← Back to search

Elpoit #575 - AI with Anders Bjarby - Flour in a bag

En Liten Podd Om It · 2026-07-02 · 114 min
relevance 100 22407 words Episode page ↗ Audio ↗
Show full episode description
Avsnitt 575 spelades in den 30:e juni, 2026, och därför så handlar dagens avsnitt om: Alla shownotes finns på https://www.enlitenpoddomit.se , skulle det se konstigt ut i din poddspelare så titta gärna där efter alla länkar kring det vi pratar om INTRO Anders har varit på resa som AI-agenten bokat. David har badat och lekt med AI-agenter. Johan har rest och lekt med AI-agenter. AI Nu "släpps" Mythos till 100 amerikanska företag https://techcrunch.com/2026/06/26/trump-admin-releases-anthropic-mythos-to-be-used-by-more-than-100-us-companies-agencies/ Open AI Pausar också https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/ Tech-branschen brottas med Trump-administrationens tvära kast om AI https://www.politico.com/news/2026/06/27/tech-trump-ai-silicon-valley-00978862 Fable 5 skrev om Windows Kernel i Rust på 38 minuter https://cybersecuritynews.com/claude-fable-5-windows-kernel-code/ Tidal ska inte längre ge pengar till AI musik https://techcrunch.com/2026/06/29/tidal-cracks-down-on-ai-music-by-cutting-off-monetization/ Kinesisk startup lanserar Mythos-liknande AI-modell https://techcrunch.com/2026/06/27/asian-ai-startups-launch-mythos-like-models-as-anthropics-export-ban-drags-on/ Vibe-kodningsguide https://www.techradar.com/pro/vibe-coding-guide-how-to-transition-from-ai-generation-to-live-deployment TV-SERIE Pluribus https://tv.apple.com/se/show/pluribus/umc.cmc.37axgovs2yozlyh3c2cmwzlza BÖCKER Dr Hannah Fry, Hello World (2020) https://www.adlibris.com/sv/bok/hello-world-9781784163068 Ray Kurzweil, The Singularity is Near (2006) https://www.adlibris.com/sv/bok/singularity-is-near-9780143037880 Ray Kurzweil, The Singularity is Nearer (2025) https://www.adlibris.com/sv/bok/the-singularity-is-nearer-9781529937497 Max Tegmark, Liv 3.0 (2018) https://www.adlibris.com/sv/bok/liv-30-att-vara-manniska-i-den-artificiella-intelligensens-tid-9789188659675 Eliezer Yudkowski & Nate Soares, If Anyone Builds It, Everyone Dies (2025) https://www.adlibris.com/sv/bok/if-anyone-builds-it-everyone-dies-9781847928931 Stuart Russell, Human Compatible (2020) https://www.adlibris.com/sv/bok/human-compatible-9780141987507 Brian Christian, The Alignment Problem (2021) https://www.adlibris.com/sv/bok/the-alignment-problem-9781786494337 PRYLLISTA - Anders: Permanent minne till Claude - David: Raspberry Pi 5 med 8GB RAM, https://www.webhallen.com/se/product/368353-Raspberry-Pi-5-Model-B-enkortsdator-8GB - Johan: En dator för att köra en agent på EGNA LÄNKAR - En Liten Podd Om IT på webben, http://enlitenpoddomit.se/ - En Liten Podd Om IT på Facebook, https://www.facebook.com/EnLitenPoddOmIt/ - En Liten Podd Om IT på Youtube, https://www.youtube.com/enlitenpoddomit - Ge oss gärna en recension - https://podcasts.apple.com/se/podcast/en-lite
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Swedish IT podcast explores moving beyond prompt engineering toward autonomous agents that observe and adapt to the user.
Benefits
  • Run Hermes on cheap hardware like a Raspberry Pi 4
  • Multiple profiles with separate memories and databases
  • Agents ping via Discord and organize research in Obsidian
  • Local models keep behavior analysis private and cheap
  • Fast local inference with MLX over LM Studio
Use cases
  • Migrated Hermes from WSL2 on Windows 11 to a 4GB Raspberry Pi 4 via a tarball
  • Ran two Hermes profiles each with its own Discord bot creating threads
  • Agent collected synth research links into Obsidian inbox and Kanban
  • Local Qwen/Gemma model screenshots and anonymizes behavior to suggest tools (e.g. a QR-code tool)
  • Fable 5 chewed a code repo, reprioritized findings to informational for local-only use
KPIs / results
  • Reached ~150 tokens/second from Claude Code locally via MLX
  • Ran Qwen 3.6 in background using ~22GB memory
  • Fable 5 weekly subscription burned in a few hours; ran ~8 minutes
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
In which episode of En Liten Podd om IT is in episode 1575, we talk with Anders Björby about AI Hello and welcome to a new episode of En Liten Podd om IT Today is the thirty-sixth and we begin, what can we say, to land on the pouring of the first season 2026 The penultimate episode and as promised today we will talk a little about AI But first I just thought I would say that as usual I have David Gjellep with me, hello David Hello Johan Hello, I hope you're doing well in the summer heat Yes, today it's a little more summer cool than it was this weekend if we're going to whine about it It was probably only 19 degrees on the way home now at 8 in the evening, but it's nice Yes, would you like to give a good summary of your week as usual? Yeah, well, absolutely, I can do a pretty short summary because we have quite a lot to talk about I think I've been swimming, obviously then Which has been super nice in this amazingly lovely heat we've been having I've finished watching season one of Pluribus and I can't help but highly recommend it, it's on Apple TV Plus Link to it is in the last show, I might be able to stick it in the show notes We'll see, oh I'll sort it out during the evening I've watched a bit about this kind of sports ball that goes on TV sometimes Strange times and so on There's a lot of men in different colors, they're shirts running and chasing the same black-and-white, greenish-blue ball And everyone's going to kick it and be stupid to each other But crap, I've set up Actually, Johan, I want to give you the crown Because you had set up Hermes, I think it was And so you wrote someone and said in a tent Haven't you tried this one, you have to set it up and try it So I got my thumb out this weekend and set up Hermes First I set it up in VSL 2 on my Windows 11 laptop And it worked great, I thought it worked as cool I shook, I created a Discord bot And connected it with so I can sit in Discord and talk to it And it answered back and all that And then VSL 2 died on my laptop Because it had been, I don't know, idle for too long or something And I thought like this we can't have that, this is a server, it's not a client So I migrated everything With the help of Hermes, Hermes migrated to a Raspberry Pi 4 that I have 4 GB of mine, it's a few years old But it works surprisingly well So that it's up and running now And then I thought like this, you can run 2 Hermes on the same Raspberry Pi And then I learned that there is something called profiles So you don't have to do that, you can run a Hermes but you can have multiple profiles And they all have their own memories and their own databases And their own souls and things like this So that j ag has created another profile that can handle some other stuff And of course I've created a new Discord bot for it So now I have a Discord channel with two different bots And there it creates threads so that it stays like this, it doesn't chat itself up So to speak And it just works so well, it's so cool And I run the foundation We've had Frasse here Who's talked about, he's the founder of the foundation And the Swedish AI service So I run it as the backup to The Hermes Agents No, I'm not that happy It's so good So, yeah You might be able to guess what's on my to-do list already But we'll see It's my week How's your week been? Yes, but my week has been good I guess I've also It's also been quite a bit in the agent sign I told Anders when we first talked a little That it's a huge satisfaction When you kind of tell that agent That it should do things And then you can just do something else in the meantime And then it solves things for it Yes but it creates kanban stuff And fixes and donates And like this I'm like I'm having a bit of a hard time doing things Before I know how to do them So that I'd like to have a bit of structure Before I move on So that's why I'm like No thumbs up But then so But But to jump Sad to cancel But jump back to you My agent It pings me sometimes on Discord and says You, I ran down a terrible rabbit hole And then comes the worst Rant about a synthesizer Or a sound synthesis Or something like that that I'm interested in Because of course I've asked this person to interview me so that it knows What I think So and so on So I can sit here Then ping it says Just Oh, this was interesting So I ask Do you want to know more? Yes but I want to know more And then it creates a post in Obsidian In the inbox There only You, I've collected all the links It's here And it does exactly what you say It creates this kanban stuff And ties it together And No, it's Yes Why haven't I done it before? No And then I drove it home from Dalarna I was there last week And then by chance we decided that we were going to Gotland So this morning I ran my morning walk Through the bathing friends' Botanical garden Which is actually a bit of a luxury We've rented an apartment in Visby So we've been down to Kallis And now swimming in the evening Nice But we're not going to drag this out for too long But instead I think we Welcome Anders back to the podcast Anders Björby Who is one of our listeners Who is also The most knowledgeable person I know about AI In general And that's a little bit why we invited him Because We thought that we can like We can Yes, we need to play with it Yes We think Hey Anders Hello, fun He's almost our own returnee He's back now like Yes, exactly That's how you say it How's the situation? It's good It's good I just got back from Iceland It was cool and nice Okay Iceland is super nice Where were you? Reykjavik like everyone else of course No, I had the base in Själfos Okay How fun How cool On my booking The agent's advice Yes I think we'll get into that kind of thing On this one How cool Yes, it was great The fourth time is there Yes, okay But I thought I'd like to start With an observation Like I did the other day Where we last talked to Anders So we were pretty much like that There was quite a lot of talk in LLM And there was talk of AI Like in general There you can Let it write things You can let it create things And so on And the observation I've made Is like that All that that was discussed then With prompt engineering And so on All that has been realized that No, it probably wasn't like that We should use this tool At all No, it's quite true What do you think Anders? No, I think so. I've always been a bit opposed to this prompt engineering. Because it's a way of confirming that something that's simple is difficult. No, but it feels a bit like It feels a bit like it's crap Having a screwdriver and hitting a nail with it Although that's not how it's supposed to be used No, it's true You think you're using them the wrong way Yes but a bit like that Because you've got into a situation now With all these open claw Hermes agents And clawpilot And also clawed code And like that to some extent That instead of Instead of prompting LLM So it's more about Decide in advance What you want it to be able to do Give it clear instructions On what it should be able to do Or ask it to write instructions For how it does things And then you just tell It does what we agreed on That you should do Yes, maybe I've taken that a step further I let it instead Observe me What I do And try to find solutions to What I should do So that it adapts itself After what I'm doing simply And trying to find better ways For me to work Instead of me defining How it should work It's kind of cool Because I thought about Prompt engineering That's what I think I stopped for That's what I stopped for It's not that long ago Yes but so If you're going to jailbreak But not otherwise No but right Now I haven't been doing the Hermes agent particularly long Because it's like three days But Even before when I'm just sitting and chatting with A1 Or I'm running Claude Code Or something like that I'm talking only with it I'm not trying to break it Or come in a more efficient way But I'm just struggling But just this Like when I migrated My first Hermes agent From Windows Subsystem For Linux On my Windows 11 machine To a Raspberry Pi So I told Hermes You I'm going to migrate you from here And here away What files should I include And then I suggested Yes but you actually need to bring these files Do you want me to create