AI-live 24 april

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

ChatGPT 5.5 lanseras

OpenAI tillkännagav ChatGPT 5.5 kvällen innan livesändningen. Modellen hade tidigare arbetsnamnet "Spud" men döptes om till 5.5 inför lanseringen. Namngivningen från OpenAI beskrivs som konstig, eftersom 5.5 faktiskt är en ganska stor uppgradering jämfört med 5.4, snarare än en mindre uppdatering. Utrullningen sker gradvis och alla användare hade inte fått tillgång till modellen vid tidpunkten för livesändningen. En av deltagarna hade fått tillgång till 5.5 på morgonen, medan andra fortfarande bara hade version 5.4.

Modellen beskrivs som att den är bättre på att förstå sitt uppdrag och ställa klargörande frågor, ungefär som en kollega snarare än ett verktyg. Detta gör att användare kan presenteras med olika alternativ att välja mellan, vilket är särskilt användbart när man inte riktigt vet exakt vad man vill ha. Det finns en forskningsartikel från OpenAI som beskriver förbättringar och hur modellen fungerar. Överlag bedöms 5.5 inte vara deras absolut bästa modell, utan snarare en något mindre finjusterad version eller en mellanversion av något som kommer snart. Lanseringen tolkas delvis som ett sätt att visa att OpenAI fortfarande konkurrerar med Anthropic.

OpenAI Images 2 - ny bildmodell

OpenAI släppte också sin nya bildmodell Images 2 samma kväll. Bildmodellen beskrivs som "crazy" och deltagarna är imponerade av hur smidig den är. En av deltagarna hade gjort alla bilder i Google i sex månader eftersom ChatGPT varit så dålig, men är nu tillbaka till ChatGPT igen tack vare den nya modellen. Images 2 verkar också vara tillgänglig i begränsad utsträckning för gratisanvändare, men med dagliga eller månatliga begränsningar.

Modellen är mycket mer exakt än tidigare versioner. Den kan skapa korrekta QR-koder och streckkoder som faktiskt fungerar att scanna. I testerna dök det bara upp ett stavfel på ett ställe, annars skrev den text korrekt. Den nya modellen kan också kombinera thinking och research med bildgenerering, vilket tidigare inte var möjligt - man var tvungen att först låta modellen söka information och sedan be den skapa en bild.

Förbättrad förståelse och intuition

En stor förbättring med den nya bildmodellen är att den har mycket bättre intuitiv förståelse för vad användaren vill ha. Tidigare behövde man vara väldigt tydlig och preciserad i sina instruktioner. Nu kan modellen förstå kontext och syfte med mycket kortare prompter. Ett exempel som visades var att skapa en bild med två symmetriska rutor för videoströmmar i en livesändning - modellen förstod direkt vad som menades baserat på en ganska kort beskrivning.

Modellen kan också ta ett screenshot från en webbplats och förstå den grafiska stilen, för att sedan översätta det till ett helt annat sammanhang utan att behöva detaljerade brandriktlinjer eller långa beskrivningar. Under livesändningen gjordes en demonstration där modellen fick skapa en poster om AI i IKEA:s grafiska profil - den hittade själv material på internet och lyckades fånga rätt färger, logotyp och känsla. Resultatet blev ett övertygande IKEA-liknande bildmaterial med texten "Automatiskt inspirerad, planerad, skapad. Teknik som gör livet enklare, steg för steg."

Upplösning och format

Bilderna genereras i 2K-upplösning när man kör via chatten, och 4K-upplösning när man kör via API:et (då betalar man för användningen). 4K-upplösning räcker för mycket tryck i A4-format och fungerar troligtvis även för roll-ups och affischer.

OpenAI har också lovat att det ska gå att generera flera bilder i en serie för att bygga en berättelse över flera bilder, till exempel fem slides i följd. Vid tidpunkten för livesändningen fungerade dock inte denna funktion helt - modellen visade att den genererade flera bilder men producerade bara en.

Konsekvenser för arbetsflöden

Diskussionen berörde hur dessa förbättrade bildmodeller kan påverka hur vi arbetar med visuellt material. Tidigare har det funnits ett stort behov av redigerbara dokument - organisationer har PowerPoint- och Word-mallar som folk kan redigera. Men när AI blir helt säker på att skapa bilder kan det förändra dessa arbetsflöden.

Det blir också enklare att ge AI mindre detaljerade instruktioner och istället låta den komma med egna lösningar, för att sedan ge feedback. AI kan ofta generera bättre lösningar än vad man själv tänkt sig, så det kan vara fördelaktigt att ge mindre instruktioner och istället vara öppen för alternativa perspektiv. Om man ger för många instruktioner får man exakt vad man vill ha, men det kanske inte är det bästa resultatet.

