Raising Europe with Tom Guinness

Francesco Pappone argues Europe risks sovereignty and economic stagnation without sovereign AI infrastructure and autonomous R&D.

1h 13m Listen to the original Episode page
Artificial IntelligenceEuropean EconomyGeopoliticsVenture CapitalScientific ResearchSemiconductorsStartup StrategyPolicy

Summary

Francesco Pappone, co-founder of Paradigma, argues that Europe risks losing personal liberties and economic independence by falling behind the US and China in frontier AI. He introduces Flywheel, a platform that automates scientific discovery by structuring research as directed cyclic graphs, aiming to compress discovery timelines from years to days. Pappone claims Paradigma can achieve frontier AI performance for tens to hundreds of millions of dollars by leveraging cost-efficient talent and culture rather than massive compute spend. He highlights Europe's compute shortage as a critical bottleneck, citing historical underinvestment in infrastructure similar to the internet and nuclear power. The episode concludes with a call for European policy changes to leverage assets like ASML and secure funding through programs like SPRIND to maintain geopolitical sovereignty.

Key takeaways

  1. 1
    Automating science is framed as expanding the total economic pie by creating novel technologies, potentially leading to non-linear economic effects.

    Unidentified speaker (speaker_0) · 32:51

Episode sections

  1. 0:00 – 1:03

    Europe's AI Risk and Paradigma's Mission

    Francesco Pappone argues that Europe risks losing personal liberties and economic independence by falling behind the US and China in frontier AI. He states that without sovereign AI infrastructure, Europe cannot defend itself or maintain its first-world status.

  2. 1:03 – 12:42

    Company Origins and Co-Founder Dynamics

    Pappone explains that Paradigma was founded to automate science by moving the bottleneck from human intelligence to compute. He details how he met co-founders Giulio Starace on Twitter and Emanuele Rodolà through academic events, leading to the company's formation after six months of collaboration.

  3. 12:42 – 16:43

    Fundraising and Early Challenges

    The team received inbound interest from VCs after posting on LinkedIn while in stealth, largely due to Starace's background at OpenAI. Pappone identifies communicating high-uncertainty, frontier problems to stakeholders as the hardest part of the startup journey, rather than fundraising or recruiting.

  4. 16:43 – 21:39

    Geopolitical Context and European Strategy

    Pappone rejects the notion that staying in Rome is due to comfort, citing a bet on Europe's need for sovereign AI capability. He argues that the post-WWII geopolitical stability is evaporating, and Europe's security and economic power now depend on mastering frontier tech like AI.

  5. 21:39 – 25:47

    Flywheel Platform and Scientific Bottlenecks

    Pappone describes current science as bottlenecked by human communication and implicit mental models. He introduces Flywheel, a platform that structures research as a directed cyclic graph, allowing hypotheses and experiments to be explicit, shared, and scaled by AI agents.

  6. 25:47 – 28:11

    Machine Learning Records and Validation

    The team has achieved five world records in machine learning speedruns, including the current record for the non-GPT speedrun and records in OpenAI's Parameter Golf. Pappone notes that the CEO can direct AI agents to achieve these records in his spare time.

  7. 28:11 – 32:05

    Flywheel vs. GitHub and Market Opportunity

    Pappone compares Flywheel to GitHub, calling it a platform for autonomous scientific discovery that expands the total economic pie by creating novel technologies and markets.

  8. 32:05 – 34:34

    Automating Science to Expand Economic Value

    Speaker_0 argues that current AI optimizes existing economic structures, whereas automating science and R&D expands the total economic pie by creating novel technologies and markets. This process could lead to non-linear economic effects similar to the Industrial Revolution.

  9. 34:34 – 36:51

    Compressing Scientific Discovery Cycles

    The speaker contrasts the 10-year timeline to commercialize the 2017 'Attention is all you need' paper with a future where AI platforms like Flywheel compress discovery into days. By making research machine-readable and visible, the platform allows for rapid scaling of promising ideas.

  10. 36:51 – 39:33

    Horizontal Intelligence vs. Human Silos

    Speaker_0 posits that AI intelligence is 'horizontal,' capable of spanning mathematics, biology, and chemistry simultaneously, unlike human experts who are siloed. This allows AI to produce a blueprint for turning problems into scientifically attackable tasks solvable via compute.