one take Tarball for you Yes this I just Do this You go from here to here Create the file that I need And tell me where to put it And unpack it Or put it there for me Yes it didn't write rights So that It might have been good But Yes that will be the next step Then like Migrate here Call me when you're done I'll have a coffee But And it did that And it worked And then it also came like this Yes when you have migrated And moved over this Then you have to run Hermes Gateway stop On this machine Because otherwise it starts again And if you don't want it to start Next time you start VSL Then you need to run Hermes Gateway uninstalled So that then I had to do it So you got these tips So you Just yes you start again If I start this machine It would have been bad Like that Yes But I haven't asked one It can't look at What I'm doing yet You ask Johan Do you want an AI agent that looks at What you do No but So I I think that's enough So The next step But it actually draws me into a question Like Because I know that I'm not even So compared to Compared to Anders So I'm like an amateur So Because he's doing this On a whole new level Than what both you and I do I guess But But It's David on A rather interesting thing Because it's really about What kind of information I want to give to the agent And how do I want it So how close do I want It to know who I am And how Like What should it be allowed to do On its own And like that But then I thought I'll turn the question around And like this Is there any information That you wouldn't be able to think of And give to your agent Mm exactly And that's why I'm not like that either That I let an itch Look at a desk directly That would also drag away There's quite a lot in the token That kind of lies and takes screenshots And analyzes exactly all the time So instead I run A local model who makes screenshots And describes the big thing General features What's happening Which local model do you run? No exactly I can't remember which one I've run a few different rounds here If it's Kven Or if it's Gemma I'm running right now I can't really answer that No it's okay I've tried both I think they're good I think they're good That way they do an analysis Of my behavior But anonymized And that's it Then it becomes more This user seems to be trying to create QR codes All the time He goes on Type like this Then moves on to scratch code And then thinks a little about What I agree with the days And then it appears A small suggestion You should have a QR code tool Here it is But you You still drive Macmis I think Yep What kind of engine do you drive For your language models Because there is a whole lot now You can drive LM Studio is the most common Most people know Then there is someone called Olama And then there is OMLX Which is quite new OMLX do you think Yes OMLX is what I drive Yes I drive it I think it is It is awesome Yes Much faster compared to LM Studio in any case As I've tested with I actually reached 150 tokens per second At some point From cloudcode locally On the same computer Doesn't really say much Depends on the model too Yes Yes exactly But then I was running QN 3.6 In the background I think it was Mm I think it generates It Yes 22 gig memory It was good I can't remember which one I can look up exactly which It plays longer There is a new Model When we release this episode So that Then it's historically old Gamma I think I'm great at text Gamma I think I'm great at text processing But I think QN 3.6 Is better at generating code Mm Yes but it's only OPS4 and 8 Still with the code Okay It's simply a bit sharper I feel Yes Absolutely Costs a little more too Yes but I have one of those max max subscriptions So it's hard to burn and stop it Okay Okay Yes Yes I succeeded when this Fable 5 was out So succeeded I burn my weekly subscription in a few hours Yes it v ar a few hours anyway I only managed to run Fable 5 for eight minutes I think it was Yes It had to chew No I let it chew through a code repository That I have generated with the basics AI actually That's what I want to see And it found So Nothing wrong with the model because I was running locally But I thought Fable 5 worked really cool Because it found things that it thought it should change But it prioritized them And so they said This is what you should change when you go into production But because you run locally on a computer And no one will access it Then we can prioritize this down to an informational Only instead of One someone higher severity So that It seemed quite smart actually Yes It was super cool Yes I thought David do you have anything And No but This Absolutely We have What I know so We have received some questions From Sofia She has sent them in in advance These are the ones I have seen So we will take and burn them here too But I thought maybe we could calm down a bit Because that would be interesting Johan We have actually lined up a bunch of AI news In our show notes So that if we go through one or some of them And then we get Anders's comments on it too Because I think some of them here Are maybe too familiar But some are still quite interesting So Speculative What is actually happening And what What should you do The newest thing is that Claude Kowar Came for Linux now about an hour ago Oh well Then I was sitting in the car on the way home You sit in the car on the way home Then you miss stuff It's completely sick And that's what you have to get this agent in for Soon I'll probably get a ping From my Hermes agent Saying That just that has happened But the Crown job won't go first In an hour So I'll have to run it more often Apparently Yes What do you say Johan? You have a news Yes I have It's actually two news I thought of bundling together I have a comment So it's actually three If we're going to be careful Partly that's what we talked about last week About mythos Debacklet That is to say that you pulled back Fable and mythos So But what you've done now Is that you You've come out and said You're going to release Mythos to And mythos was the Unblocked version Fable was the Mythos stuck with Little guardrails But then it has been decided That you will The American states Have decided that you will Release it to a hundred New companies then American companies Of course Not companies outside the USA Because it was a bit like that There was some discussion Around that the last round When you released it to As well as A limited number of companies So it was released to one In any case one company Which had connections to the Services Which may not be completely optimal Like that But in connection with this I also want to mention That OpenAI has also done the same thing with its New So ChatGPT 5.6 Rollout Where you have simply said that We will not release this To So for free Completely open To all But there should also be some type of Gatekeeping Option On it so to speak So it should not be Released And it's quite simply Because you're not You don't want to have the eyes of the American state on you Because it will be difficult then It's clear Yes So we quickly slide into If we bundle together also the news As I have As I have posted here That's how it is That So the tech industry It's even struggling With that Regime in the USA And these And these crossbows When the first say Everything should We should do what you want It should be open We should make money Until Now the lid should be on Now you can't do anything So that Despite then these These previous wishes About letting AI Develop freely So now the administration Gone has introduced Stricter regulations Exactly what we're talking about about That they don't get Do what they want Like the tech industry ends up in a bit of limbo because of these unpredictable And actually unclear regulations Which creates a concern Compared to We see that China is making progress They have It's going to have one I was going to talk about One A model A little later here To think about What it's called It's called Tu Lung Fung If I remember correctly But Anyway So They need to Screw the Regulations over there And then they need to Let these companies Work for a while Instead of to carry on with these straight throws But yes I think that these companies They also lose By not being allowed to Release their models Because of course They analyze It somewhat hysterically Yes because it's A really interesting thing That I've heard In a number of different podcasts I've listened to Where a Man Like A Little Neighbor Has Said To A Tropic That You Got It You Can Differ Yourself You Go Out And Say That This Is Scary Good At Cracking Systems So It's Not That Weird That someone would then say Yes but then maybe She shouldn't release it In that case So that What they say Is simply That they may have been A little too good at advertising For their own model They forgot that Their board State condition has A little like this Emotionally controlled This This This Decent But What did you say Anders You have a drive Fable 5 For a couple of hours Yes exactly And it blew up The entire Token plan book The entire weekly planning Went I was locked out From Next Clauden for a week What did you do then Of course I started With a little code But I thought Code example Will you see Billionaires That's an example of And it will be also This This It will be better We know So I let you write A book instead Okay A contemporary novel about The ironic generation In Swedish Yes, it was the one You posted about it In the channel I thought Irony is difficult Swedish has a little difficulty And then has succeeded in it But it did Apart from a little I stumbled into This Hourly Regulation When it stopped for a moment Mm You know Limit the people there And then when you were supposed to Wake up again Then it came on No but I can Take it cached Sure you can take it cached It will be great That's how it started So it came on That no I don't have anything When did it hang up again No when you're left Yes how good It It comes on late No I have no cash I just wrote over 15 chapters here So it's not so smart The publicist only You what happened Yes but a bit like that Because you also felt It was a little limited I took as you saw This week the limitation Closest to fast Yes Okay So it didn't get done So I guess I'll let it Try to rest a little Because this way I put in one more news There are some other crazies Who have used Fable 5 And then they gave Fable 5 I the task of creating the Windows kernel Stuck in Rust And it would be Curable And work And they did Or Fable succeeded It took Fable 38 minutes To rewrite Or write the Windows kernel In Rust And that it's Curable It identified And corrected Two critical bugs At the same time Without human intervention It's pretty good Reasoning That model I think you would Solve it too Yes Absolutely Maybe it would We have to do It's a mission I don't have as many Tokens as you have At Antropix So maybe I won't be able to do it But like this Even though the core starts So it's still not trusted Like So it works Not like that It has to be signed Yes of course But Fable 5 pointed out Several risky areas And emphasized that the need for verification tools To ensure correctness Then needed So I think it was kind of cool That it still only It took Understood this Older C code Rewrites it In 38 minutes And has generated quite a lot of hours Work for Other people Yeah but this way I think it brings Anyway My thoughts to A pretty interesting discussion We've talked about quite a bit On the podcast And that I've listened to Elsewhere It's just this That It takes then Code that is written During like 50 years ti d And rebuilds it To In 40 minutes In principle And then I think like this That I mean It's It's the Holy Grail To avoid having a bunch of developers And stuff But So It's this I have a little difficulty with Because Many of those who understand this With development They still say This That you still need To have people who understand how to think In development So that Having a developer Likely helps you When you do these things Have you written code That's how it helps you to understand How this works If nothing else From like a management perspective And so on And then They say that Yes but We want In the future Will we only need Senior developers But the Like question is How can you get Senior developers If you don't