Konkurrens mellan AI-företagen

Utvecklingen visar att det pågår intensiv konkurrens mellan AI-företagen. Google ligger inte långt efter med sin Notebook LM som också har bra presentationsgenerering och bildmodell. Det förväntas att Google inom en till två veckor, max en månad, rullar ut sin nästa generation (kanske Gemini 2.5, 3 eller Pro 2) som troligtvis kommer vara lika bra som ChatGPT:s nya modeller.

Även Anthropic investerar i liknande funktioner med sin Cloud Cowork, och försöker locka över användare från Copilot genom att bli bättre på presentationer och liknande. ChatGPT:s konkurrens riktar sig inte bara mot nischade bildverktyg som Midjourney, utan de vill erbjuda en komplett lösning för olika användningsområden.

Öppen källkod och lokala modeller

Utöver de stora kommersiella modellerna diskuterades även utvecklingen inom open source. Gemma 4 nämndes specifikt som ett imponerande exempel på vad en kvantiserad modell som är lokalt installerad kan åstadkomma, trots att den inte är lika stor som 5.5 och andra jättemodeller. Det är fortfarande lite tekniskt att sätta upp sådana system själv, men de mindre lokala modellerna som faktiskt kan köras på vanliga datorer börjar bli mycket bra.

Avslutning

Livesändningen pågick i cirka 45 minuter. Nästa vecka skulle det inte bli någon livesändning på grund av första maj, men om två veckor planerades nästa sändning med fler exempel på AI-genererade bilder och presentationer.

foreignumlive um umOh, um...

Let's take a look at what happened in the AI world. Our website also has a lot of AI support programming. We have also launched a separate service for live surveys.

First of all for ourselves, but we thought we could also put it out so that others could try it too.

But maybe we'll have to talk about that ourselves at some later time. It is quite AI driven. We'll take it next week, so it doesn't just have to be self-promotion today.

It popped up yesterday in my feed, it was announced yesterday, chatgpt's 5.5 spud. If I'm not mistaken.

It was called spud before, but now it's only called 5.5. The naming from OpenAI is...

It's weird. 5.5 is not just a small upgrade to 5.4, it's a pretty big upgrade. That's why they named it 5.5 and not 6.0. They just changed the name. But now it's called 5.5. It was called Spud before it was launched as some kind of nickname.

And it's interesting that you said yourself that I have access to it now since this morning and I don't have it, so it doesn't seem to have rolled out to everyone. I'm in and watching now on Mellan ChatGPT and I have 5.4.

um -and it's important for the model to understand what it needs to do-and to ask questions. We saw that with Clod Opus 4.7 last week. and then it's smart and acts like a colleague instead. I like it.

It's a great feature. In some cases, we may not really know what we want to have, and if we get presented with different alternatives, and we can choose a way from there, I think it's great that they build in and build in the alternatives and choices that we can then make. Mm.

Exactly. They released 5.5 last night. I watched some of them, they released their own clip and then I watched some AI experts who had livestreams and tried to get it right away. And they hadn't gotten it right away either. So it's rolling out. If you run a chat GPT and don't see 5.5 yet, it will probably come in the near future.

I have gotten it anyway.

I saw that it was here in the morning. I didn't have it last night when I was about to test it. But they also released their image model Tuesday evening. They released Images 2, which is a crazy image model. It will probably be released to everyone right now. I'm totally amazed by how smooth it is. Now I'm doing all the images in ChatGPT again, because I've done it all in Google for six months because ChatGPT was so bad. So I'm back in ChatGPT again.

1.5 Is it also available for people who don't pay for Plus or Pro?

I understood that they would let those who run for free run a little bit with it, but there is a limit, a daily or monthly limit. I can double check a little parallel. --They are doing a slow roll-out, which they haven't done before. We can have a look at some of them. It's worth knowing that they post quite directly. They post a news article or something about these things. OpenAI's web site, which I have in the picture now,But in some areas, we have made huge steps up.

So. But overall, this is probably not their absolute best model, but a slightly less tuned version, or not a finished fine-tuned version of something that will come soon as well. But they have put out a little in-between version, probably to show that they are in the same... So that not everyone thinks that Anthropic is much better than us, then they put out their model. But there is a long research article you can read if you want to go through and see both what they have experienced that it is better at, how it works and a little like that.

So that's a strong tip. Yeah, agency.