  11. 39:33 – 48:29

    Business Model: R&D as a Factory

    Paradigma targets corporate R&D rather than individual scientists, aiming to transform fragmented, high-risk R&D into a streamlined 'factory' process. The model involves a setup fee and a token-based pricing structure for inference and orchestration, allowing companies to retain IP.

  12. 48:29 – 51:24

    Cost Efficiency and the Frontier

    The speaker claims that reaching close-to-frontier AI performance can be achieved for tens to hundreds of millions of dollars, a fraction of the billions spent by major labs. Citing RC and DeepSeek R1, they argue that capital efficiency and culture, rather than just compute spend, are key to competing with US and Chinese labs.

  13. 51:24 – 57:05

    Demographics and Talent Plasticity

    Speaker_0 asserts that AI talent is becoming younger, with 19-year-olds offering greater 'plasticity' than experienced veterans. This demographic shift allows European teams to compete on cost and speed, as younger talent is abundant globally and less bound by traditional institutional structures.

  14. 57:05 – 1:12:43

    Investment Thesis and Valuation

    The speaker suggests investors should value Paradigma based on the 'model of the world' they are betting on—specifically, the future existence of autonomous scientific discovery—rather than current revenue. The goal is to prove the bet is right to attract capital, similar to OpenAI's early scaling loss bets.

Assets discussed

AssetKindStanceContext
Anthropicprivate companyneutralRight. So it's similar to what you're mentioning, Frontier like Anthropic or OpenAI.
OpenAIprivate companyneutralRight. So it's similar to what you're mentioning, Frontier like Anthropic or OpenAI.
DeepSeekprivate companyneutraldeep seed. Seek paper, the DeepSeek R1 paper, is January 2025.

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

Automating science and R&D is the process by which we make the economic pie larger, opening up non-linear effects similar to the Industrial Revolution.

Unidentified speaker (UNKNOWN) ·32:44

How to verify: Check economic literature on the Industrial Revolution's impact on GDP growth rates and compare with current AI productivity studies (e.g., NBER working papers on AI and productivity). For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

The 'Attention is all you need' paper took roughly 10 years to result in hundred-billion to trillion-dollar companies.

Unidentified speaker (speaker_0) ·34:34

How to verify: Verify the publication date of 'Attention Is All You Need' (2017) and the valuation milestones of companies like OpenAI, Anthropic, or Microsoft's AI investments. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Paradigma was founded in February 2026.

Unidentified speaker (speaker_0) ·42:52

How to verify: Check the company's incorporation records or press releases from February 2026. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

RC trained a set of models close to the frontier for under $20 million.

Unidentified speaker (speaker_0) ·50:43

How to verify: Search for RC's funding announcements or technical reports detailing their training costs and model performance benchmarks. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

DeepSeek R1 paper was published in January 2025.

Unidentified speaker (speaker_0) ·50:02

How to verify: Locate the DeepSeek R1 technical report or blog post dated January 2025. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Meta failed to build a frontier model with Llama due to culture, not capital.

Unidentified speaker (speaker_0) ·59:02

How to verify: Review Meta's Llama release history and public statements from Meta executives regarding the development challenges and team structure. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

xAI had trouble building Grok to a frontier level due to culture.

Unidentified speaker (speaker_0) ·59:02

How to verify: Review xAI's Grok release history and public statements from xAI executives regarding development challenges. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Paradigma is the only Italian team selected for the SPRIND Frontier AI Challenge.

Unidentified speaker (speaker_0) ·58:27

How to verify: Check the official SPRIND Frontier AI Challenge winner announcements or press releases. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Thomas Wolf, co-founder of Hugging Face, recently exited to NVIDIA.

Tom Guinness ·53:44

How to verify: Check SEC filings or press releases from NVIDIA and Hugging Face regarding the acquisition or exit. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Primary source

OpenAI solved a very large Millennium Prize problem.

Unidentified speaker (UNKNOWN) ·1:04:23

How to verify: Search for news reports or OpenAI announcements regarding the solution of a Millennium Prize problem (e.g., Navier-Stokes, Poincaré Conjecture) around the time of the podcast release (Nov 2026) or prior. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Paradigma is the only Italian team selected in the SPRIND Frontier AI Challenge.

Unidentified speaker (speaker_0) ·1:05:11

How to verify: Check SPRIND program announcements or press releases for the list of selected teams in the Frontier AI Challenge to confirm Paradigma's status and nationality. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

SPRIND Phase 3 unlocks €26.5 million in equity-free funding.