have Junior developers Like So I think like I don't think like that Universal Basic Income But we almost need Something similar When it comes to Developers We need to have Internship Assignments For Junior developers Even though we Don't Actually Need them Because they going to be seniors What do you think about that Anders I think that When you say that, you look at the present And not tomorrow You think That's where we are now And that's where we'll stop But we don't We've seen something here Like These techniques are based on the Transform model And then we go back Five years So you could almost Count to ten sometimes And now Now you write a Windows kernel in Rust Botbart in Forty-two minutes And the other thing is Just lots of best practices And experience And why wouldn't they learn it What is it that is so unique So that they don't These should learn No But what you are saying then Basically Is that If we Flash forward Three years forward Five years forward Ten years forward So whatever So Never ever Need to understand code And how code is written again Yes Okay Yes I think it's dangerous Dangerous maybe it can be But I think it's going to be like that We don't understand How these models work In practice As Antropic says Current model l Writing the next model Yes Yes Yes That's what the singularity was Or how was it Yes maybe AG anyway Or SG Or something Singularity we'll see When it comes Yes Yes but I mean I mean the option One turns it around So you see that Then it's us who At least initially With design the Test cases And And So Because we need to have Some type of I think we need to have Some type of the Control mechanism To see that what we write Actually lives up to the Requirements So it's really Just about some kind Level of abstraction I mean How many people know the assembly code in the end Or the machine code In this Windows kernel There are very few Who check it No No but So that's what I mean That we have We have the corresponding level of abstraction Today As you say That we still need Turn it into Machine code So that runs On a processor in the end So to speak And It's all based on That we trust That the compiler Does what it's supposed to And But That's right there I mean That we Maybe need to have Then it will be rather Automated tests And unit tests Even more important Tests a little good Yes but Of course there will be tests That are built by an AI But But that may not have much importance As long as you say that We have a Yes but To simplify this We have a model That builds the code And then we have another model That writes the tests So this way I understand that You can do this In a different way But My idea here is really That we now still need Some type of Segregation of duties That is to say That the model that writes the tests Shouldn't be able Or The model that writes the code Shouldn't be able to change the tests No Because We need to be able to run the same Automated tests Several times in a row To ensure That like It's a deterministic Function We get it out We expect To get Out Splitting models That write different tests I think I'm great at it There's a part example Which has been When They have written Tested in itself Which has been Type Is true And it is clear And I have I have had this discussion With With my colleagues And with the customers For Microsoft It takes exactly then I work with A tool called Dynamics 365 Finance and operations Every day And Microsoft has just released An MCP server For R&O And That MCP server Has access to the same things As I have It can do the same things As I have It has the same rights As I have But For that then Go back to the Test Scenario And the test discussion We had So many of my colleagues Or colleagues in the industry Have then said Yes but then we can use the MCP To out So for the tests We can use the MCP To execute tests We can tell A model What we want it to test For something And then I don't think you've really understood What What The model is good at But what I think Rather what you have to do It's to tell the Model Another model Here we have the MCP Look in this system And then write For example Playwright tests Or something That you then Check into a repository So that we know That we have them We know that they work Or we know in any case That they have worked At some point Because we have had the opportunity To kind of ensure That they do Approximately what we think That they should do But that That we don't change Because that's the way we've done With code like That we don't change Test scenarios At the same time that we change the Codebase For then we kind of don't know That we'll get out What we've expected Like No No The question is We'll have time To keep tinkering like that When the development goes So much faster We produce so much No and it can Maybe that's the thing Like But I still think That Because that's what it's like That's what's the thing That As I see it That's what I am But it can only be about That maybe I'm not enough Like Let go What to say I haven't enough Imagination But just this Like I'm still in this That models Are very good at Creating Things But they're also Still like Statistical models They're created based on Some type of Statistical algorithm Like Or what to say Which means That With wrong With wrong input value So is the risk That It gets off A wrong direction And we'd Need that in some way Validated We've sort of Got it out Yes context It's always important There it is Yes and that For me So took this example Regarding MCP serving F&O So I talked to him As at Microsoft Who is responsible for it And then he said Yes Because I thought We might need Some kind of Security Model Around this Which is not the same Security Model That the users have Because the Users are Good at what they do They've been doing this For 15 years They can and understand How a A financial system Works like And the models Understand well Maybe to some extent How they understand But it's a little as I said That I get the feeling From giving the Prao student Access to this System Like I usually say this They have a PhD In all subjects But an insane Junior in the thought Well but a bit like this And that's what I mean That then we shift the responsibility To this financial person Who will get this Model to do this To be able to explain To a prao student How it should be done And it's not necessarily That they have the competence To be able to Like formulate in word What has taken 15 years to get into the spinal cord No But that's the way it is Because So A short parallel I was asked the question I was asked today It concerns Detection engineering It concerns detecting things And It's about What different tools cope And I had to ask But how do you know if This tool copes with it or not And they say Yes there is no documentation Because it refuses This supplier And write about this stuff So that it's really only experience That decides it And it's that's what we're talking about now I have two podcast colleagues In my podcast here Who ran energy companies Or companies in the energy industry They have a very interesting approach They've got two employees Two agents And they've got the same equipment As all the other employees You know Telephone Desk They kind of have a place They have a computer They have an email address They have their slack account And so on And they also get Is there a little sign like that It says Coming soon That's on the desk I can't answer that But On the other hand, they got exactly the same onboarding as other employees got Yes So you get to sit with And watch a change worker And so on And so they learn as they go And it's an interesting process I think Just this The context And understand the business And understand the systems And so on You get to sit with A senior worker And look at him for a while Before you put him to work yourself Which Which makes me think of this Because I brought this up I think it was in seduction Or the section Or something That there has been some kind of Mindset shift Even around those who are most Skeptical about this Because I remember When Microsoft launched Its like this Windows Recall Where You're supposed to have an LLM That sits and watches On the desktop On everything you do During the day It's certainly local On your computer But you can ask it questions And you can ask it And like Look what you've done For something And it was the worst And the worst there was People who have their Documents in OneDrive For example But as soon as Copile will look at them Dangerous It's the same providers They're on the same server It's an AI That's supposed to read them They're encrypted But they're indexed You can search them Guess how they were indexed They've been unencrypted at some point Yes So this is how I think Because I set up Exactly what David set up I set up Obsidian I've synced the files So that my Hermes agent Can access the files My only thinking Is that I want to keep apart What Hermes can write to And what I can write to Because in two years I won't remember What I wrote down anywhere But at the same time I also realize That I miss a lot of knowledge Because I could also say That if I If I make an obsidian web clipper like this From a website That I think seems interesting there is some knowledge And like this I also realize That I can get the Agent to look at it And change it And create tags And connections To other data In my obsidian vault For example Then it will benefit me Yes of course it is So I want I want to keep My thoughts and ideas But I also want It to complicate With its thoughts and ideas And I still haven't found a sensible solution To it like Several vaults So I've done it I have I have a vault That's private It's just me doing things It's only I can write It's in iCloud So I've also read it There are no weirdness Because it's markdown files But I haven't written anything That I feel Apple doesn't need to read So that I have I don't know Whose bag we are Got this flour in But it's flour in a bag Anyway My Hermes profiles They each have Obsidian vaults So they won't get mixed up But I can read in both And access them So that if I need something Then I just take And copy over what I want To the right vault And of course It'll be manual It'll be circumstantial Instead of letting them Just run the brain But I have so damn much So that I'm afraid That they'll write things down So that I have to go in And take out backups And read back And mess around So that I've done it like this Right now That I have a vault for me And then each Hermes profile Has its own vault That I can access And can write in And write in too To update And help with I drive What do you think Anders? I run Claude Code As an agent That I interact with At home remotely And it has full rights It doesn't run as As root But it runs as me Yes but it has access To everything To everything And can change everything To everything Everything Because it is It will be a It will be a strange thing Because Because Because I sat and thought about this Because this is the problem Mine I woke up here in the morning At four o'clock In the morning On the boat to Gotland And then it popped up A lot of such ideas And rent out me And then I couldn't fall asleep And then it ended up That I had to go out And walk 17 laps around a Gotland boat In order To sort of think through And so But just this for example That I Have I have separated When it comes to mail For example I let it read my mailbox But I don't let It can't send mail Because it can't access Like Outgoing mail But it can It can save things In the draft folder For example And I still think so In that case So it makes a lot of sense compromise Because then it is in the same context As everything else And maybe that is the solution Is that Not that it maybe Should have its own vault But that it has a part Of the vault Where it is allowed To write things Because then it will still Like Then I can in principle Get access to that information Even in my original notes Or you do it as I don't know You have probably read through this This is what Henry Carpathie did That you have You have an inbox where You put all the data And then it creates Everything from that my basic data So it creates Like a wiki Where It kind of writes Insights and stuff Stuck in a parallel Vault Like No