# AI Race: Superintelligence, Distribution, and Market Winners

**Podcast:** Tech and Tales
**Published:** 2026-03-21

## Transcript

Tech and Tails.
Herzlich willkommen zur 72.
Folge von Tech on Tales.
Heute hört ihr den Mitschnitt eines Panels von der Start Summit Conference in St.
Gallen, auf der ich gerade.
Der Titel des Panelties Race to Superintelligence.
Which AI will beat the rest.
And the panel had auf English statt gefunden, daher hoffe ich, dass ihr mich vergibt, dass es in der Originalsprache is.
But before we dazu kommen, möchte ich euch gerne einmal ein Überblick geben.
Und zwar ein Überblick über die verschiedenen KI-Strukturen, über die Modellarchitekturen, die es gibt, um später so ein bisschen besser zu verstehen, was wir da eigentlich diskutiert haben.
Und die Annahme, dass nur LLMs das Rennen entscheiden werden, is natürlich naheliegend, aber auch nicht unbedingt zu 100% war.
And um das zu verstehen, lasst uns einmal da ein bisschen tiefer reingehen.
Also KI is nämlich einfach nicht gleich KI.
Wenn wir heute über KI sprechen, meinen die meisten eigentlich eine ganz spezielle oder bestimmte Art, nämlich generative KI, also die Systeme, die Texte schreiben oder Bilder generieren oder Code generieren können im großen Ganzen.
Aber das ist eigentlich nur ein kleiner Teilbereich.
Denn wenn man KI-Modelle grob in drei categorien einteilt, dann gibt es erstens die klassischen Machine Learning Modelle.
Das sind so Entscheidungsbäume oder lineare Modelle.
Und das sind dann immer Wenn-Dann-Bäume, wenn man so möchte.
Und die sind super gut geeignet für bereits strukturierte Daten.
Also alles, was Supply Chain is und was relativ lange schon strukturiert ist.
Und sind natürlich super effizient, aber in der Komplexität sind sie sehr begrenzt.
Das heißt, es geht alles eigentlich nur linear und nicht, wenn sie Dinge erschaffen müssen oder selber vielleicht Entscheidungen treffen sollen.
Das ist mit denen in der Regel nicht möglich.
Dann zweitens gibt es Deep Learning-Modelle, das sind neuronale Netze.
Die funktionieren auf verschiedenen Dimensionen und sind dem Gehirn sehr ähnlich.
Das Gehirn dient auch für die neuronale Netze als Vorlage quasi und die sind super für Sprache, für Bilder und für Audio.
Und dann gibt es die generativen Modelle.
Das heißt, das sind so ein bisschen die Stars der letzten Jahre, die erzeugen neue Inhalte statt nur zu klassifizieren und effizienzen zu bringen.
Und eventually inhalte in allen Bereichen und multimodale Modal modell, also das heißt auch Bilder und Videos und so weiter.
Und in dieser dritten Kategory, da findet eigentlich in gewisser Weise schon das Wettrennen statt.
Also nicht nur LLMs, wie gesagt, sondern die gesamten generativen Modelle.
Und innerhalb dieser generativen KI, wenn wir uns dann angucken, was es da gibt, day gibt es wiederum auch drei Modellklassen.
So, das erste sind diese Sprachmodelle.
Das sind die Alleskönner, von denen ich eben sprach.
Das sind die Modelle, die mit Text arbeiten, Textwahrscheinlichkeiten mit Gewichtungen im Text.
Das heißt, die können aufgrund der einzelnen Wahrscheinlichkeiten der Wörter zueinander text generieren.
That's why they schreiben, sie analysieren, sie übersetzen anders programmieren.
And programmieren is momentan der größte anders use case.
And architektonisch basieren die all of the gleichen struktur or architektur.
This is the transformer-architektur.
And the central idea dahinter is that Modell eben nicht nur Wort für Wort ansieht, but the zusammenhänge im Kontext verstehen möchte.
And then the Wahrscheinlichkeiten errechnet.