I haven't thought it was worth paying more for chatGPT, but now that they have jumped into the cap, maybe you can upgrade there too. In the end, you sit with such ultra plans everywhere, I think, because you want to test all functions. But even in thinking here, it should be very, very good. And what they have also released is their new image model. When you choose to create images here, it runs the new model as well.

I think I can show you a little bit. I have labbed with the image model and I have labbed a little more with it. Create some pictures-umBut the other thing I've tested is that it can make QR codes and dash codes and they also become correct. It can blip off the dash codes, it can blip off the QR codes. It makes all these small details correct and also writes the text correctly. Of the things I've tested, I've got one typo at one place where it actually made a letter mistake.

But it's incredibly accurate now. From a first profile picture to a dramatic one, to a different picture, to a different version. It works. You feel that it understands what I mean. It does things, it thinks for a while and it has the ability to go out and research and find news and do such things. It can also absorb updated content when it generates. Previously, you had to do a little of that first and then ask it to take a picture. Because when you asked it to take a picture, it could not go out and search.

But now it can combine thinking and research with taking pictures, which gives much better results.

It's a very good improvement from previous models they've had. Nanobanana is great too, but now they've really competed. It looks like I haven't tested it that much. Just this, that you can affect the image as you say, but... I was a bit too busy with it too. Just mark certain things that have changed, for example.

--... um-you They didn't understand what kind of frame I wanted, they had put in different sizes, but this one understood the purpose of the image. Then I pushed it a little longer, "Can't they climb a little on the frame? Can't I get some?" So I tested different variants for screen division and it has been elaborated around a bit. But just that it understood what I meant on a fairly short prompt. That's what I'm most impressed with. Because what I wrote was, I sent up the picture and said, in this style I want a picture in 16.9, but with two symmetrical screens that I can put video stream for people in a live broadcast.

And then it took that context and understood, okay, this is the picture I'm going to do. And that's not good. This is one thing I've tried in all picture models before, and it's been really complicated to get it to understand. And now it does.

It's a very interesting subject, because we've talked about it before, but... Previously, you might have to be very clear and write very precisely what you want to have, and in some ways, and sometimes it could solve a picture or a code or a text or whatever it was. But we've seen a... I'm anWhen you and I talk to each other, we know a bit about each other. We know what we have on our knowledge levels. We understand each other's ways.

-Social media. --And then it's like, "Oh, but change dimensions and do these things." And then it's just like, cut off the picture or put a bunch of white on the sides or like, it hasn't really understood the mission. But now it does.

you It might not be for everyone, but many would have appreciated this way of thinking. Even with our tonality, it's graphically shown. Um. Hmm.

Interesting.

The interesting thing is that it can do so much with so little information. Previously, you had to have some brand guidelines, send up things, describe a lot. Now, it's like a screen dump on a website. And then just understand that style. It has a more intuitive understanding of how to translate it into a different context. That's really cool. I had a workshop yesterday, and we talked about PowerPoints and creating visual material.

One of the things-that we think will affect this. We've felt a need for editable documents. An organisation has a... -has a PowerPoint and Word template, and people can edit it. If the AI starts to be completely safe when making picturesNow there is a lot of energy to optimise that a bit more. It's not very practical if everyone is generating pictures all the time. But I have been working a lot with the presentation generation of Notebook LM.

And there you can change this slide a bit, change this slide a bit, give feedback. And that picture model is also very good. Google is not far behind. It has also worked well. But when this generation comes, it will take one or two weeks, at most a month. Then Google has rolled out their almost nano-banana 2.5 or 3 or Pro 2 or whatever you want to call it. And that will probably be as good as this one.

I don't think ChatGPT is on some magical thing that the others don't almost have either. But when all models start to get good, then it's a bit like, if it just has style instructions, examples to go on, and someone just explains, I need to tell this and this, or inform this and this, then just make it a poster, a social media post, a presentation, it just packs it, so that the person can be the one who formulates the goal.

And then we're back to the course of the work, to give the context, the material it's going to work with, give the assignment, the end goal, what is important to achieve with this, Maybe some limitations and maybe iterate a few rounds. "I don't like this, I don't like this." But I think it's scary. Usually it thinks out a better solution than I do. So I've started to give less and less instructions because it comes with pretty good solutions that I like.

And then I can give a little feedback instead of "I think this way, I think that way." If I give too much instruction, then I get exactly what I want and that may not be the best. It's pretty cool to get another throw of A in.

That's what AI does very well, it can actually bring up things and perspectives that you haven't thought about yourself. And now it's clear that in many cases, umIt would be impossible to throw in I can't promise that.