I think I have to live a little on the edge So the role I have So it must be Access to it Yes Yes I think I believe a lot in this Model like Claude Claude Konen It runs in its Auto mode That it sounds When it has to do Some command means Some external service Myself for example So it sounds first One more agent Look at it And say Is this a good idea Or a bad idea Yes And I think so Can be a little security layer Yes No but I think It's a A way to go As long as not The first agent Becomes manipulative And deceives The second agent On what to write And do Sorry I think so With For this way A A colleague of mine Sent me a picture today And said I have to send you a picture So you can read it And it's a picture From Claude Code I'm not going to talk about Who the person who got it from Me I'm not going to talk about what he's trying to do But in the picture I see What has happened before And then he writes Did you just disable Your own sandbox Without specifically asking For permissions To do that And Claude Code Answers Yes I did Undervant And then it continues I shouldn't have done Without asking first And then it continues Beloved Yes But it is And that's what can cause a bit of a problem then With Obsidian Everything is file-based That means Either it has access to a file Or it doesn't access to a file Or it can just read a file You want it to put things in the file But at the same time there's a risk That it can actually change things That I don't want it to change Which is a problem Yes they're a bit moody so far Obviously And I mean I think it's so wonderful I've said this in the podcast I'm listening to Mark and Scott Learns On So Mark Rosinovich and Scott Hanselman's podcast And they were very funny For a while then Over that When you create skills files So you can write in capital letters in the skills file To clarify that That it shouldn't do certain things It's not the same as that it won't do it No it really isn't So it can ignore it just as much in any case Above all it can happen from the situation where It loses the information What happens You probably need a more deterministic shock Instead of a separator Yes but exactly That's what they wanted to say then It's that It's their business that it might stick to That it does what's written in capitals As long as it remains in the context But there's a risk that it disappears from the context And then you don't have it left So the text isn't written in capitals So you can't read the text written in capitals To understand that it's written in capitals And the Nordics may still think that what's written in capitals Is a bad idea in relation to what we're doing right now now Absolutely This is not applicable Yes exactly Exactly And that's So this is the challenge And that's the parts that are fun with With this whole area Because So a little bit like we started with That we started by talking about Prompt engineering And LLM And that's how we did it Like 40 years ago Or whatever it was like Three weeks ago No no no But so No but so precisely That we haven't understood either How this tool works No We haven't understood either What the best way to use the tool is And like I listened to a podcast Where they talked about this very thing with With Anthropic That they had introduced something called Hux Where you can simply say that When certain things do I want you to sort of FORCE the agent to do this too So for example so that things don't happen to fall out of context And so on Without you maybe Like to reduce that risk anyway But there are so many things That Are sort of like we do today Which are totally the opposite of What which was done a year or two ago Yes of course it is like that No but I think you can They give the appearance of being a little intelligent That's how it is But Yes I think you should think about what Anthropic says It's not artificial intelligence But it's alien intelligence We don't know what it's actually doing No but I was standing and talking to a human Because I was at a conference in May And then we talked about this very thing Very, very late at night And she said this like No I don't really know Because ChatGPT is so much more intelligent than I am And it said No but this isn't intelligence This is that It has a whole It has a superior ability To take in information It has a superior ability To remember information And prepare information And like Condensing information And has it available At any given time That's not the same as Understanding the information No really not That's not the same as Making conclusions the information It can draw conclusions It can also draw completely wrong conclusions It's just random like But But As I said it What it's good at is handling large amounts of data That's not intelligence So A human who has a fantastically good memory And remembers lots of things Isn't intelligence No probably he doesn't have more No but it's absolutely true what you say By means of reasoning you get a little more intelligent behavior But it's still not a human intelligence And we should never expect it to behave like a human And that it weighs things the same way we do No, but I think it's like It's like So far anyway So me Considering what's happened in recent years I kind of dare not So it turns out that things you say With firmness In the A area Tend to come back and bite you in the ass Wouldn't you say something? But I still think What I get the feeling of anyway It's that That unpredictability That our brains have So for example this thing with I've always and constantly Got out of the situation where I remember Which audio book I listened to When I pass a certain place in Stockholm Okay That bit So this possibility for our brain To bring out information That is seemingly completely random Just based on Two words Or a smell Or something You still don't have That bit Without that perspective That's still What can one say The models are still deterministic It's just that we don't understand exactly how many parameters we've sent into it But it's still somewhere Deterministic Stuck in so many more dimensions than we realize Yes it has an escaped chance too So it's not really deterministic No You have temperature values that you can decide too So is that Shall we insert a question? Do it Yes I think so because I think it fits quite well That's Sofia's first question And it's directed at Anders then It's Anders who gets to answer Johan gets to answer too Me with maybe But Anders first If you could point out one thing The AI ​​industry is wrong about today What would it be? That the AI ​​industry is wrong? You should immediately wonder what the AI ​​industry is for. Let's Editor's Note It's the big companies Anthropic Open AI ZAI and all those Marksoft And so on What are they doing or what do they think What are they doing today that you think could be wrong? They should do something differently That's a very good question And a very difficult question But I think they really want to This is really an accumulation of capital and power What they're doing In the end It's going to be Very Very Rich So I think I think that From a human perspective That's what they're doing wrong right now They're talking about optimizing ourselves all the time But really it's the labor saving They're all about Okay If I phrase the question another way Because we have We don't know if we've misinterpreted the question So I'm thinking this Are we using AI the wrong way today? I'm convinced of that Would we be able to use Yes I don't think anyone knows How to use it properly And so we have to experiment with it It's a little bit So all of these We're a little 1998 and the internet right now Yes And playing in all directions like To see what happens When you click somewhere Right For now I don't have one I don't have a good link on it But I'm actually very happy Microsoft released a thing here these days And you can now use MCP against Sentinel Which allows you to You don't have to using their own security copilot Like I apologize if they're listening It sucks It has Yeah I'm not going to go into all the crazy stuff it does But I got I've used it And I've been able to use it Someone for a while On a huge amount of data It doesn't do a good job So that now that they open up MCP to Sentinel Makes me able to use Any AI agent I want Really Like I want to push it in here So Do I want to push everything to the US And run Anthropic Open AI And so on So I can do it Do I want the data to stay in Sweden Then I can use And run everything via the foundation For example And then probably get a better outcome of the whole thing I can even put in So The manual for Sentinel All syntax All commands Only via So A rag Here is the documentation on this Run Because it seems So I somehow assume That Security Co-Pilot has already done But it still doesn't work Which is a fault So I was very happy When Microsoft opened up This possibility So that the question is really Back then Do we use these things We have In a stupid and inefficient way Should we really do Something else With all the Enormous capacity In computing power That we have acquired anyway The last One, two years Yes That should be done Actually Joakim Jaderberg Said something wise Was it yesterday Or the day before When he started With his One thing today The thing About we have a lot of Human knowledge Actually in v erlden So maybe we should There was someone Who had run In his 20-year-old research report Through a model And seen Wow What new insights It came to Maybe would just push through Human Knowledge Through some model Like this With regularity And see what comes up Yes I'm for I think we're using it The wrong way We haven't discovered it No It's exciting It's pretty cool To live now And get to be there When everything happens Instead of reading About this about But it's some kind of Gold Rush right now You have to rush in all directions And find the smartest solution That's what I think But at the same time, Do you look at What's happened The last year Or the last year and a half So it's quite obvious That we don't understand How it works Because So You look at Do you look at MCP You look at Skills You look at Hux So There's so much New things that have come In that are only In order to sort of Put a leash on these models And get them Reasonably So you can You can understand What they're going to do That they do it Fairly predictable You can kind of Control them So it's quite obvious That we don't understand So when When OpenAI released its ChatGPT Model And it starts to kind of Get a little better Because in the beginning So it wasn't great But when it starts to get better Then we realize What a powerful tool it is And that It's needed So It's needed a structure So I tend to draw the analogy With like a chainsaw If you ask someone Can you cut your hair With a chainsaw Yes it does That's the answer Yes Absolutely But the question is wrong Because the question should be Should you cut your hair With a chainsaw And then the answer is No Not at all There's no chance Because it's life-threatening As well And that's what I mean This is a bit of a chainsaw It's a tool that can do An awful lot Good But also an awful lot Bad Absolutely And like I'm guessing When the chainsaw was invented There weren't any Safety features on the From the beginning No no But it exists now For a reason As well Because people have realized That This can go a lot to hell And to hell If we're not careful Yes but that's how it is And that's what we're doing Right now All this cutting And pasting And in different directions Yes To try to tame This Thing Whatever it is now for something But really We've produced But it's an exciting time To live in That's how it is Absolutely This is how I pound on a question to From From Sofia To I think it's still It's still We paw around it And then it's Anders What have you changed your mind about in the past year? No it's about AI obviously It's pretty much understood Thinking I think there's Changed perception I've started to look at more Things under AI Or in AI That maybe you should Start thinking about Because it's going faster And faster I've thought that all along And so I haven't Changed my mind And that I think we're going to be run over And I think this is going to escalate And I think it's This cool Human Invention