now with the COVID-19 situation, we are in a web-based environment and we write to AI. We don't know how we will interact with AI in the next year. It is important to keep up with the development, but it may not look like it does today. It may be wearable with AI, or it may be just a little grunt. Uh, I sometimes wonder what this information, because it is pure information we want to include in our digital systems, Where does it live and how do we access it?

Is there a world in which we can have a dashboard to look at? How to present the information. Now it's very far ahead, but I don't think it's true that they put a lot of energy, time and money on that.

Mm.

Mm, mm.

Hmm.

A horror-struck enjoyment in the future in some way.

Yeah, yeah. There are two other big themes, I think, maybe even in the last week.

Wow.

This week, two things have come to light. The American politics has finally come to life. It may be this right-wing, it is some kind of mythological model. A lot of things have happened where you are now like this. Okay The American superpowers have also made some kind of statement about, I don't know if you can remember, but there are many actors who have started-to become concrete things.

Oh.

Yeah.

That will probably stop the fast-moving, and that's probably good.

China might not do the same, because the Chinese models are a bit of a hack. It's not far between, so maybe more Chinese models... I I think you can show a video of it, it's better that people go in and watch it themselves.

Hello.

We might add something there too, but there is a company called Figure that we usually24 per day, but we're going around the world. But in all this, more people have come there and visited and done some interviews and stuff. I watched one of those the other day, yesterday, the day before, maybe. It's pretty good. If you can look up the link, we can post it in a comment. We can put it on the website together with the show notes.

Because then it's someone who goes around and checks the entire factory floor and interviews. It's incredibly fascinating what tests and things they do with these and how carefully everything is built. ---I haven't confirmed it yet, because it's not on their website.

Hmm.

Oh, I put that out.

I was thinking to myself, You put in a simple sentence: "If you could afford to buy a new car, 500,000 SEK would cost you a car.

It sounds expensive, but it's so damn cheap for an autonomous robot that has the worst operating time. It's just crazy.

It's on the sales list that I can find.

Hmm.

Mm-hmm.

It can carry 20 kg and it has a runtime of 5 hours.

You would have rather bought a robot for 300,000 SEK, which I can just ask you to go down to the store and buy."Oh.

It has a speed of 1.2 m/s, so it's slower than a human. It's only 5 hours autonomous and when it charges a bit it starts again.

It's easy to buy a home, but as I said in previous lectures, for the housing association, which I've been the chairman of for several years, Trapphusstäd kostar 300-400 000 om året för fem trapphus.

Oh.

Den här kan ju städa trapphusen varenda dag, inte en gång i veckan. Och så kan du byta glödlampor. Den kan göra så mycket saker. Så det kommer ju bara bli... Vi är ju breakeven på ett halvår. om vi kan köpa det här för 300 000-400 000.

But, but... I haven't seen that video, I'll have to check it out myself.

Hmm. My assessment is that they have been so transparent and shown so much and now they let people in their test facilities and you can see them doing things there.

Uh,I'm gonna talk to you later.

But how well it can actually do these things, or if it is a little CS points as well, for now.

Yes, I think they have really broken it. They have trained their own neural networks. So I'm how you stand here like an input and control Hampton Prata become at end of the desk he features the other for talking about it all things are on device a robot and for the phone to be the latest upon the internet can you ask a kind of funka in Allah lagin some internet or so sorry like me to a follow-up So all compute here even hair also some of you can roll the network all horde vara satin shower Locals and for a number of copper and for a kind of your a flier soccer man, thenAnd you can't see that in any video I've made, but they claim that they've broken it too.

Hmm.

I see a few people. Hmm.

I myself come from a background in automation and I think it is very importantSo nothing with AI, but what humans have coded, what is available with sensors and other things.

So, somewhere... Hmm? Classic robots, simply put. Hmm. Hmm. Hmm.

And I have such a hard time seeing how that...

There must be such extreme... I will definitely check out this video.

Hmm.

Speaking of robots, I usually show a They are not transparent, more than they have shown. But it is a company from from Lund, which is doing It's actually an AI technology behind it, they've put it in a small robot dog. And what they mean is that this AI model learns from experience. The difference from today is that when we train these AI models, they are trained in the end system and then they are released into the community and then they are run as pre-trained.

