Speed vs. Safety

Transcript

00:00:00Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:00:07Two technology-loving minds who don't just talk about artificial intelligence, but live it.

00:00:14Here you get clear classifications, real practical insights, and a fresh perspective on what is possible.

00:00:20Understandable, critical, and always with a wink.

00:00:24An invitation to think, to smile, and above all to join the conversation.

00:00:33A warm welcome to Think Different, Think AI.

00:00:37Jens and I have met up in person again.

00:00:40Unfortunately not in such a nice hotel lobby as last time.

00:00:43But nevertheless with a cold drink in hand after these hot days.

00:00:47And we realized for ourselves,

00:00:49that the last episode we recorded,

00:00:52the one you got to hear last week,

00:00:54is so current that we had to build on it once more. Jens, last time we talked about things

00:01:02like prompting, skills, loops, and in the episode before that we had at some point

00:01:08also talked about Fable 5, and now we said to ourselves, this all hangs together

00:01:12somehow. So what hasn't happened since then, or what has happened? Yes, of course we have this,

00:01:21We recorded the loop episode and I think it fits quite well that we now

00:01:24do this current addition to that loop episode, because the listener among you

00:01:29and the AI enthusiast, the way we are, will certainly have noticed it.

00:01:34Fable had, or rather Anthropic became really well known with Fable a few days ago, then

00:01:40it was shut down, we did a whole episode about it, you surely know

00:01:43that topic.

00:01:44Now this weekend again, such things apparently always happen

00:01:48on the weekend.

00:01:49And it has nothing to do with VW or BSF.

00:01:52No, but those are existing stories.

00:01:54Apparently, where things from America spill over here and get decided.

00:01:57Because there is another model,

00:02:00that has been unlocked again,

00:02:03and another model announced that they, their model,

00:02:06the one they had planned, the way they are, if I've got it right,

00:02:08the one from OpenAI, that would have been the 5.5, which was actually supposed to come out.

00:02:13No, 5.6, there is a 5.6.

00:02:15It's totally funny, right, because one of them gets a phone call,

00:02:18saying he should please hold his model back a bit and not

00:02:21release it just yet, along the lines of, you don't want something to happen to you, do you.

00:02:25And the other one released Fable 5.

00:02:28Unlike what we have known so far, the whole situation

00:02:32around AI, where things were actually just cheerfully posted out into the world,

00:02:36it seems we are now entering a new phase.

00:02:40Which of course speaks for the quality of the models.

00:02:42These models are being continuously developed further.

00:02:46That means they keep getting better, and accordingly of course also more dangerous, in that

00:02:53they surpass us in many capabilities, and above all now in capabilities

00:02:58that we used to show off, we thought we had programmed something well, we had

00:03:00programmed something securely back then.

00:03:02Now these things come around the corner and discover fairly quickly that this is not the case, yes,

00:03:06that we godlike programmers we used to be were not so error-free after all

00:03:10Error-free. Exactly, error-free, were not so error-free after all, and accordingly there are relatively many security holes somewhere that get cracked in no time at all by these new models.

00:03:22I think somehow the government also said, the American government, that their systems too were cracked by Mythos within a few hours.

00:03:29So that really is a scope that definitely calls for regulation.

00:03:34But on the other hand it also shows a bit how dependent we have become there by now.

00:03:40I mean, loyal listeners of our podcast will certainly have noticed as well.

00:03:45There are not only the American models, there are also Chinese models, for example.

00:03:51Yes, let's leave the French model with Ralf out of the explanation here for now

00:03:58and not have that be an insinuation or anything like that.

00:04:02But the Chinese models will probably also notice what's going on here, because one

00:04:07or the other model, at least that's what one has read or heard, came about through distillation

00:04:13of the big models.

00:04:15So you basically have to imagine it like this, that providers, let's say, ran a lot of requests

00:04:21and conversations with the US American models, automated, in order to

00:04:26basically rebuild and replicate these models.

00:04:29A kind of reverse engineering.

00:04:31They actually prompted, watched, so you say, I send 100,000 requests

00:04:35on the topic of how best to go fishing, then these models answer in 100,000

00:04:43variants, which are however very similar.

00:04:45So that in principle the right instructions come out, that I take a fishing rod with me or something

00:04:47else and the bait and whatever.

00:04:51And set off some dynamite, you've seen so many movies.