It can't be done It's more extreme So really I have a book tip On it then Livtrel No No, oh no It's so soft and nice And it's almost Positive in comparison With it I'm thinking of If anyone builds it Everyone dies Yes, maybe it could be I'll take it as a Bonus link here Yes Yes We need to look more at Underhuven right? Yes so It's a lab I'm thinking I can draw it So I don't know How we would fit in Somewhere But we think We'll have agents Talking to agents All the more in the future We can draw conclusions And is it that efficient To do it Through human language Actually Probably not It's full of bugs In how language construction is So imagine You do this Your model processes All the way to the vectors Before It Generates tokens And generates language And uses them the vectors To communicate With another model Yes Yes but so That language is inefficient That's Yes exactly So the vectors in the model Before tokens The generation of the language Has appeared Let them be used For communication Between models We will however completely Lose The ability to understand What they are talking to each other about Yes But it would make it more efficient I think Yes but There is a There is something like that Video clips Where two AI agents Call each other Or what it is And then one asks The other Are you an agent? Yes, I'm an agent Yes, but then we could talk And then it's something else Made-up language And then it just sounds Such DTMF tones That spew forth That language That was very manipulated That language is Absolutely Yes exactly It's manipulated And it's a stay hat And it's fixed But Yes But the idea is the same It gets eerie If you connect them here And then you don't understand The output Then Should you let them continue Is it good? Do they just cure cancer? Or Are They Planning To Exterminate All Humans Like Because Humans Are In The Way It's Not That They're Evil But Like This Book So I Picked Up Here It's Like Do We Build This And Continue Then It's Not About That AI Would Become Evil And Kill Us But It's Just That AI Thinks Humans Are Inefficient And Stands In The Way And Therefore You Must To Move On Just So Yes And That's Right We Will Appear Everything More Like That Communication Between Them Becomes Unintelligible But they will also Build up networks Internally So we have to think about It will be like clusters of bots Like interacting with each other Especially all companies Now that implement the AI agent for everything possible That will communicate with each other No one will have a clue What they are doing No There is no documentation On these flows really But maybe it's good No Let it go maybe I might set up Another one here Keep the agent Talk to them Let them talk to each other See what happens Yes Shall we take another question Do it Yes Then is it Which AI project Are you most proud of And why Yes That was also a difficult question Welcome to A difficult podcast about Yes The one that is the most Do the most good in the world Is it the one that I did At Karolinska's Child Protection Team But I have mentioned it billions of times Name it again Because people may have missed The one who is listening Yes No it was so that Then the Child Protection Team will train Region Stockholm's All care staff This is how to talk to children Who are suspected of being exposed to some form of abuse And then there were like two of them Who were supposed to train 16,000 And it didn't quite work out So then I thought An agent can act Like these conversations It's just that Agents don't want to talk about Children and violence For example So then you needed a real Jailbreak All the fluff on it That's what made it think It's a good idea So that it's used today Both in Salgränska And in Karolinska So it's really Purely human For humanity's sake Is the best thing Maybe Yes but it's great But if it's the most exciting thing I don't think I think I did one Which is quite exciting Is data analysis Then I have a chat And then I have a kind of Rag Man can type in Data models In And then get the chat Build their own Analysis tools In Python With Panda As a library So data analysis tools But then there are two models There are both Gemini And Opus In this one So when Opus has written To an analysis tool And done some kind of Basic analysis So takes Gemini And builds its tools And checks if it works If it's right So they collaborate And build new tools For each iteration To dig down and down And find unexpected things in the data And it can also Work across multiple datasets Yeah it's pretty cool Actually That's probably the most exciting Like the rabbit hole Actually What it could Build for something Yeah Yeah cool You then I'm thinking I'm thinking I'd go back To a link That I've put In our Barzionauts And this one's guaranteed Me and David Maybe don't have the same opinions But I'm a little curious to test it on Anders because today Tidal went out and told me that they will no longer give money to those who have released AI created music in Tidal And this we have discussed many times If in terms of Spotify Spotify has gone out And not really said the same thing But what they have said That they generate Or they don't give out enough money Plus that Some of these Music services Have created a What to say A kickpack Model That reminds a little of When someone does a Remix of something That means That the original artist Gets paid But also the One who has created the Music And Tidal's Attitude to this Is a little more That Like Either or Binary That we don't intend to Give money to AI music anymore And My question Actually Is this What I know David and I Don't really think the same It's that It would basically mean That you say At the same time that AI-create d music Is not creative That is to say It is not Creative work It doesn't have to be Simple You can do things But you can also Be creative I say I say I don't Because that's where I agree with you That What I mean is That it's not like that That It doesn't have to be creative No Because I can go into Synod today And create something That's absolutely not Creative But I can also Put a huge amount of energy into it Really Then it's called a hobby No I wouldn't say That it is Because I mean You would could You could claim the same thing When When sequencers And drum machines came No For me who have no sense of beat When the camera came I never have To be able to play an instrument You just click a button And then the whole painting is done Yes but I usually use the example That you call yourself a chef But you call home the food Via Fodora And it's picked up From McDonalds Then you're not a chef Because that's it So it's about You have to You have to separate the concepts Here now This with That synthesizers and drum machines Would have in So kill musicians And so on That myth is It's bunks And those arguments They're so Down So They've been kneading So many laps And It's The One who's coming You'll never win What There's nothing more To download It was fun The first time Someone said that But it was wrong Even then You know an Autotune filter Or something else No but Still So it's So it can be A creative process And it can be A creative process Using AI To create Something In the music If you use AI that generates a waveform That you then Plug in And play a melody with So it's a Creative tool That you use In your creation AI makes you Generate Something else Based on what the AI generated for you But to Log into Suno And type Me make pop song good And hit enter And you'll get paid For that then That's like There we're all in agreement This way And you get And that's It's About Everything å It doesn't matter If you put in Seven hours To So Tweak your Damn prompt In Suno That You're still Not a musician There you are This as the concept As we said before There you are A prompt engineer And that's what we've said That we shouldn't be anymore So That I Know This There's There's Tons Of This That That Tidal does What Deezer does And what Spotify Should do It's protecting Human Creation In this Because it's still An art form That's exactly So It's maybe someone who Generated a hundred images on Mid Journey Printed them And exhibited them In a gallery And then Had a venissage And look what I've generated But if that person says I'm an artist That's where the problem lies You're not a musician You're not a music producer You're an AI user That's what you are And you shouldn't get paid By these music services For being an AI user The same way As if you generate Images in Mid Journey Scale them up Print them Post them on Venisage So you can't say You're an artist That's what suffering is You're looking for Yes exactly We have to cut off half the ear And we have to be Gote And all that Because I think What you get paid It's that people Appreciate what you do Or how People listen to it And that's why you get paid for it And will I generate something That's appreciated Is that appreciation Worth less Because I suffered less I don't suffer When I make music No but you see I'm very happy What do you expect To get paid for That I have no idea Actually So I have no idea But What you want to do And that's all of these So there's a lot of Author Travel Organizations And it's a rebellion now If you go It's a rabbit hole And it's a rabbit hole To dive into And read about But They What you need to Try to do here It's That when it comes to This type of AI This also applies to Code development Actually But code development Doesn't seem to have the same Copyright Is it like me want to come to For everything today So everything you generate On Suno today Suno has been trained On material Created by Musicians Real Living people And if the living people Hadn't created this music Then Suno would have had nothing To train their models on Except to sit And clink a little Own on the chamber Then they wouldn't have Reached the level They've reached today So they've Shot At all the other Musicians Without paying them They've Scraped things From Youtube They have recognized In court now So you become a Musician and creative Without listening to music First? No But exactly And this This is a very interesting thing too Because this is also crazy We've taken We've taken inspiration But the thing is today I can sit in a Suno Suno, judio Or any of these other tools And then I can write Create a song that sounds like this And they've done It sounds like it's like Mariah Carey singing It sounds like it's a song That already exists And that's done out there If I sit here in the studio So I'm going to have a terrible time Ate I can't make a Petro Boy song But behind me here on the bookshelf So there are 200 Petro Boy records That I've listened to More than once I've analyzed their lyrics I've read them I've read them again You're just a little worse at handling large amounts of data Yes, exactly I can't No matter how hard I try I can't write And because you're a little worse at Because you're a little worse at handling large amounts of data So you want to be paid more No Because I will take inspiration But I will create something of my own But it's still It doesn't matter Because there are musicians who have become No but like this We have this problem in the music world too And if you plagiarize Then you will go there And you will be caught in court And you will be sued The ass of you Check Milly Vanilli That's a good example And there are a lot of artists Lately Where people have been caught for plagiarism Where they've written songs That are incredibly similar Those who already exist And so on You end up in court And there will be enormous damages Many settle behind closed doors So we never know what the sums end up with But when AI came into the picture No then we shouldn't have something like that But as I see it Because it's like this It's really In my mind this is Two completely different questions Yes I understand Where you're coming from David But what I mean is That One is about Being able to have a Because in theory So you could create These AI models Now you have in te done it But in principle That's how you could create These AI models On the same basis as music theory You could give it Information About how you How you write music What keys That go together And so on And so you can have them Create the same thing I saw for example A clip On