Unidentified speaker (speaker_0) ·1:05:43

How to verify: Review SPRIND program guidelines, official documentation, or press releases detailing the funding tiers and amounts for Phase 3 participants. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Europe's policymakers have historically underestimated the necessity of investing in the internet and nuclear power.

Unidentified speaker (speaker_0) ·1:08:03

How to verify: Research historical European policy debates and investment records regarding nuclear energy adoption and early internet infrastructure development to assess the accuracy of this characterization. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

ASML has a market cap of over €700 billion at the time of recording.

Tom Guinness ·1:10:05

How to verify: Check ASML Holding N.V. stock price and market capitalization data for the date of the podcast recording (prior to Nov 20, 2026) via financial data providers like Bloomberg or Reuters. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Accelera AI is producing frontier chips from Eindhoven.

Tom Guinness ·1:10:16

How to verify: Search for Accelera AI's product announcements, press releases, or technical papers confirming the development and location of their chip manufacturing or design operations. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Francesco Pappone's Twitter handle is TensorQT.

Unidentified speaker (speaker_0) ·1:11:29

How to verify: Search for the user profile 'TensorQT' on X (Twitter) and verify the bio links to Francesco Pappone or Paradigma. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Paradigma's headquarters are in Garbatella, Rome.

Unidentified speaker (speaker_0) ·1:11:29

How to verify: Check Paradigma's official website, LinkedIn page, or business registration documents for their registered office address. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the claim that OpenAI solved a Millennium Prize problem by checking OpenAI's official blog or peer-reviewed publications for the specific problem and solution date.
  • Confirm the current status and eligibility criteria for the SPRIND Frontier AI Challenge Phase 3 funding via the official SPRIND website or German government announcements.
  • Check Paradigma's LinkedIn page and GitHub repositories for evidence of the five claimed machine learning speedrun records and their verification sources.
  • Review ASML's latest quarterly financial filings to verify the market capitalization figure of over €700 billion mentioned in the episode.
  • Investigate the 'non-GPT speedrun' and 'OpenAI's Parameter Golf' records to understand the specific benchmarks and whether they represent general AI capability or narrow tasks.
  • Search for academic or industry reports on the cost of training frontier AI models to assess the validity of the 'tens to hundreds of millions' cost claim versus industry standards.
  • Verify the existence and current hiring trends of 'Accelera AI' in Eindhoven to assess the claim of emerging European chip innovation.
  • Check Francesco Pappone's X (Twitter) handle @TensorQT for additional technical claims or philosophical positions that may provide context to his investment thesis.
Full transcriptexpand

0:00Unidentified speaker (speaker_0): Unfortunately, today in Europe, if you go look at the next big thing that's happening, which is the intelligence revolution, as many call it,

0:08Unidentified speaker (UNKNOWN): we

0:08Unidentified speaker (speaker_0): are not positioned to be in the first world. I think there's a world in which we end up being losing our personal liberties because we cannot project the force to defend them. Because we end up being a nation that's simply incapable of that.

0:24Tom Guinness: So other nations will be providing the things we need. day to day to survive.

0:29Unidentified speaker (speaker_0): That's almost the best case.

0:30Unidentified speaker (UNKNOWN): Hi,

0:30Tom Guinness: I'm Tom Guinness, and welcome to the Raising Europe podcast. Here we go behind the scenes with the people building the future of European tech.

0:38Unidentified speaker (speaker_2): Today our guest is Francesco Paponi, the CEO and co-founder of Paradigma. In this episode, we discuss how to automate science, the mechanics behind autonomous research factories, and what it takes. to build sovereign frontier AI. Please enjoy this conversation with Francesco Pappone.

0:57Tom Guinness: You don't want to miss this.

1:02Unidentified speaker (UNKNOWN): My

1:03Unidentified speaker (speaker_0): name is Francesco Pappone. I am the CEO and co-founder of Paradigma. A Paradigma, we are working on automating science. And there is a very simple reason for which it's important. It's because the biggest problems, the large imagine, you know, going outside of the solar system, imagine solving nuclear fusion, imagine curing cancer. All those problems are big scientific problems with the bottleneck. The reason we're not getting there yet is we don't have enough smart people thinking about that.

1:39Unidentified speaker (UNKNOWN): them.

1:39Unidentified speaker (speaker_0): If you could double the amount of smart people, if you could triple the amount of smart people thinking about them, you'll get you would get there so much faster. And what we're doing, essentially in Paradigma, is moving this bottleneck from just people to compute. So that, you know, it's something that we can scale much better than amount of smart people that we have in the world.