And that's neue Dinge wie flüssige Gespräche, logische Schlussfolgerungen, or even complex problemlösing.
And that man LLMs, so large language models.
The twenty growth group are the diffusionsmodelle.
That's so the Künstler under den Modellen, when man so möchte.
When you have a build generier, then steckt dahinter meistens or hochwahrscheinlich das Diffusionsmodell.
And the funktionsweise is eighteenth faszinierend, denn man sieht am Anfang, also when man es langsam aufbauen lässt, dann seht man am Anfang einfach ein reines Rauschen, so ein bisschen wie ein stark verpixeltes Bild or week in the fernsehen up 23 Uhr, next, but we all are.
And then the dritte growth are the multimodal modell.
This is the next a build analysis, then a presentation, whereby I think that I do not relative rudimentary in my malta.
But you can use the Diagramme bespreching and so on.
And this is even the greatest spring that passified, then the zucchine is not the entire modellen, the LLMs, but even complex systems that all combinieren.
And we see also on the university, not in the moment in the LLM-Phase.
So Big Tech, they are language multimodal in any year.
And they bought systems that all made another combination.
So the frame is just architecture, so to say the one lets the Rennen gewinnt.
And that we can discuss it.
But before we get into this discussion, so I can get to what these models are, so to say, what are the central architecture, beziehungswork, what are the designents in the modern.
And when we can get monolithische models, that's a riesige Modelle, modell that all quasi in einem System vereinen.
And it has herausgestelled, that this tatsächlich die performantesten Modelle sind, auch for kleine Teilbereichs, wie zum Beispiel medicine or physics, but they are extrem teuer.
They are to control, it is also Guard Rates and Energie.
So dafür sind sie aber sehr leistungsfähig.
Also es kommt immer so ein bisschen darauf an, in welcher Richtung man gehen möchte, aber momentan is that the legal artist, what all verwenden.
This is the absolute standardansatz.
But the antworten.
So the zweite category in Architectural, and this is the mixture of experts modelling.
That he's strategic so much interessanter, weil this is not a gross model, but it's verschiedene kleine, specialisier expertens.
And it is went energie effizienter.
And this is meant to the riesige for the zucchini, that man of Teilbereiche quasi activities and not this riesenmonster in Bewegung setzen muss.
So and then the drift variation are the agent-system.
This is the next shit.
That's this is not really a modell, but this is a system of mehreren Modellen, die dann gemeinsam zusammenarbeiten.
So that heinwegen ein Modell plant, an anderes.
And the eighteentliche Differentiator ist das Post-Training.
Das heißt, um ein Post zum Post-Training, yeah.
So when the pre-Training fertig ist, kommt das Post-Training.
Dazu gehört das Fine-Tuning, dazu gehört die Feedback-Schleifen, und dazu gehört das Alignment.
Und dieses Post-Processing oder das Post-Training ist der absolut entscheidende Wettbewerbsvorteil.
Das heißt, wie gut die Firma im Post-Training sind, macht dann letztendlich den Wert aus oder die Performance des Modells.
Und um abzuschließen, es gibt noch zwei andere sehr strategische Aspekte in dem Ganzen.
Das sind einmal Open oder Closed Architekturen und das ist ein sehr entscheidender Aspekt.
Es gibt geschlossene Modelle, die sind sehr kontrolliert.
They sind meistens Leistungsstärker, aber am Ende eine Black Box.
Das heißt, sie sind nicht sehr wandelbar, beziehungsweise keiner weiß am Ende, wie sie funktionieren, wie die Weights gesetzt sind, wie das Post-Training war und dergleichen.
Oder es gibt eben offene Modelle.
Meta hat einige und die sind natürlich sehr flexibel, anpassbar für die jeweiligen Use Cases und haben sich natürlich in letzter Zeit schneller verbreitet, weil sie nichts kosten.
Und man sich schon als relativ einfacher User ein Modell runterladen kann, wenn man einen guten Server hat und das Ganze dann laufen lassen kann.