00:04:53And out of this, well this, with a certain probability, because we

00:04:56we all know the topic, that probability calculation happens in the background there, the token prediction.

00:05:01And to get that back out again, very, very many requests were sent to these American models.

00:05:06And that way you basically understood how this model works and then built it accordingly.

00:05:11There were stories that some of the Chinese models are set up like that.

00:05:15And what I also found quite funny is the other thing, that apparently there are even companies going to court,

00:05:23because immediately after Fable appeared they had recognized the potential and had already optimized their business processes

00:05:32for it and said, this is basically a model that is so good in

00:05:37text analysis, I think it was a law firm, was so good at text analysis that they

00:05:43didn't want to do without it anymore, switched everything over to it, and then it's

00:05:46of course doubly annoying when all of a sudden it's like, right, we're discontinuing

00:05:51the model, although to be fair one has to say the model was not officially discontinued

00:05:56by the US government. They only said, you have to make sure that only US American

00:06:00citizens use it. That is of course a bit difficult in times of Amazon Bedrock access,

00:06:05APKs and what have you. So they shut it down in general. But then there

00:06:09you stand, as a company. And that also closes the circle a bit back to this

00:06:13loop episode. If you always try to be right at the bleeding edge, then

00:06:18you can not only fall flat on your face because, well, you're going into new territory and cause

00:06:23either costs or you have to learn to deal with it.

00:06:26Never!

00:06:27It can even be that your business process comes to a standstill afterwards because the model got switched off.

00:06:32Another good point here, because I wasn't really aware of that at all, it only became

00:06:35clear to me, only dawned on me, as you were talking.

00:06:37Other model providers too, whether through distillation or through their own model development, will

00:06:41most likely reach the status that we currently have with the big

00:06:45models.

00:06:46nice, with how much deliberation you go into this and say the American

00:06:52government has, as always, thought it through very thoroughly. What I

00:06:57found quite funny about it: when Mythos appeared and it was still said that it is only

00:07:02available to a very limited group of users, we also reported on that

00:07:06in the podcast, it was also listed as a model at Amazon Bedrock in the EU zone, in

00:07:13Frankfurt, where I then also thought, hey, if it really is this ominous super-duper and

00:07:18I-break-out-of-everything model, whether the euro zone Frankfurt at Bedrock is really the

00:07:23best place. I don't know about that. We'll have to check in a moment whether it's still there. Yes,

00:07:29sorry. But do you still remember what this model is called that has now come out,

00:07:33the one that sort of claims it has Fable 5 capabilities, because it isn't a model,

00:07:40it's more of an orchestrator. Yes, yes, it's Japanese, but the name escapes me right now,

00:07:44I think something with Fuku, I'm afraid I'll have to look it up. Yes, yes, so honestly I

00:07:50have no idea what it's called either, maybe it'll come to us in a moment,

00:07:54we'll certainly keep talking. I also found that quite funny, how models suddenly

00:07:58sprout out of the ground, or rather methods sprout out of the ground, that claim of themselves

00:08:04to beat the benchmarks that these big models had presented for a few days, and from

00:08:14that side, it's coming to me, what's it called? Sakana, AI, Fuku. Yes, that's exactly why I know

00:08:21that I always really enjoy doing the podcast with you, because you're the educated one. Yes,

00:08:27but I'd like to come back once more to the topic I just briefly touched on with

00:08:31looping, because we did talk about loops a bit.

00:08:35Looping, yes, we do loops, then we have looping.

00:08:38Looping.

00:08:39Just remember that.

00:08:40Exactly.

00:08:41Yes, that of course has, as has just become clear, consequences that you now have to

00:08:45worry about, if we, the way we introduced the loops, are in a

00:08:49world where I can no longer be sure that a big model that is released

00:08:56out there is still on the market when I wake up in the morning.

00:09:00So this nice storyline, I'll set up a loop here and maybe hundreds of them,

00:09:05well, that's also what we always hear when we look at it, hardly any of the big models

00:09:10is even programmed by programmers anymore, that is only looped as well, so many

00:09:14of the big AI programmers and big names who sit in the model companies say, we

00:09:20don't do it ourselves at all anymore, I'm basically only there to build loops.

00:09:25So, but if the essential part of a loop then disappears, meaning the AI model in the background,

00:09:32well then your loops are simply worthless too, honestly,

00:09:35they don't get any further from there.