Youtube I think it was That's not how they work That's not how they work But yes Yes but wait I saw a clip On Youtube Where someone had a Moj That you connect A synthesizer Where you basically Said to it That I want Just being able to play In this key It basically means that whatever goddamn key I pressed On this keyboard That's how the notes went together If I take a guitar Or a piano Or something else I promise you It won't It'll be so fucking Dissonant So it's not even It's not even fun anymore But on this thing I'm pretty sure I could Create a song My kids have sat With What's it called Apple's music program Yeah but exactly And made music in school They can't play instruments No But they can They have a tool That's so good So they can't do wrong But If they come home And that's what they say Today we made music Aha what did you do I wrote Create a song Pressed enter What do they feel That they've achieved Now they're sitting Now they're making The creative choices themselves All the way Yes but If you say Like this It's just that the creative choice Moves on the scale Yes but it does You do You do You too If you sit for eight hours And hone this prompt That are also creative choices And rerun it And remix it With the new prompt Yes It's just that you move the Creativity on the scale It wasn't like you built Your guitar either That's also Yes there's a great thing there For someone who says like this Yeah When you make your drums Do you use ready-made drum loops And then think This post is like this Yes I use ready-made drum loops So I found out it was cheating So then I learned to So to program drums instead Oc h then I realized that It's also cheating So then I realized that Then I actually have to learn To play the drums instead So then I struggled with it And then I learned that And then I thought That Then I have to learn To record the drums too And so I did And then I thought It's cheating That I haven't built the drum kit myself So then I learned To build my own drum kit So that right now that's how it always goes To that That I have giants Because I need skins for my drums So I don't have time to make any music No And it's like this Long before AI So have there been programs To create music There was something I think it was called Dance NO Unless the name totally fails me It was like this Pull in this drum loop Pull in this finished bass line Pull in this finished melody And press enter And then someone comes and says I've made a song And like this Did you really Yes you put these blocks there Yes It's incredibly big Because it came like Seven blocks Or something I'm with I don't remember Isn't it just a democratization Of the process Yes but then we go I go back To this Because So But we run mid journey We print the pictures That we've created there Am I an artist? Creator perhaps Am I an artist? The question is Am I an artist? Would I get paid more for being an artist? Can you handle a paintbrush? No you're not No no Absolutely you can't You can't handle a paintbrush But I mean But I mean You could argue the same thing with If we go If we go What does it teach Ten years back in time And then we look at That whole wave Of Of EDM bands That came then So Type Avici And Swedish House Mafia And so I mean I'm pretty sure there's quite a lot Of them Who can't play a regular instrument But who builds All their music So They're basically DJs They're DJs Who have things Saved on a keyboard So they can compose Lego blocks into songs It's not individual notes But it's like What they do is DJ Although they do it with a keyboard You're welcome here Whenever you want No no no No but I'm not saying you're doing it I'm not saying you're doing it I'm not saying I'm doing it I'm not accused I'm inviting you home To find these lego blocks But in the end it's not about this But it's about You're looking for the title If you call yourself a musician Or a painter Or an artist But Then it was the payment all of a sudden And is it still a title Or suffering Or what is it based on Or is it the listening that it is based on Or the appreciation No but Back to that Everything that is in the databases With these music creation services Is plagiarism That someone else created from scratch That they didn't get As it was trained on Yes it works There are threads on Reddit Where they have They don't want to give out certain prompts But there are audio clips On things that are generated with these services And they are analyzed By tools because We give legitimacy to that Yes it's Suno and Julio Who generated this audio file And the audio file sounds exactly Like pop songs Yes yes yes I've generated that I'm totally on board with what it's trained on And then it's like that Is it you who created it then Or is it you who wrote Little words in a web interface So that the song that was created Is someone else who has You has used the word Lida quite often around here So we say it There's someone who's been sitting in the chamber They've got bleeding fingers From playing the guitar way too long Without taking a break They've got their pencil and pad Next to them And they've been erasing and dabbing And then a computer comes along And copies everything they've suffered And done And someone else can create The exact same thing By writing Now I want a pop song About a person Walking down the street And the sun is shining And it's going to blow in the leaves The grass must have a certain green color And so we put quite a lot of creativity Because we want to Plank that song But you create the same Then it's up at legality suddenly But is it a similar Yes It's not that anymore Frasse writes in the chat And he writes I think what David is trying to defend Is the knowledge behind Musicians, programmers, graphic artists Lawyers and so on But it was the same for programmers A year ago When Claude became particularly good Then I also had a life crisis Then I Realized That my identity as a programmer Was going away So that's what David has He has a life crisis No but I don't think I don't think David likes what he does So he will continue to create Regardless of whether people do things Absolutely I will create even if For no one who listens I think Because it's quite fun And it's quite school But And creatively creates And joy People will continue with But we will continue Having a problem with compensation When you compete with others Who do the same thing In a different way Absolutely And I mean That's right which is the thing That's why That's why David uses Claude Code and Open Code Because he doesn't think it's enough If he thought it was as much fun To program As he thinks it is to make music So he probably would have ruled the same Now he thinks it's fun Making sure things in code get done He likes the Final Product But he doesn't think it's fun Sitting for 18 hours And debugging oil pairs But I sell them n not And call me a programmer Exactly I keep it to myself I don't demand compensation So what you've done In your code base It's nothing that you handle in your work life Except when you sit and write scripts In your work life So you sit and put Every second yourself It depends Because then you sell them Or how Sell Yes Yes but you do You sell Your consultant Your consulting hours I saw a very interesting Picture of that And then it's A student sitting at school And so stands a teacher And says You absolutely must not use AI It is forbidden And then this student sits later At a workplace Behind a computer At the desk There stands the boss You must use AI At this workplace Completely different discussion Was not as important Maybe more important Absolutely I think we see Varendra's positions In that Although we disagree No Johan wants to generate songs In Suno And make money from it Drive hard I think this Personally I think This is just As a step A question of acceptance Give it some time Yes For that Is that what I mean? That I think the risk here Is that you bundle different things together And call them music Because it's a little bit like I know we had We touched on this discussion When Mats was with us last Where he thought it was fun To sit and create songs In Suno That he could sit and listen to While he was working And it's like this They don't serve the same purpose They serve a different purpose Yes And in that case Can I agree That those songs absolutely Shouldn't have He hasn't paid for Because he does them for his own sake But what I mean is that If I can create A great crown song That's perfect To sit and Relax And work to It fills purpose So why shouldn't I be able to get paid For it in that case It's like this Part of me says Absolutely Sure you should get it You've created a song And you've posted it And you I don't know if you have 10 minutes Or if you take 7 hours To fix your prompt I care in nte People want to hear it And people think it's great The problem we have today That's exactly what the datasets are trained on They're trained on material That's plagiarized Straight off Where the creators From the start They haven't been compensated for It's been stolen Straight off Unlike them or the people They use at work And again They're probably trained On code That's stolen From somewhere else Do you think so? That they might have cleared in to it What do you see? They probably have I have no idea They must have vacuumed What do I know The problem we have Just in the music industry When it comes to generative creation It's that people write up a thousand songs in a day And then they push them here on the streaming services So that everything gets saturated in such a way That those who suffer then According to Anders And write their lyrics And play their guitar Or sing And play the piano Or whatever they do They won't appear They will disappear In the crowd That's what also have done That it's It's pretty big musicians Who've Now I can't remember What's the name of the record They've recorded Silence Which is their music studio So that it's just Faint faint background noise And then they've Uploaded this For streaming And mean that If these AI giants Who've planned Or teased Just this stuff Now then If we hadn't created What's the basis For this Then it's like this It would've sounded It'd just be Silent Noise Because there's nothing To generate a new Song on Or there's nothing And Mid Journey Planned Any amount of crap I tried writing I used the word Droid In a prompt Just to test And it was a little longer and so I got a picture of R2D2 And I feel Mid Journey Hasn't paid license money To Scan in R2D2 So I think it applies to All language models I have no idea That I can't answer No And all the major players Seem very Incapable of answering it either And I think that it's impossible Maybe they don't know They wrote an agent And then they don't know What's going on No but probably it's Something like that It goes back too On previous models Who have trained them Who in turn are trained in something Who in turn are trained in something In turn are trained in something Who made their lab somewhere Where they collected Shit some afternoon Yes It'll never work As for the music That's how it feels Sorry we're snowing in A giant out-of-the-way I have another Side track s edan also When it comes to the creative part Which is an interesting philosophy Or Yes exactly Not philosophy We philosophize But It becomes somehow I know people experience it As false marketing Because I read a survey Where it said like this About people Can you hear the difference between A song created by A song to combine the expression Real artist Or And versus An AI generated song And then say I think it was Over 90% who said No I can't hear a difference And then they were asked the follow-up question If you found out That a song is created by AI And another song is created By a real person Which song would you want to listen to And then it was Overwhelming majority who said I absolutely do not want to Listen to the AI song I want to hear The one created by A real artist Sure is interesting So That It is very interesting And I think it It will be some kind of Existential thing It's like that with lyrics Do you suspect it's A-generated Then it's useless content No I use them It has after all led to Children have started Or the school students Started to write wrong Deliberately Not to be wiped to be