But then this Intuicel, which they are called, train and learn from themselves. They have this dog, they obviously have a goal to set up. Meanwhile, the sensor started to adjust itself and how it should move in order to be able to stand up. It fell at first and then it stood up. And if they are now, not transparent, but if they are now a bit on that track I see an opportunity for AI, and it opens up the way for a robot to get a "Persolleet" I talked to this person and they got angry and then they got a personality.

Yeah.

Very exciting and quite scary. It happens a lot within I'm fascinated by the robotics.

Yeah.

All right.

Hmm.

-...

-Mm-hmm.

Nästa vecka har jag AI-kurser igen. Den här veckan har jag inte fångat upp det innan jag höll den i tisdags. It's hypothetical within a year.

Mike.

Hmm.

And all these companies also talk about selling start 2026 or later 2027. And they're on track with that. I would say that they're maybe a little before what I thought. Because I thought they were at the end of 2026 or something like that.

Oh.

But now it's just May. And then they start to come up in mass production. It goes faster than you think.

Yes, and of course they will be expensive in the beginning, but like everything else...

Yeah.

Um, Okay.

Yeah.

They are both very fast and they have dared to be very transparent during the journey. Not with everything, they don't tell exactly how their technique works because they still want toand can do all types of homework and factory tasks and they have trained him for everything directly.

I was thinking about this iRobot with Will Smith, where the robots just turn red and shoot at the entire humanity.

When it comes out, it's as a general person in their work.

Then there will be specialist professions and small motor things that they may not be able to do. Hmm? Yeah.

Oh, well.

Yeah. Yeah. Yes, and that's also weird.

Why do they still look so fucking scary? If I were to buy one of these, it would probably look like something else than a scary murdering robot. Because now they do. We should make them a little cuter, at least. And the other thing, I can just show it, and then I'm not going to make a video of it, but One X has at least made a little cuter robot. I didn't follow them that much. Hmm. Um Yeah. Um,Hi.

As long as we know it's a robot, we have a little...

We talked about this, that people get a strange feeling that something is human, and it's not.

figures It's incredibly smooth.

It has a little robot movement, but it's also a little... And this one has two eyes, at least.

Okay.

It doesn't have a black visor. And somewhere... In the end, we end up in a form factor that we also assume is with us. One X seems to be more focused on that, how the psychological experience is.

Whereas Figurine is more like, "Do you want to go to a production floor or go around to an office?" That doesn't really matter. Their primary goal group is to sell to someone's home. So. Yeah. Jan?

um I just wanted to throw in The last time it was the same, but it was the last time.

I remember that it was a crux of getting I mean...

Hmm?

several image generators. But I have continued on this and it actually works really well in ChatGPT.

Yes, that's it. Hmm.

with Image Gen 2 or Image 2 or whatever it's called So if you ask it for a series of pictures, like "I would like to have a carousel over this campaign on..."-Oh.

I don't believe it.

It was a bug in the beginning, I didn't get it to work the first few days, but it's good to say that. Hmm. Yeah. -spotlessly today.

Oh.

That'll be it.

So, bill your nearing shot, KPT. In so many new use cases, you can do so much now.

The text was good, but I noticed that it was a bit weird.

Hmm. Hmm. What is it called now?

umOh.

Images 2, I think it's called. But in chat, the government is just outstanding.

Okay. Very good.

No.

We've been talking for a long time, as usual. Only 40 minutes.

Yes. Yes. Yes. Yes. We're going a little down on the territory.

Better than 50. -Ja, då kommer vi inte gå live.

--foreignSomeone was looking inside, but wrote...

Nej, det är bra. No, we probably won't. It will probably be two weeks of keeping it up now. But feel free to leave comments and even find something. And then you can find this as a recording and all the links we have talked about and transcriptions and things on klo.se, in our knowledge bank. So feel free to go there if you want to look at earlier episodes as well, because on LinkedIn it is really hard to find back to the old links.

So that's why we put everything on klo.se now.

Super! Thank you for today. Have a nice weekend!