1:59Tom Guinness: Right. So it's it's helping unlock the amount of smart people. in the world by giving them additional tooling via paradigm.

2:07Unidentified speaker (speaker_0): Exactly. The entire point is that then now people will be able to increase their output, not just based on how many they are, but also how much compute they have.

2:15Tom Guinness: Right. And so when we think about your story,

2:18Unidentified speaker (speaker_0): you

2:18Tom Guinness: know, you trained in physics at Bologna and mathematical engineering at Politecnico de Torino, and you are were also on a machine learning. PhD track. You know, what made you leave that path to find a company?

2:32Unidentified speaker (speaker_0): You know, it's pretty uh interesting because when you see my profile, when you see my trajectory in life, it might look almost like a zigzag, really. If you're looking closely, I did my bachelor's in physics and then move moved on to mathematical engineering, which is not too distant, but it's almost a little bit of a uh it goes in some specific direction. And then it it's usually I don't usually say it, but before my PhD, I actually did an MBA, which is something that I don't I think it shows up on the LinkedIn, but people usually miss. Reason for this Because ever since I was in my bachelor's, I was you know, mainly attracted by two things. On one side, the hardest problem I could find. So I went into physics. Of course, just like every other uh arrogant uh physics bachelor's student, I wanted to be the next big theoretical physicist. And but I then in hindsight, I realized that it was my attraction for it the most what I perceived to be. the most difficult problem. So that's one side of it, and I think that accompanied me in the meantime. And on the other side, uh I have always had somewhat of an entrepreneurial

3:48direction. And the first manifestation of it was I founded a non-profit when I was in my bachelor's. A

3:54Unidentified speaker (UNKNOWN): non-profit. A

3:55Unidentified speaker (speaker_0): nonprofit. And it the entire point of the nonprofit was to mitigate the brain drain from Italy. And it's still something that kind of uh is important to me, and I guess we'll discuss this quite a bit. Uh but uh there I already knew that I wanted to make something of my own. And of course, the MBA side was part of building the capability of one day building my own company. And so when when time came, when the time seemed right, my machine learning PhD seemed like a, let's say, a temporary uh activity that I would be doing to be at the frontier of the next large big problem, which was machine learning, and I think still is, of course. And then, of course, fortunately enough, I had the luck of meeting my co-founders and and we started a company.

4:45Tom Guinness: So, yeah, tell us about that. So, how did you know Paradigma come together?

4:49Unidentified speaker (speaker_0): You

4:49Tom Guinness: know, your co-founders you know, finding each other, you know, Giulio out of OpenAI, Emanuele as you know, a Sapienza professor. How did the three of you end up building this together?

5:02Unidentified speaker (UNKNOWN): So

5:03Unidentified speaker (speaker_0): this is a pretty funny story because I think it goes very much uh it runs very different from whatever you usually hear from startups. because me and Julio met on Twitter.

5:18Tom Guinness: On Twitter? Yes.

5:19Unidentified speaker (speaker_0): Was

5:20Unidentified speaker (UNKNOWN): it

5:20Tom Guinness: called Twitter or was it called Twitter?

5:21Unidentified speaker (UNKNOWN): It was X

5:22Unidentified speaker (speaker_0): already, but it's already we still call it Twitter now.

5:24Tom Guinness: Okay.

5:25Unidentified speaker (speaker_0): Out of habit, there's no really no no other reason. But basically we uh so I have this anonymous account, which is now not anonymous anymore. People know who I am on Twitter. Uh and I'm usually posting about technical stuff about AI and trying to be on also the kind of small bubble that exists on these on these websites, and it's really helpful and useful. And there was this other account whose name was Julio. Uh but actually wasn't at I don't I don't think it was the actual name of the account county yet, but now it is Julio, Giulio. and and we're interacting quite a bit in English. We didn't know. we're both Italian at all, you know, let alone in the same city. And and then I ended up organizing an AI event in Rome, which was the first AI tinkerers in Italy. And I ended up showing up. And I recognized him because when you apply, you have the the possibility of inserting your hand your hex handle. And so basically.

6:28being the detail oriented person that he is, he was the only one who put in the details and I recognized him.

6:33Tom Guinness: The only one.