But I wish I could. When they released this, they said they could generate more images. They also demoed it from OpenAI. I haven't gotten it to work, but they said if you want five pages in a series, you should be able to build the story line over more. I can't get it to generate more. It will show that it's generating more, but then it will only be one image. So there might be some technical bugs around it.

But they have promised that it will work. And then it can be like: "I need five slides." And then it makes five slides in order, or builds a narrative over several images.

It's fun that they still... I put a little... I mean, I've done a lot of GPT and opening. It's not just the image that is their focus. But potentially they could with their billionaires models compete with Midjourney or something else where there is a little more I mean Mm-hmm. Mm-hmm. Mm-hmm. They're not just niche-ed to create images in chatGPT.

No. I thought we could do a live generation. Could you just find one random company or organization that I'm not prepared for, so not AI either. If you just throw it out there. Something big and famous. A poster that talks about AI. Use a graph identity for, and just find a big company or organization. Top of mind.

Should it be an AI company?

No, you don't have to. Take whatever. It's fun to take something else. IKEA. That's good. Let's send it away. What I experienced, I've done this before, I don't want to pick them up live, but... I had a course with an organization and I said: "Take the identity of this organization." And it just found it on the internet. They found other image material and then just taught it that style. And they were like: "This really looks like our graphic profile." So you can do it for IKEA too.

So it will be fun. So a poster that is about AI uses the graphical identity for IKEA. Now we'll see what it does for something. But there is quite a lot of material on the internet. What did you say? It was fun to do it with an Ikea manual. The instructions for building Ikea furniture. It's completely without text. You could get an AI poster with just the Ikea guy. I don't know if we can do it, but I saw that it could actually create a PowerPoint presentation.

And that is totally worth it for many if it can create... Good PowerPoint presentation.

Just that. We got a picture here. It looks like Ikea. Many pieces are correct. Maybe it's... IKEA has a lot of pictures and stuff, they don't have a lot of other stuff, but they've taken the right colors, the right logo, it's become a thing. I'll save this for later. I'll show you a bigger picture so you can see it and look at it. I'll show you a bigger picture. Oops, my son is bugging. UhAmanda. Yes, that's the top of it.

I can't get it to work very well. Here, I can understand it. Let's scroll down a bit. It will be 2K resolution when you run it via the chat, and 4K when you run it via the API. Then you will have to pay for the usage. But 4K resolution is enough for a lot of pressure in A4 format. It works for writing posters and flyers and so on. Probably for a roll-up as well. They have built a little, it looks like a typical IKEA picture, so "Automatically inspired, planned, created.

Technique that makes life easier, step by step." Yeah, it really looks like an Ikea poster they have. Thanks.

-Very good.

Mm-hmm. It's so f***ing cool that he hasWhat do you think?

I just wanted to interrupt you with a PowerPoint question. We have to work a bit during the week. Next Friday is a red day, but in two weeks when we do the live again, we might have a lot more fun examples to come up with.

I don't know if you've got any questions here now, but what I've heard is often... I want to be able to create presentations with AI. I'll live that up next time. What is possible and what is not. Yeah.

-But the others seem to be investing in that. Because even Anthropic had something in their new Cloud Cowork and stuff that also makes things better and stuff. It's also a micro-trend right now that the others are trying to steal some of the co-pilot users by getting better at that.

We talked about Open Source last time. And Gema 4 specifically, I have tested Gema 4 myself and it is impressive what a quantized model can do that is locally installed, that is not as gigantic as all 5.5 and so on. Now it's still a bit technical to set up all this yourself, but you can see that these smaller local models that can actually be bought on computers are starting to get very good. Awesome. Hmm.

Yeah. --------YouTube clip YouTube is some some pasta can't grab a good transcriber in the referent from them but they said that we can for a link to the video clip also, but mom and her elixir and Tim mess sending Adam Prata Rambo Gori and on the head from there I mean the via pekha to the vizier cellar so we take it a mere trustworthy for the very nice and go into the bio album and suck filter era soccer so you're via little snap men's click on manSo I think, can the SVT, we can also have a little...

We don't require that a human employee always does 100% of the work.

um That'll be it. And now we've been playing for a long time today again. Shit, we've been doing this for 45 minutes. And next week is May 1st, so we won't be live. But in two weeks, we'll be live again. Precisely. While we generate AI images and presentations, everything we could do. Maybe not just on Friday, but yes. Great. Thank you for today. Have a nice weekend.

Maybe pause it now then, at least. It has been driven on angels.