Dieser Konflikt erinnert stark an so frühere Tech-Diskussionen, so ein bisschen Windows versus Linux oder iOS versus Android.
Ich glaube, Windows hat den Kampf gewonnen, weil das Problem bei diesen Open versus Closed Source Architekturen ist immer das Funding.
But it's funny if that's long frisday sound, die hinter den offenen Architekturen stehen, haben meistens nicht hinreichend genug Kapital, um das eben we'll see the framework.
Yes, with your vision can you enter.
This had actually geopolitical that come.
So that's not the best technology, but the distribution has.
So the fading of intelligence in the stuff of the all things.
And should not this is the greatest effect.
So I wish I was in English.
Welcome to Start Summit.
I think this is our first panel this morning.
I hope you all had a good journey here.
And we're gonna discuss the race to superintelligence.
Which AI will beat the rest.
And I would like to start with Liz.
Elizabeth.
What's the one thing that everyone here needs to know about you that is relevant for today's discussion about the race to superintelligence?
Which AI will beat the others.
Hi, so first of all, good morning as well.
First panel, we are all waking up.
Until Lena upgraded herself and went to the NZ set.
So the one thing that would be relevant, I guess, would be my take on AGI or LLMs.
I personally feel or I'm in the opinion that LLMs will not bring us further or bring us forward that much longer.
So my take is that AGI will be based on different models and different architecture, and that LLMs will just be part of it.
I make my first machine learning experiences in 2011 or something when I was working with the team at Ladenseile, and for the last three years I've been holding these uh beyond the AI hype speeches uh every year and gotta do this again this May in Hamburg at the OMR festival.
Yeah, I think that's the most important uh stuff about me.
And regarding LIST statement, I think it doesn't matter that much if we reach AGI in three or five years, or if we do it with LLM or more world models, uh Mambo models, whatever.
I think LLMs already like enable very significant value creation.
Uh we just have to learn to adapt and to to implement it.
All right.
Let's start by setting the stage.
So we are talking about a race between AIs.
So we should determine where is the finish line.
So, Liz, how do we know who won the AI race in the end?
What is the finish line?
Is it superintelligence?
Is it AGI?
Is it something completely different?
I think there's uh one big aspect is the geopolitical aspect.
So at the end of the day, it's an arms race between China and the United States.
And I think if you look into the advancements of weapons, if you look into the advancements of robotics, then there are clear signs for you know, certain countries leading the race.
And once that gap becomes too big, you can tell that you know somebody won.
If you want to speak about winning, I think China's demonstration of their robotic skills for their new year celebrations was pretty impressive.
And that was one of the, I mean, it was also very, you know, demonstrating power in a way because they showed the advancements that had happened over the past 12 months, leading, you know, coming from the very basic robots to robots that can do backflips and become part of an autonomous army.
So I think if you look at it from that perspective, to me, it's kind of clear who's going to win or who's going to lead.
Because in my opinion, at least on that aspect, on the robotics side, China has has definitely an advancement.
Okay, so you connect the winning line or the finish line to something that we can actually see.
So robotics or or some kind of use case of AI.
Yes.
Apply basically embodied and applied AI.
Okay.
Pip, can you give us an overview?
Who's even competing in this race?
And who has an actual chance of coming in first?
Yeah, so I agree with the geopolitical views.
Uh from an economic point of view, I think there are two finish lines.
One is the companies that will be like create a sustainable business so that will economically survive.
And I think not everyone uh will.
Um, and the other thing is, I think the escape velocity of AI is when you reach that point where software is able to create software completely autonomous.
Because then you will have some exponential uh growth.
Then you just have to make sure you you get access to money, but that's gonna be much much easier than, and that would be the finish line.
So whichever company is closest to having software writing itself, um, I think that is escape velocity, if you if you will, uh, or the finish line, or very close to the finish line.
Who's in the race for that?
I think you would need to build kind of a metrics for enterprise B2B usage, professional usage, and free-for-all usage.
I think the free-for-all market, most people would argue OpenAI is leading it.