6:34Unidentified speaker (speaker_0): The only one. And then, as it turns out, he came in and we started organizing events together. He really wanted to organize this kind of event as well. And then at the second iteration of the event, Emanuela Emanuele showed

6:47Unidentified speaker (UNKNOWN): up.

6:48Unidentified speaker (speaker_0): And before that, I only knew him by face. fame because I did my bachelor's thesis in one of the fields that he pioneered, which is geometric deep learning. And then he showed up and he told me, look, you should come do a PhD with me because we're really interested in the stuff kind of stuff that we work on. And I actually started my PhD with him. And then this ended up,

7:11Unidentified speaker (UNKNOWN): you

7:11Unidentified speaker (speaker_0): know, leading us to start the company and in just a I think six months later after this happened. Yeah.

7:18Tom Guinness: So a couple of meetups, and then, you know, your PhD starts. Yeah.

7:23Unidentified speaker (UNKNOWN): And

7:23Tom Guinness: then you decide, you know, now's the time to turn this into a company. I

7:28Unidentified speaker (speaker_0): mean, yeah, there is a kind of lost part to this, which is of course we started working together. That's the main thing that convinced me really. Because as I told you, I already had a few uh okay. occasions in which I co-founded something. Uh and

7:44Unidentified speaker (UNKNOWN): uh

7:44Unidentified speaker (speaker_0): we started going to hackathons with Julio, and me and Emanuele started working together on machine learning papers. And there was super clearly a very good connection. And immediately we we started feeling the need of something in immaterial material that wasn't named yet, of making of disclosure. this collaboration more than just an academic one. And that's kind of what led us to the start of the company.

8:09Tom Guinness: Hey everyone, a quick favor before we jump back in. A huge percentage of you who tune into the show regularly haven't actually hit the subscribe button yet. If you've ever gotten any value from Raising Europe, whether that's a key takeaway for your business or inspiration from the people building European tech, please take five seconds. to hit subscribe and drop a quick thought in the comment section below. It's completely free, it helps our channel massively, and it lets us keep bringing you the best founders and investors across Europe every week. Thanks so much for your support, and now back to the episode. Was there a specific moment that convinced you that, you know, automating science, you know, rather than another model or another app? layer, you know, this was the problem that was worth solving? Was there a moment where that became clear?

8:57Unidentified speaker (speaker_0): More than a moment, and I think it was a day.

9:01Unidentified speaker (UNKNOWN): A

9:01Tom Guinness: day.

9:02Unidentified speaker (speaker_0): A day, because alongside my PhD, I was still organizing events together with Julio Giulio and Emanuela Emanuele was participating as well. And we had the the luck to be uh the organize a side event in what is still today the largest Italian tech event, which is the Italian Tech Week. Now it has a different name, it's Wave. Uh and basically we were thinking amongst ourselves: what is the next big thing in AI? W So, and we also thought about this in a very European kind of perspective. seeing that there was not much movement in Europe. And we're like, okay, we cannot just chase the existing players. We need a bet on something that's coming. And so we grouped together uh friends from Frontier Labs. There were people from XDMind, X Meta, there was someone from Hugging Face and Julio and Manale were on the panel as well. And this we had everybody pitch what they thought the next big thing would be. And at the beginning, nothing was going in the necessarily in the

10:04direction of autonomous science. But the interesting side is the discussion converged to it. So we all started from the from positions and then we converged to this kind of we have to make autonomous or more, let's say, streamlined the process of producing novel knowledge, novel technologies instead of just utilizing the ones that we have very efficiently. And this became a consensus that was now running

10:32Unidentified speaker (UNKNOWN): uh,

10:33Unidentified speaker (speaker_0): you know, in our discussions. And that was really what kicked it off.

10:37Tom Guinness: Right. So it w it wasn't uh an aha moment. It was more of a a group consensus over a day.

10:43Unidentified speaker (speaker_0): Yep. It was kind of uh uh let's say growing into just like when you slowly understand. some theory or when you slowly understand some novel uh idea that you have, it was something that clicked over time. And then it has, of course, is still clicking, if you if you one might argue, right? Or in early stages of the company,

11:04Unidentified speaker (UNKNOWN): but

11:04Unidentified speaker (speaker_0): it did work like this.

11:06Tom Guinness: Right. And so why the name Paradigma?

11:08Unidentified speaker (speaker_0): Yeah. There's the there's the nice story, which I will tell you later, but the reality is it would just felt like it was a great name. It's a

11:16Unidentified speaker (UNKNOWN): cool name.