00:09:36So accordingly we have to, everyone has to think about that, I believe.

00:09:40And we have often talked about that, the topic of harness engineering,

00:09:44how do I basically build up my whole AI setup, what does my AI architecture actually look like?

00:09:49You have to think about that for maybe two or three milliseconds longer and not

00:09:54just prompt what you could do there, so that you basically build a system that

00:09:58is then also robust and can prevent or avoid such outages in an emergency

00:10:04or at least catch them in such a way that it doesn't lead to critical situations

00:10:08somewhere.

00:10:09That reminds me of another thing that can go badly wrong

00:10:14with loops.

00:10:15If you follow me, you'll have seen that I got forcibly blocked again recently, because all the tokens that Anthropic grants me in the big Max plan were used up.

00:10:27Well, no, but I was somehow banned for several days because I had used up my weekly limit.

00:10:34You really should watch out with these /workflow slash code things.

00:10:39That's the one thing, okay, fair point, when the models hit token limits, as long as you

00:10:47don't have pay per use, that's the one thing. The other thing that struck me, what came up in

00:10:52my attempts to work with loops.

00:10:57You give the system large amounts of data and say, run your loops over it.

00:11:02If you don't make sure from the outset that, for example, blocks of text are sufficiently

00:11:11large or small, whereby sufficiently large or small is deliberately chosen to be vague,

00:11:15because I don't know what large or small is, but you'll notice in a moment what

00:11:18I'm getting at.

00:11:19Because I told the thing, listen up, here are big books, chew your way

00:11:23through them and evaluate the following things for me, please do that with a goal loop,

00:11:28so keep going until the following questions are answered, and all of a sudden everything

00:11:32had aborted. Everything shot off into nirvana. It said on the screen in yellow letters.

00:11:37API limit reached, open parenthesis. This is not about the model limit, you have

00:11:44quite simply bombarded the API. We're not continuing. Right, so then you say to it,

00:11:49keep going. And then it says, oh yes, totally great. Listen, I'll continue, blah blah blah. It then

00:11:55worked again, errors again. At some point it noticed, oh listen, maybe I should reduce my

00:11:58queries so this doesn't run into the limit, and at the very, very, very end it says,

00:12:04you're done with everything. By the way, we had eight crashes. Should I repair

00:12:08what broke in the process, and you think to yourself, okay, the goal loop is broken,

00:12:14then you constantly have to say continue, and the result was as if I had written a normal

00:12:19prompt, everything I put in about verify yourself and make sure and these are

00:12:25the KPIs. It pushed all of that out over the airwaves during these abort moments.

00:12:30Yes. So I also thought, I don't even need an American government

00:12:34that shuts a model down. My own incompetence is enough as well, and with that I'm calling

00:12:38no longer for prompt engineering, but for the loop engineer. Yes, but the loop engineer,

00:12:45that is, I think, a special point, I think with this whole AI hype topic and

00:12:50the two of us are always right at the front too when it comes to throwing terms

00:12:53around, we'll keep coming around the corner with new topics. I think the

00:12:58experience of the last two or three years with these topics simply shows, we are all

00:13:04still early. All the things are emerging relatively new right now. That means there isn't yet

00:13:12the optimal loop engineering school that tells you, if you do it like this, then it

00:13:18will keep running.

00:13:19There's definitely an influencer who will sell you that as a course.

00:13:21Exactly, absolutely. And you can immediately make millions with that if you do it.

00:13:26Anyone who has ever tried to build a workflow with, let's say just two AI models in a row.

00:13:33will notice that they have to rework things again and again.

00:13:36There are always these places where we currently rework, where we, the famous human, always have to be in the loop,

00:13:42who can't just be on the loop yet and somehow sit on his hacienda

00:13:47and enjoy the sunshine and count the coyotes.

00:13:50No, we still always have to, after an indefinite time, we don't know exactly how

00:13:56long that is, that was also our Temporal UX episode that we did once.

00:13:59We also don't know one hundred percent exactly when we actually have to step in again,

00:14:02maybe a notification that is pleasant, that says the loop is finished, or else

00:14:07the way you just described it, that says, hm, there's a problem here.

00:14:10Everything's on fire here.

00:14:11Everything's on fire.