I beg to differ.
I think Google is leading it, because in 2024, OpenAI removed the authentication wall.
So you don't need to log in anymore into OpenAI.
That means if OpenAI says they have 900 million users, they don't have a clue how many users they have.
Because if I use OpenAI on my iPad, my phone and my desktop, I'm three users because I don't need to be logged in.
Or most users don't need to be logged in.
And if you take that definition, then Google with the AI mode and AI overviews actually has two billion users using their AI already every day.
And I think the free-for-all market will be with Google because they are the only company that can economically sustain it.
They don't have a churn problem, they don't have a retention problem.
Everyone is still using Google more and more every day.
So I think Google will win that because of their financial resources, their extremely sustainable core business with advertising, etc.
So that's the free-for-all layer.
The enterprise layer, I think the main combatants would be OpenAI and Entropic, obviously.
And I think Entropic is leading in enterprise adoption quite clearly.
If you look at data from uh fintechs, uh like uh billing data, credit card data, uh, even including buyer transfers, you see that OpenAI uh is not only leading but acceleration, uh accelerating in B2B adoption and enterprise adoption.
Um that might change every day, even like especially with the new um status of being deemed in um supply chain risk in the in the US, uh which they fight in court.
And I uh wish them the best of luck to to win that in court.
Uh Entropic is leading that.
And then in between is the professional use layers.
That's consumers and professionals like us paying for their own subscription.
And I think that's tied between Google and Tropic and OpenAI.
So everyone can still win that.
Uh that's a lot about taste.
Uh, that might even be a logopoly.
Uh so three three to four different players which are um well differentiated, uh, taste matters, UI matters, uh, and everyone will use the AI that fits his or her purposes most.
Sorry for the long answer.
But it's a very complicated question.
Yeah, Liz.
I may I ask a follow-up question.
Why why do you believe on the enterprise B2B sort of tier?
Why why does why is Google not leading then?
Because you said like it's either open AI or enthropic.
Because Entropic has focused on the most advanced use case, like with the best penetration or value creation, uh, which is coding.
So I I I think I would say 70 to 80 percent, that's a guesstimate, but of enterprise usage is for coding and a bit of customer service, maybe.
And Google's and OpenAI's like OpenAI Codex is getting quite good now and sees good adoption as well.
I think the CFO said they have quadrupled uh the users since the beginning of the year.
But as I would expect, 70 to 80% of the enterprise market that is sustainable, that's actually increasing revenue expansion is mostly coding.
Uh so that's the most valuable market right now, and that's clearly leading uh led by by Enthropic.
But you have you've mentioned all of the companies that probably most of us know right now.
What about the lab we are not thinking of?
What if the one AI developer that reaches super intelligence is not a developer that works for Anthropic or OpenAI or even Google, but for some unknown lab that we are not even thinking of.
How about that scenario?
I think that's a very possible scenario because the early the early disruptors have typically not been the companies taking the market or dominating the market.
So the first browser was something called Mosaic, which nobody knows about anymore.
The first social network was six degrees, which nobody knows.
One of the first search engines was um Archie, which nobody knows about anymore.
Or the companies uh sort of opening the markets for mobile phones were Motorola and Nokia, yeah.
The real first brick mobile phone was invented in 1983 when I was born actually.
So 43 years ago, and they're none of these companies have managed to keep their market dominance, and they've lost their market share because what they've done is they they cleared out, you know, they cleared out the market, they cleared out the innovation.
Something similar is happening right now with OpenAI.
They they've built the foundation and the foundational models, but they have also built a foundation for AI that now has been sort of copied from China.
Um the big Chinese models have all been trained on Western or American models as a as a foundation.
So I think it's very possible that somebody else will swoop in based on the existing systems.
However, the barrier of entry is very, very high as the yeah, as you know, like the the spend or the um capital you need to train is is extremely high.
Like the the US spent over 250 billion in AI software investments, just even letting out leaving out the hardware investment in uh GPUs.
So you need a you need a lot of capital.