11:17Unidentified speaker (speaker_0): It's a cool name. I agree. And it wasn't necessarily tied to the what it actually ended up meaning. Um or maybe our collective unconscious knew this and blurted out the nice name, and then we rationalized it. Uh but there's a very nice, I think, explanation for this because the last time we changed our paradigm. for science was quite a few hundred years ago. And one key thing that we argue as a company, and we argue this with also Flywheel, which is our platform, which is online and people can use that, is that if we want to do this thing, which is moving the process of discovering things from just a human bottleneck to a compute bottleneck, we need to a rethink part, at least parts of the scientific paradigm. And so we need a new paradigma in Italian. And so that's the origin of the name and the explanation of it.

12:10Tom Guinness: You know, it's so clear with the name and what exactly it is as well. So you've kind of encapsulated that well within the name.

12:18Unidentified speaker (UNKNOWN): But

12:18Tom Guinness: you know, when you think about going from you know, research to founding a business, I know you've done your MBA, so you're kind of almost preparing yourself. for this before and you've co-founded things in the past. You know, what's been the hardest part of going from research and writing these academic papers to founding the business? You know, is it recruiting, is it funding, or is it just being believed by people in the market?

12:42Unidentified speaker (speaker_0): I think this is a very personal answer. I will not be speaking for many founders, I think, because we had, for example, we had a very good, very I think peculiar, let's say, um foundraising dynamic. Because of course Julia was leaving OpenAI and we had inbound from every VC you can think about. We never had to do any outreach anywhere.

13:07Unidentified speaker (UNKNOWN): Uh

13:08Tom Guinness: did he put it under his stealth on LinkedIn?

13:10Unidentified speaker (speaker_0): We put it under stealth. We did a post on LinkedIn and our inbox blew

13:13Unidentified speaker (UNKNOWN): up.

13:14Unidentified speaker (speaker_0): Um And I guess that's part of the strategy, but we didn't do it consciously. I think the hardest part is definitely communicating with on a very frontier problem in which there's a very high uncertainty, even in researchers, among researchers, clearly and effectively

13:38Unidentified speaker (UNKNOWN): to say

13:38Unidentified speaker (speaker_0): stakeholders, cool. which might not necessarily be ready to absorb. Not even a technical explanation, but just a sophisticated one. And so that's the thing we're finding the hardest among along, let's say, the bunch of hard things that you have during the when you when you start a company. Probably that's also because we are less used to struggling in that side. We are as researchers, you're actually learn quite a few things about companies, I think, because you learn to kill ideas. That's that's that's not obvious, but if you want to run a nice lab, you need to learn how to kill bad ideas. And and hiring is also something you do, especially in a field in which we still hire very strong technical talent that's still in the research kind of side. But um fundraising we had to learn. We kind of I think we were getting better at it. I think we still have some some work to do on the messaging.

14:36Tom Guinness: Right. And when I think about you know running an early stage startup, you know, from all the conversations I've had, it's it it now seems like you know, for the highest performing organizations, they're more like high high-performance performance sports teams or Olympic teams. for example, where you're attracting all this great talent, but you know, you need to make sure that you're doing the right things. You know, oftentimes Olympians are defined by what they don't do rather than what they do.

15:05Unidentified speaker (speaker_0): Yeah. When you have these kind of you know, intellectually gifted people, it's really their in their nature to explore and be attracted to kind of shiny ideas and shiny things.

15:19Unidentified speaker (UNKNOWN): And

15:19Unidentified speaker (speaker_0): oftentimes they're right. And those ideas are really exciting. But there can be just so many exciting ideas you can have in a company. And a key part of this is to for them to consider them, consider you at least in the ballpark Park of being an intellectual peer and to trust your judgment from your bird's eye view. and to develop the capability of making that judgment over time.

15:41Unidentified speaker (UNKNOWN): Uh,

15:42Unidentified speaker (speaker_0): that's really hard. And I I fully agree. I think that it's very similar to an Olympic team, in which every person has a very strong personality, and you also have to bend your goals alongside their strengths. Uh, as you bend And their strengths alongside the goals, those are things that co-evolved. You cannot think of either. as a fixed element of your business.

16:05Tom Guinness: And I think Jeff Bezos also has a saying for the amount of good ideas that he comes up with is that you can only give as many good ideas to the business as they're able to execute. Otherwise, it gets very confusing.

16:18Unidentified speaker (speaker_0): I fully agree on this.