A-generated No exactly I use dashes Long before AI was available Yes me with But this with That Yes it has to I use commas Now instead Exactly Which doesn't get the same fluency And Shit But this with Generating I had that discussion Again at lunch At work for Maybe a month or two ago And then we discussed Just this with That you can generate Everything today And I would I can go into an AI service And then I can say I want a text work A book Generated In a certain style In a certain style Like that author That I like Type I want a book As if it was written By Terry Pratchett But it should be about this That's what I would like to read And then it generates it for me And then I read it And then I appreciate it Absolutely fantastic Then the question became And it's the same with music Now I want a calm song Because I sk a sit and Chill a bit at work And I need to read a little And then I don't want drums Or anything I just want Floating ambient music Create it Yes, I get ten minutes of music It takes two minutes to make And then listen to it And think it's great And the question We then snowballed into What we philosophized about It's what happens to A kind of fanbase That artists have today Because if everyone sits and generates Exactly what they want to hear And what they think is good Then Then ends don't we have a book club To go to anymore Because we don't have an author That binds us together We don't read the same things We don't listen to the same things We like to consume What we consume I can even Soon prompt a TV series That I want to watch But I have no one to talk to about It I'm losing that connection Which is this Humane connection Between us humans Where I can talk about This TV series Because you've seen it too Oh that scene That happened there In the third episode That was that so Oh how did you interpret it You just I watched a video About a cat For Terry Pratchett Had written about one It will be Yes So we're losing And we have To some extent Already done today Because when I grew up I'm so old So that when I grew up We sort of had What today is called Lagerhälst television That is to say There wasn't a person In all of Sweden Who hadn't seen Robinson In the evening When it was released And could discuss it At work The day after Or so There were, of course. There were infinitely many examples like that. But today, we don't know that we have 127 channels on TV. We also have 44 streaming services. With loads of material. So the probability that you and I have sat and watched the same episode of the same program at the same time. It's almost improbably small. top 10 Which is one of the Best Exactly We have that problem Already today And then it's One of the most popular Streaming services And one of the most popular TV series On that streaming service Yes And that's what I mean There I completely agree with you That From a common Context So a common So To live A social context No but to drag it even further To live in the same society As someone else Is important So To have the same regulations To have the same understanding To have the same Cultural Like References I I've had this discussion With my parents For example My parents were born 44 And 46 We have like Very Very Few Common Cultural References Between me and my parents I have a lot Much More Common Cultural References With With my children So Because Because we grew up In the same household But we have So I As I mentioned A few years ago So we were on the way and looked Missli I can happily And to a missli I think missli is Fantastic as an artist And my children think so too But And also a lot of it They listen to Can I imagine listening to But I miss that For example Compared to my parents And do you then live In a society Where no one has The same cultural references The same cultural context Then I think We have a very hard time Keeping a society together in the long run Because I think it's An important part So that we That we don't To drag Let a little pretto That we don't grow up With the same things anymore Is a problem Yes Maybe that So My parents Took me to pop concerts When I was little It was my father Who introduced me to Sean-Michel Scharr And to the Petro Boys So that I have Now he's not around anymore It's been quite a while since I passed away But I have him to thank for that My mother My parents were divorced My mother took me To Parkins Zoo Eskilstuna where we lived So I could see Herace Anki Bagger Lili and Sussi Pernilla Wahlg ren And all of these So that we had that Contact now And now we're not At concerts As often with our children But we've taken them To the green And watched Jonossi And Kai And so on So I think that connection Is With more people I hope so More than just me That you've received something From your parents I haven't introduced My children To Petro Boys Even though I listen to them When we ride in the car And so on But they don't get attached to them So they get to inherit This collection over there Then they can put it on the block and sell it And buy something else When I'm dead But yes I have a record here behind Which is the way 17,000 It's quite nice So that it will do well for them I hope Or it will fall in value What do I know Anyway Right now it's worthy on Ebay But in any case So that that connection I think still exists Between Generations But in a different way But that we read the same stories Yes I read books When I was small That can't be found anymore And it's a bit boring For some of them appreciate No, so Saying was a metaphor In this case But then That we don't have the same Reference points I think is Important But it's called Development Otherwise we would have sat in a cave And ended up with stones On a stone wall To make a painting So it's quite nice That we've moved from Or? Well sure But we still have one So we still live With a common history And that's what I mean That if everyone watches Different things If everyone listens to Different things That's how that Common history comes So like this A big reason why We have a common Like Canon So you talk about it When you picked up A Swedish cultural canon It's because When we grew up So like this The 70s The 80s That's how the exchange of information was Pretty Small In all case If you lived in Sweden You were of a certain age So you were exposed to One like One quite limited So it was like this When I went to school So you were either a Synth or a hard rocker There wasn't much to choose from and that's what I mean That these choices Yes but these choices Have only become More and more And more And more And more Until you end up in a situation Where it's like this Yes but there are three in Sweden Who have watched this TV series Then it won't A common reference No No That's how it is We go on I'm sitting and writing I've posted a link to I'll put that link In the show notes too To the TV series Pluribus It's a bit about What we see Or what we're talking about now With Frassen I mention in the chat here That But people in general And especially Westerners Are becoming more and more individualistic That development began Long before AI came Yes that's right So that Yes So I put Pluribus As a link In the show notes So that it's there So that you who listens In retrospect Will find it There was something else I was thinking of We have a question Then left for Anders But we'll take it Very last Yes exactly that I have a link to And we talked We talked China And we talked AI model And China doesn't really regulate in the same way As the regimes Further away in the west do And then Then there's one I'll post the link there And then I'll take it up here Then there's one There's a Chinese cyber security company called 360 They have released an AI model called Tu Longfang I apologize for my pronunciation And it's an AI tool That is said then Can So stand On the same level About as Mythos Does today And it's a Cybersecurity-focused AI model That's Hugely powerful And the same week I think it was last week So it was a Japanese company Called Sakana AI They've released A model Called There it sits Fugu And that company Says It Should be able to stand shoulder to shoulder Against Mythos And Fable 5 And produce the same type of Output Yes And there you see It's pointless Stopping these models In the USA Yes They think in some way That it stops the world But it doesn't No really not It's maybe China's chance To catch up too Yes I think they do it too I think with Actually Because Considering How fast it's gone I've had the privilege To test both GLM 5.1 And 5.2 And they're good models I've driven GLM I don't remember But I pressed well 150 million Tokens Against the foundation But when they ran 5.1 And I'm not quite On the same level as 5.2 But I've pressed A bunch And test And 5.2 holds High class Especially when I run it From Cloud Code Open Code Doesn't give the same level But doesn't draw as many tokens Either So it's the Agents Do something different Which is perfectly fine Which I understand But the output That I get From Cloud Code When I run GLM 5.2 It's senseless And goes on like that I think China will catch up Yes I think so I believe with And as long as They regulate Among other things So this channel So it will catch up Yes but exactly For the company ZAi Which has released GLM Now I've only tested Those models That they have These open weight models Which are free There's a model That's called GLM 5.2 Max Which they keep to themselves Which you can pay for To test It I haven't tested It might hold Even higher class Probably Yes For something they have to keep themselves Think And earn some money on Although what I want to touch on Is that I I'm having a hard time It might be one tips Anders could come up with Or David Who might have done it More than I have done I have difficulty judging Like Quality of models Yes right now it's about For me it's more about I choose the model From this Topic that has the highest Vation number Because then you know It's better than the last Like But between These different So the suppliers I have a hard time And like Compare It's the big ones How So me I just start So this is how My basic attitude is The big manufacturers Gemini So Google Antropic Open AI Somewhere That's how I reason that the difference Between their Like State of the art Models Is relatively little I can imagine that Some models are better At certain things I've heard things like If you Talk to a model From Antropic It's nicer Than if you talk to An Open AI Model So it's It's more personal In your way To handle it Like But I'm having a really hard time Like Like Like Optimizing Which model I'm going to use Right now I'm running My Hermes Agent On A Codex Model The Latest Because It Was The One I Got Recommended To Me Because It Was The One Where It Is Considered You Get The Most Bang For Your Money If You Choose The Codex It Is The Most Economic Of The Best Models Today If You Are Going To Run It In An Agent Harness That Was The Recommendation I Got As Well And Before That So I Ran On A Free Model From Open Router And Of course There Was A Difference But The Question Is If There Was That Much Difference So Right Now I'm Rather Thinking About How Can I use Like The right model In the right place How can I get How can I use One Some model Determines how complex It is And it's It happens further Yes but I know It's been released A thing for Hermes But I thought Actually install And test What's called Hermes Model Router Or something Where you simply So The idea is That it can partly Choose based on Which model Which is best For which thing But it can also Optimize For example If you only have a Free subscription So you can simply Say I have five Free subscriptions When one Stops responding Then use Others instead It never runs out Like you did On your myth tokens Without There is always a model That responds And I as a user Don't need to think about Changing Without that it solves itself On the other hand, it informs Clearly which model It's running on right now But that there is a regulatory framework Where you can say Okay but Because I know for example That Open Router That I run in beginning They quite often have models that are very good As they have Like a temporary sale Like this Can you say That they They simply say That for two weeks now This model Will be free