However, if the architecture changes, and Deep Seek has has demonstrated that with the Deep Seek moment, that if you build up an architecture that is more efficient, that is um sort of crafty in a way, and uh can use resources in a in a more in a better way, then we might see like an innovation jump too.
And also the German startup Black Forest Labs, which is not 100% German, but German engineers sitting in uh close to Freiburg.
So I think there are there are possibilities, and again, history has shown that um the disruptors are not necessarily the winners.
We're gonna talk about the tech in a moment.
I would actually like to hear from our audience right now, which AI are you using?
Who who's using Google?
Can you raise your hand?
Interesting.
Okay.
Who's using open AI?
What about Claude Anthropek?
Okay, and everyone else, what are you using?
Anything else?
Ah.
Okay.
I have no idea how many people those were.
There was one Kimmy.
Uh I'd like to address.
There was one Kimmy lecture easily.
Okay.
Um I'd like to add to that.
Like I mostly agree with um Elizabeth's analysis, but we may not forget that it's not only about building the most capable model.
The core difference to the earlier years, the browser wars, etc., is that now like distribution is much more important.
Google isn't only winning because they have the best AI lab with deepMind, they also have distribution.
They have nine apps that have a billion or more users, uh, they have uh more data, they have their own hardware.
Um so even if you had like you had this little lab with the brilliant idea, you distill a big model and you advance it to build something like AGI, or you have a new approach, you will not have the access to capital, to hardware, uh, and not the distribution that the big tech players uh have.
So whoever, like if that would happen, one of the big companies would be fast enough to copy that and use their own distribution.
So while all of what Elizabeth said is true, I would expect in that case that one of the big players, maybe not OpenAI, but maybe Google, maybe even Apple or Facebook, would be very able to catch up very quick.
But Seren, you're still considering the models that we have today.
You were not thinking about this one super intelligent model that is smarter than everyone here in this room together, because kind of this model, I would suspect, would be able to figure out how to get the hardware, the distribution and the money.
That's like it's a very philosophical question.
The question is, why would that model still need distribution?
I mean, it could just build its own economy.
Why would it need why would it want to interact with humans at all?
So the that's a very hard question.
Like well, it's the topic of our of our panel.
We are racing to superintelligence.
And maybe we need to figure out how do we get there.
Actually, let's talk about the tech for a moment.
We have another few minutes.
Elizabeth, you've already mentioned uh large language models, but also new approaches to actually reaching the next level of AI capability and superintelligence.
How do you think will that happen?
Are the current models sufficient enough to supply that to reach this stage, or do we need completely new approaches?
I think the approach to define the world over text is very limiting.
Because uh, what makes us humans human is not necessarily text or the logic behind text.
And that's at the end what LLMs are.
I I think you can explain a lot, you can work, you can work very well with them, but to sort of reflect or to I mean the ultimate goal is to build intelligence that can replace us in a way, you know, that can replace us at work that works like humans.
But humans are a lot more than text, I think, and a lot more than words.
Humans are emotions, humans are the understanding of the physical world, humans are intuition, humans humans have a whole a whole lot more capabilities and and uh abilities also connected to the brain.
So there's uh Jan Lacan, the former head of AI from Meta, who is now building uh at uh world models at with Army Labs, um startup he founded in Paris.
And the interesting thing is that these models are based on on diverse data, so sensor data, visual data, vision models, basically, text data, of course.
And right now, uh a whole ecosystem is building around these word models.
So there are startups in India.
One is called Human Archive that is collecting physical data.
So people are running around with a camera around their neck like a necklace, and are recording their day-to-day lives, are now mapping that data and preparing that data for to be training data for the models.
I think it's also, by the way, it's a very lucrative business.
If I didn't have a job, I'd probably get into the data collection of uh world data.
Because these models will be so hungry for that data.
If you look at how how they vacuumed up the internet over the past 10 years, they will then they need to vacuum up physical data from autonomous driving, from anything they can get their hands on, videos and and whatnot.
So I think that's that's the alternative.
If that leads us to AGI, I'm not sure.