16:20Tom Guinness: You've said that AI will split a new first world from the rest economically military. militarily, and politically. And that Europe only recently realized history hadn't ended. You know, why build this startup in Rome rather than in the US or UK? And what have the benefits been of building here?

16:43Unidentified speaker (speaker_0): You know, as soon as we started raising this question, it came up all the

16:48Unidentified speaker (UNKNOWN): time. Me

16:48Unidentified speaker (speaker_0): and Giulio, Julio, we all we both spoke. English very well, and we didn't have any trouble even culturally being aligned with uh many of the you know general Silicon Valley uh culture clusters because we're basically all day on the internet, so we know many of the people in Silicon Valley today.

17:08Unidentified speaker (UNKNOWN): Um

17:09Unidentified speaker (speaker_0): and so everybody was telling us, Why are you guys not going to a SF? stuff? What's going on? What you're your you guys. are are kind of shooting yourself in the foot by by staying in Rome.

17:21Unidentified speaker (UNKNOWN): Now,

17:21Unidentified speaker (speaker_0): of course, one might say, yes, you of course you guys are doing this, is because you just like to stay comfortable in the place that you are. Honestly, I think I have a pretty strong counterpoint to this, which is in Julia wasn't up in AI. I had myself a pretty cool offer from a big company uh to go and work. in pre-training, which is I really like, so it's not really about comfort. Uh but there's a second bet I think we're making here. There's a bet,

17:47Unidentified speaker (UNKNOWN): which

17:48Unidentified speaker (speaker_0): is the following. And I think you kind of introduced it in the words that you read that I wrote.

17:53Unidentified speaker (UNKNOWN): Um

17:55Unidentified speaker (speaker_0): it's very clear now, if you're a European, that we've been convinced historically, unfortunately, uh of a political reality, a geopolitical reality that is now evaporating right in front of our eyes. We're convinced of a kind of end of history after World War II, or a few more occasions after

18:16Unidentified speaker (UNKNOWN): that.

18:17Unidentified speaker (speaker_0): And in which really we were we had found our steady growth as a geopolitical entity, as Europe. We had found our place as a market in a sense. And that there were countries that were the first world and countries that were the developing countries or the third world.

18:39Unidentified speaker (UNKNOWN): And

18:39Unidentified speaker (speaker_0): that was kind of set in stone. Now, um, I would like to uh you know to take you

18:49Unidentified speaker (UNKNOWN): to

18:49Unidentified speaker (speaker_0): the reason for which those are countries that are developing countries is those countries missed an industrial revolution.

18:56Unidentified speaker (UNKNOWN): They

18:56Unidentified speaker (speaker_0): couldn't for some reason organized or they didn't have the resources, the leverage ongoing economic change. Unfortunately, today in Europe, if you go look at the next big thing that's happening, which is the intelligence revolution, as many call it, we are not positioned to be in the first world. We are not on the first world trajectory of this new age of economics. And many things that we think that we kind of assume are given are security as a as a as Europe, our ability to manufacture the best goods in the world and to consume them, those are dependent. on your economic capital, on your economic capability, and on your capability to deter an attack or to defend yourself. All of those things, of course, that's my prediction, but I think I will gladly defend it, is will increasingly be dependent on our ability to have frontier tech, a large chunk of which is

20:01AI infrastructure and. capability of training models. And we don't have it. And that puts us in a trajectory in which all those things that we give they have as a given are at risk of being jeopardized. And so we think this will only not only just be an opportunity cost, not only just be, you know, a nightmare in the future, it will be something that triggers a response. that triggers a necessity, a market for a company that can build frontier next generation technology in Europe. And that's why Paradigma is it exists. And that's also why it tackles a problem that see looks like it comes from the future.

20:42Unidentified speaker (speaker_2): Right.

20:43Tom Guinness: And I think, you know, here at Raising Europe, you know, we're all about you know helping founders build great businesses. here in Europe so we don't miss this next revolution. You know, we did so well in the industrial revolution, we really shouldn't let this one go by. Otherwise, the rest of the world might think of us as, you know, a museum state or a museum bunch of states that you go visit on holiday rather than leading the global conversation.

21:13Unidentified speaker (speaker_0): Yep, that's definitely your a big risk. And as long as we're here, uh we will die trying.

21:22Tom Guinness: Exactly. So when we think about you know what you're building at paradigma, you know, to help solve this problem, you know, what problem is flywheel solving? Yeah. You know, why is scientific research still bottlenecked when compute and AI capacity have grown so quickly?