00:14:12Just briefly, briefly, briefly.

00:14:13Could you come over for a second, yes?

00:14:14So that means, yes, so if you want to do things right now, then you also have to

00:14:18be in something like 24/7 standby, if you were to let that loose on critical systems

00:14:24as well. If I have some kind of book project there, yes, something different, then

00:14:28it's not so bad if you restart the thing again on the weekend

00:14:31or something. But with all other topics we are, I think, at the point where you have to say,

00:14:38think about things a bit, look at things a bit from an IT perspective,

00:14:43what we said earlier, try to build robustly, maybe don't just let the loop

00:14:49run wild on things, but there too, cut the small part into smaller

00:14:53pieces, make it monitorable, have it made verifiable, so that you look at what can

00:14:59go wrong there.

00:15:00And I think that's a bit of what I'm taking away at the moment from this AI phase

00:15:04we're in, now without digging out some kind of Gartner hype cycle

00:15:09thing.

00:15:10But I think we're now a little bit...

00:15:11In the valley of tears?

00:15:12Yes, valley of tears, I don't even know.

00:15:13It doesn't feel like that.

00:15:14I've been through that a few times already.

00:15:15Both professionally and privately.

00:15:18I think I've been through that with prompt engineering and with loop engineering and with all the other

00:15:22things over and over again, and gone back up into the hype cycle and things like that.

00:15:25I think it's actually more that we are already on this path in the direction of the plateau

00:15:30of productivity.

00:15:31But things are simply missing.

00:15:33We talked about it recently.

00:15:35We're like in '97 or something.

00:15:36You know, the internet is there, things work, websites are generally there.

00:15:40But Google hasn't taken shape yet, big usability stories aren't

00:15:46worked out yet and we don't know yet where a navigation is best placed or something

00:15:49like that.

00:15:50We're still trying things out in many places, and that leads to cool creative results

00:15:55honestly.

00:15:56I mean, there's so much hot stuff being made out there with AI too, where things

00:16:00get rendered, animations get made, info websites get built, which back in

00:16:05our day we couldn't have imagined, at a speed like it's

00:16:08being done.

00:16:09lots of creative topics are being done right now. Loops are set up, marketing systems are

00:16:14automated, program forges are built. In many other places things already

00:16:19work, but it's not the case that you can copy it one to one. So we don't yet have

00:16:24this, there's an Amazon and I can build a Zalando out of it. So that moment

00:16:28isn't here yet, because you can't copy it over one to one like that yet. I think

00:16:31we're in an early phase, the way the internet actually is right now.

00:16:34You also notice that with another topic, namely today we had a bit of

00:16:39a closing note about models that can suddenly break away.

00:16:43Then we talked about skills and loops as a continuation, and you've certainly also heard

00:16:47the topic of Agent Harness in the other episodes, and at the end of the day it always

00:16:52seems to come down to the fact that today you find a lot of commodity harnesses,

00:16:58You have a ChatGPT, you have a Gemini, you have the next Siri AI, you have

00:17:05Cowork from Anthropic, but this whole topic of how do I serve the service worker,

00:17:10how do I serve the scientist, how do I serve the professional person

00:17:15who tries to create added value with these tools, without, well, do I install

00:17:21OpenAI today, do I take out the next subscription tomorrow, the next subscription, the next subscription, or do I install

00:17:26Open Code from GitHub here, or who knows, or do I maybe even sit down and

00:17:32build my own harness, and then it gets really funny when these environments keep

00:17:38changing. But on the other hand you build something that wants to absorb this constant change of

00:17:42the environment, wants to make sure that you can work productively, brings calm into the system,

00:17:48models interchangeable, skills still functional, standardized loops,

00:17:54without the user having to say so separately.

00:17:56How do I present it to my user?

00:17:58Someone recently said to me, hey, great Agent Harness you built there.

00:18:03But he always tells me that I have some kind of JSON app.

00:18:08Yes, that leaves me wondering too, what does that person want?

00:18:11Well okay, it just said there, JSON generated.

00:18:13Whereas, so, how do I talk to my user?

00:18:16That you also need a thick skin there, where I then also have the question,

00:18:20How do you run operations in a world where basically every three days new models, new techniques, new channels, who knows what, come out and get banned?

00:18:27Come out and get banned?

00:18:29Oh yes, if my harness ever gets banned, then it would be famous.