I personally don't think so.
I think it will be a part of the whole AGI concept, you know, like also LLMs will be part of it, but it will be sort of the pathway to it.
So it will be like uh, you know, one step ahead, so one next step in the evolution.
So I think that will be the next innovation, and then we will build on top of that.
So in my opinion, it will take uh a few more years before we are actually there, and somebody will win and we have ADI.
Do you agree?
What was the original question?
I agree with what you said, but the origin the original question was Um if the current models can actually add the current AI models can actually take us to a stage where we can say, okay, this is superintelligence, or do we need completely new approaches and technology?
Uh yeah.
So I would assume uh I mean nobody knows, but I think there's a high likelihood, if you think in scenarios, that LLMs will only bring us so far.
However, I think they will create enormous value even before AGI, as I said before.
But there might be new approaches like world models that are more likely to more efficiently lead us to AGI over time.
However, I think you will be able to use LLM to replace large part of the bullshit economy.
So jobs that like entry and you said the bullshit economy.
I love that.
What is the bullshit economy?
Like research work, creating power like uh economic research, market research, creating bullshit slides for meetings, typical consulting work, uh large parts of the financial industry, large parts of the legal industry.
I think that can be augmented and partly replaced with LLMs already.
However, I wouldn't look at too much at the current state, uh, but more try to anticipate what you can see already.
Like it's super hard to see what's gonna happen in three years.
It's quite clear to see what will happen in the next six to eighteen months, maybe.
So scaling still works, the next generation of processors or GPUs will be much better.
We will have much more and much more cleaner data in the future.
Um we get uh lots of reinforcement learning data.
So the models are only now or for the last 12 months starting to use your interaction data.
So every time you're using a model, you're training it.
You've you're fine-tuning or post-training it.
You're doing reinforcement learning, that's why you are asked to do the sums up, thumbs downs, or your next interaction with the model is continuing to train the model that's happening just now, basically.
Uh in the models we are using right now, that's almost not included yet.
So for the next six, six to twelve months, I still expect uh large linear improvements in capabilities, and that will get us very, very far in terms of replacing white, typically typical white collar uh work.
Yeah, I'd say that doesn't mean you will replace whole departments or something, but you have you will have one person with AI doing the work of three to four people within the next 18 to 24 months, I think.
In in companies that are open to it.
Obviously, there are like there's a lot of inertia in corporate uh in corporates in uh in government institutions, etc.
But startups will be able to do that work with like one person doing the work of four to five people.
I'm 100% sure.
I mean, that already takes us to our uh kind of last topic: the power that comes with this replacement, right?
Because if we zoom out of this race for a moment, um, because whoever wins, the next question is going to be: what do they win power over?
Let's say the best model is being provided by by Google or by OpenAI or even by Anthropic.
Considering that they will be able to replace a lot of these jobs, or even you already mentioned the military, like military decisions.
I mean, we have to start considering how much power we are giving one certain company or a couple of companies.
And I would like to hear your opinion on one current AI affair.
I think you already mentioned anthropic and how they drew a red line when it came to autonomous weapons and to mass surveillance.
What do you think?
Will this intermezzo with the White House reflect badly on anthropic?
Will they will they lose, or will that actually bring them out even stronger in the end?
So it will like we only have a little time left.
I'll try to be fast.
They they it will help them with the consumer market, so there's uh sympathy from the consumer market.
Uh it will impede their growth in the comp uh in the B2B market for a certain amount of time.
However, right now their model is so good and so popular that people people companies want to work with them, and they have the back of the big uh hyperscaler companies.
So Microsoft, Amazon, uh, and Google basically filed abicus uh briefs, so they were coming up in support of entropic because lots of their own uh hyperscaler capacity is used by entropic Models, so they they don't want uh entropic to to to not continue to grow anymore.
So I'm quite optimistic that entropic can win this.
And I think our time is up.
Thank you so much for showing up this morning.
Thank you two so much for coming.
It's been a pleasure.
And enjoy Start Summit.