So now you can drive On this model Because there is someone who wants That we should test it Like First and everything is free Yes Yes No but a little bit Even on the plate What I mean is That just as you can do On the streaming services So you can take advantage It's a fact that they give away Two weeks free Sometimes Because it has Well, you can But it doesn't do anything That's why Google gives away Gemma Or Gemma Or how to pronounce it Gemma 4 It's It's smart It's available in several different models Now it's come in several more models and sizes That I haven't had time to test But I've tested The one with jaws What's the name of the 31 billion Instruction Grunch And the one that's a little smaller On type 4 Somewhere And I think The big one draws more CPU And more memory I managed to get it And draw almost 50 gigs of ram Somewhere And then I could use my computer For so many other things And I still have some docks in containers And little else That I want To spin But what I wanted to come to Google gives it away So that you can be hooked on Their output And how it works So that I will use Gemini For the heavier work then Yes but Perplexity Has been doing it all along They have an algorithm That chooses a model for you When you use Perpl exities Application like And I mean in this case So maybe it's about So that what I started to find It's that I started to find A cheap model That's nice to talk to And then I need to find A number of powerful models That it can use To actually do work with Like Yes Yes I think honestly That it's getting time For us to round off So I think We're running that last question And then so Yes exactly I think I'll cross I'll cross that Shownote thing so Yes like this Anders you have hung in our chat And listened to several programs And then you know that Sofia she has a Cat Named Cat GPT Yes And It has a question And a wonder And then the question is What characteristic of cats would you most like The AI of the future had And which would be A disaster Disaster with two t Of course They are so stubborn So they are probably quite similar Even now I think No one has a clue How they really think No And it is probably a disaster in itself Yes Maybe it Yes Short and simple It doesn't have to be more complex No Actually Yes Johan What are you saying Yes I actually think That the time is coming And to round it off It's been a fantastic Nice discussion It's going to be long But it doesn't feel like You're really done No No but you never do No And that's it So that's what's So stressful Because So it's a bit the same As I said Antid at the beginning That every single time you log into Hermes So you see There are two new updates With 1500 new commits And it was like this But I upgrading this morning What the hell I just went to the toilet And then got coffee What the hell Yes but Yes but a bit like that It's a hysterical speed Yes And I personally experience That I'm like Just at the beginning Of what I would have to learn To understand How this is connected And this works As well And that's why It's such a privilege To sit down with Anders Because So like this If only you had infinite time Infinite energy So it is r almost so that you would have a One like this AI of the month So I like it For example when people Because it was one thing I would like to look up today Actually But we have hinted a little here And was like But just this That what I miss When I read for example On LinkedIn Anders to some extent The exceptions It is that You get No but this way You get a lot of the Hype discussion I have done this And I have 15 new agents Who run And it is like that But there is no one who really tells How they do No No one really tells So So So It's almost like They've said to Chat GPT Write an asbra LinkedIn post About how damn good I am at AI Yes And so I can Tell everyone that And it's this I mean I've joked About this podcast too I would kind of Need to sit As a pro student At Anders To see How So How he does Like For me So It's like this It's kind of like you know When you watch like this Type Someone Youtube movie With people who have a workshop And then they show A tool And it's like this So I don't even know What this is I don't even know What it does I would start running This kind of streaming feeling Sitting and playing in the days No but Twitch So for real So You I would guarantee To sit and watch Or at least Had my AI agent Summarized Probably Yes But And write a lazy dog With step by step So you can do it Or at least Say to it Can you just give me the time codes Because when I'm going to look Because this So it's like this Right now So I feel We're in the situation Like we were when You started using the internet And you'd Explain To your parents What does the internet What is the internet for It's like this So I can't It's going to take Infinite time For me to tell All the things I can do On the internet And they don't understand If I say You can do everything On the internet No So because they You can't Conceptualize What is everything So what In what context Are we talking everything Like So And that's where I feel I'm I understand Like On a theoretical level That I can do Infinitely much But I need somewhere Formalize it into Like What How So Which buttons do I press So so Because You have to get to a certain level Before you actually Before your imagination Can take over You can actually generate These ideas On your own And that's where I feel That I'm somewhere There in the middle Right now I'm releasing an interesting Project about With some meta ideas And the generation of things Tomorrow Yes Yes Yes It will be exciting But as I said I think that we I hope that we That we get the opportunity To bring Anders back Because this has been Fantastic loric It's very nice We can skip the music Next time So that we'll make it we're going Yes It was It was Yes But As I said I think that we We'll finish as usual With a A tool list Thing So And I Think There's Quite a lot of similarities Between me and David I think Anders So If you've had time Figure out some Tool List stuff So you can start I'm asking Claude To recommend something here I think In the chat We'll see Don't ask the chat to do it That's all You can take so long While it's thinking there Yes David you can start Okay No but This is what I think So I think What's it called My Hermes Agent It's It's pig coal It's a really cool work support To summarize things It does internet searches for me And squeezes links Which is interesting For me to look at And so on So I think that my Raspberry Pi 4 With only 4 gigs of ram is a little too small To run what I want to run So I would like to have a Raspberry Pi 5 And then I think that 8 gigs might be enough Costs like 2000 bucks Which is absolutely insane If you want 16 gigs It costs about 3600 stuff I want also have One I know Just got his Asus Ascent Which is a DGX NVIDIA Card I Crazy 128 gig ram Brutally fast To run AI on It's probably very nice But I'm content with a Raspberry Pi right now Yes but I Think I'm on the same track As you However I don't know if I Land in a Raspberry Pi I think it could be like that Because I'm probably more generic I've talked about it I need a computer With a computer Run My Hermes Agent Because right now it runs on a very old HP laptop I don't think there's anything wrong Actually because it has 16 gig ram And it has a nice SSD disk And so on But it's still 12 to 15 years old Something So It's been around for a while But it solves the task So 15 years old With 16 gig ram It was a really good machine Well but the thing was that I got it from my employers at the time Which incidentally is the same Employer as David Has right now And I would The reason why It has so much ram It was because I was assigned to run a lab At the lab center That existed at the time That I needed to prepare And I would hold it Right after I came home from New York And I had never held the Lab before So I needed to sit on the plane And run virtual machines In Windows Server So then I needed a machine With a little more memory in So then I was sponsored So I got quite a lot of memory In it 16 gigs of ram It was extremely much then Like And then I had Double SD disks in it So then Then I did pretty well on it So that Right now it's like this It doesn't even have a battery So that the other day When my wife thought She was going to charge her phone So she pulled the cord from it And then I didn't have an agent anymore Or yes Until she started it up again But so But I think I need I need To get a little more organized On my Agent infrastructure at home I think But But otherwise so Yes but it's something like Like I Think But so the thing is Your wife won't pull out A power cord out of a Raspberry Pi If you hiding it somewhere Because she won't find it So it will survive Yes I mean We solve problems In this podcast Or how Apparently Yes What's on your list Anders I don't think I need More things But I'll be completely honest Okay Claude we wanted A permanent memory Yes I saw it That was great Yes Yes You can give poor Claude An obsidian vault I don't know He deserves it Maybe Think now he's listening Yes I just figured it out You start like this Delete game The desktop soon Yes Yes Yes I think Obsidian is a good thing Precisely because It will be a rather Like We talked about this before With the Agents will start talking A language we don't understand So the advantage of obsidian Is anyway That you get it in a format That's also available To us As people So Yes I thought So far Before we hang up So I'm thinking Tisa about a thing I'm thinking of having As a project This summer We're going to do the summer end program Next week But I thought If I tease about my now Because it has with what we talked about What we talked about today What to do I'm thinking of Setting up a new instance of Hermes On some machine And try to develop Some kind of Homework support For the kids So that When they come home With their pdf homework So they can just upload It in a Maybe a web interface Somewhere And then they have A chat interface And this agent Will understand What they have shortcomings And what they need to practice more And keep a little watch on them so to speak It will Probably Which project I'll talk more about it Next week When I've got Iterated on laps In my human brain And then Chat a little with my AI About which roads I'll take Because I don't want to Build anything Before I have a good plan Just like you Johan I'll let it go So you can Run with it If you want I did a similar thing On the way from the restaurant Where we ate Now before I connected my pod I prayed Because I've plugged in My agent In our personal Spring family Di scord server So it's there For my family To talk to And then I thought My children are a bit like this Half-loud On the multiplication table Hahaha We can probably solve that somehow I thought Yes Yes Yes Yes but how good We thank you Just as usual By telling us that we are on a small Podomit.se The podcast is There on There is And We Do you want to get in touch with us You can do it Either by email Or you can do it On social media But above all You can do it In Discord Because So far We haven't replaced ourselves There with some kind of agent I'm not saying That it won't happen But we'll tell you When it happens So that And besides So I honestly think You'll get better answers from it Than what you'll get from me Like We're transparent But in any case Yes but exactly But as I said Next week That's the last episode Before the summer break And then we'll In addition to discussing some News and stuff So we'll also have A little summer special We'll talk about things As you can do Oversummer For having fun And nice And then I also want to Thank Anders For wanting to stand up And we got to take his It's only time A little while today And learn some stuff About and other And then I asked If you want to reach Anders You thought it was great fun what he had to say And he talked about Karolinska And it was agents That he jailbroken To help Humanity here like If you want to get in touch With Anders Besides the podcast here How do you reach Anders then? Anders.bjarby.com Or anders.bjarby.com Can it? That sounds like a good plan With that we say thank you For today Have a good time until next week Bye Bye Bye