00:18:32So at that point, yes, for marketing that would be good.

00:18:36But I think Agent Harness would be another episode, we haven't announced episodes in a long time.

00:18:41Oh well, who knows, so feel free to...

00:18:43Yes, but in any case a harness engineering episode.

00:18:45A harness engineering episode, yes, I'm looking forward to it.

00:18:48Yes, but otherwise, as I said, with a look at the clock and the train breathing down your neck, the,

00:18:55so I just said that we're moving towards the plateau of productivity,

00:19:00that in this Gartner hype cycle, that's then the phase where everything

00:19:04works well, that you're out of this super hype phase, out of the valley of tears

00:19:08and that you now move on.

00:19:10I think this path is currently so fast with AI that you can already see this productivity plateau.

00:19:19But the path will still get a bit rocky.

00:19:21I think it will still have a few bumps along the way, some harnesses will break, some loops will, hopefully undetected, do the wrong loopings.

00:19:32And there, I think, a lot will still happen before we perhaps have something like we used to have,

00:19:39that we then have a stable internet that all people can use,

00:19:43that is then also available to the general public.

00:19:47That will take a while longer, I think, actually, until we have that consistently everywhere.

00:19:52But nevertheless I want to, as always, end on a positive note.

00:19:55I'm looking forward to it, because there too completely new things will emerge that we don't even know yet.

00:20:01So if you think now of the iPhone moment, which software emerged afterwards, which

00:20:07interaction possibilities emerged afterwards, things like WhatsApp and co., this arranging to meet

00:20:11and things like that.

00:20:12All of that only came about once this technology was there.

00:20:14And I think we're now still in a situation where a lot of new

00:20:18technology, a lot of new use cases are actually only going to come for us, ones

00:20:23that we're not even thinking about today, that simply, you know, software

00:20:27that I generate on the fly, because I don't have to buy a graphics program off the

00:20:31shelf anymore, but instead simply get the graphic that I need in that moment,

00:20:35generate it. You see more and more today that people say, oh, I built myself the PDF shrinker

00:20:41somehow, or did something else with PDFs, and built a Splunking

00:20:46and you read about things like that here and there. I think a lot, a lot of cool things will still happen.

00:20:49But it'll all definitely be awesome, and while you were talking I had a

00:20:54business idea. Oh, now I'm curious. Should we stop or do we want to go to the

00:20:57No, no, of course I want to share it with my listeners, that's also a bit, I don't know.

00:21:02Just imagine the topic really did pick up speed, that models are always only available in America first, or in China,

00:21:09because, I mean, there is such a thing as export control, right?

00:21:12Maybe in Europe too, yes, we're keeping our fingers crossed for everyone.

00:21:15So, now just imagine that this were to come.

00:21:17I have memories of very old times, well, only ever from what listeners say, right?

00:21:21Pirate Bay and the likes of Napster. I somehow have this strange feeling that at some point

00:21:26we'll see people who then live somehow in the countries in question, are citizens of those countries

00:21:30and put some kind of rack systems in the cupboard and basically route their personal,

00:21:36well, not exactly personal, but route them through their state-citizen access, whether

00:21:41that's then legal, probably not, because with that you'd be up against export controls,

00:21:44I could imagine that something like that will get built at some point, then we'll have to

00:21:47ask Thomas with his Ford GPT and so on.

00:21:49Maybe there's already something out there.

00:21:51Thomas, if you're hearing this, we need to talk.

00:21:53Then you basically smuggle the model locally onto your hard drive.

00:21:56I carried it on paper.

00:21:57Just like the KGB.

00:21:58A printed-out model, that'll definitely be funny.

00:22:01On that note, today was a colorful potpourri of the most varied things, Jens is now rushing

00:22:07off to his train, which thanks to its classic delay is on time,

00:22:12if he takes that long to get there.

00:22:13We wish you a good time and look forward to it.

00:22:16Guests again soon, and tune in again when it's called Think Different.

00:22:19Think AI. Ciao.

00:22:23Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:22:28Two technology-loving minds who don't just talk about artificial intelligence,

00:22:33but live it.

00:22:35Here you get clear classifications, real practical insights

00:22:39and a fresh perspective on what is possible.

00:22:41Understandable, critical, and always with a wink.

00:22:46to think, to smile, and above all to join the conversation.