21:39Unidentified speaker (speaker_0): Think about it this way. Today, the way science works is small groups of people work on their scientific problems and things they've been trained on for a long time. And it takes a long time for us to train a person to get to that level, to the frontier, if you will. And each of those people kind of has in their mind their own little map of what they're trying to do. They're, of course, everybody agrees kind of on a scientific paradigm and up on the scientific method. And each of those researchers across disciplines in their mind has:

22:21Unidentified speaker (UNKNOWN): if

22:21Unidentified speaker (speaker_0): I I have this hypothesis, I made this observation, now we'll do this bunch of hypotheses. to test these and these and these, because I have this idea in mind and this theory, this this approximate uh model of what the world could look like. And then if those are true, I do this, if those are false, I do this. And so each person in their mind has this tiny map, this tiny tree of things that could happen, of hypotheses and experiments and how they went. And then they wrap up their work, they run the experiments, and they write a paper. And so in the paper, they pick a few trajectories of this map that they built and they write it in the paper. It's kind of post-doc, right? Fan fiction of what really happened in in the science. And they and then they go to a conference, they talk about the paper, and then they talk to another person. just like we are doing. And they try and describe this map that is in their mind and pass it on to the other person that has their own map in their mind and have the two things coincide. And this is kind of how we build consensus in the world, right? Now, as you can see, this is very human bottlenecked. This is very human communication bottlenecked. If the entirety of our ability to progress in science is dependent on how fast can I get my little

23:39map and your little map degree. This is this doesn't allow me to leverage the technology that we now have. Now, the technology that we now have is instead allows us to know what problems are interesting to us and to direct compute to try and solve them and test out if they were solved. And so what we do with Flywheel is we make the tiny map that we have. each of us kind of usually as a researcher has in their mind explicit. So if you go look into Flywheel, we have that we organize research in terms of what we call a direct thecyclic graph, which is a very fancy way of saying it's a network in which every so-called node, so every point in this network, is connected to the others and represents an hypothesis with an experiment. And so now you have an explicit actual map of what experiments I tried, where they brought me, and what failed. And now I can instead of just sending you a small, let's say, um

24:46curated version of what happened, I can you can directly start from my graph. and if you have a new idea, you can branch off of it. And so this makes science much more collaborative, much more immediate, we think, but profoundly more uh capable of scaling with compute in the sense that now this shared digital structured uh memory, if you will, graph in Flywheel can be looked at at the same time by tens of thousands of agents that can be directed to attack the problem from all kinds of different directions. And so this is something we're now using to attack really difficult problem in machine problems in machine learning. We've uh managed to get five world records and very different machine learning challenges doing this. We are using this internally for our RND, Indie, and we're starting to serve companies with the uh with a mechanism that leverages this to expand and amplify their existing RD.

25:45Tom Guinness: And so what are these world records that you broke?

25:47Unidentified speaker (speaker_0): In machine learning, in this, there's this very, I think, healthy practice of defining some specific um, many call them speed speedruns, runs, in which basically you have a very clear metric about if a very clear problem. So, for example, the most famous one is the GP2, GPT2 uh uh the nano GPT speedrun, which is based on the GP2 GPT 2 architecture, a very old architecture of the current GPT models. And the entire point is that this neural network, and you know, all the LLMs that we see today are just big neural networks, uh trying to train it in as little time as possible. It's pretty simple speed run, if you will. It's very

26:30Unidentified speaker (UNKNOWN): easy

26:30Unidentified speaker (speaker_0): to understand. And of course, the entire world is working on trying to make this faster because if you can make it very, very fast, you can get

26:40Unidentified speaker (UNKNOWN): the

26:40Unidentified speaker (speaker_0): same amount of GPUs to train battle models. Because one of the key things is the more you train them, the better they are. And so you have to have a finite amount of time, you can just train them more and make them better.

26:49Unidentified speaker (UNKNOWN): Uh,

26:49Unidentified speaker (speaker_0): It's super complex. competitive, of course, as you may imagine, because it's the big thing in AI. And now, if you get a very good record in it, you might get hired by one of the big labs. And for example, in non-GPT speedrun, we have the current standing as of today world record. Uh so we made a quite a sophisticated change in one On a specific point of the architecture that gave us a very nice speed up. And we have,

27:17Unidentified speaker (UNKNOWN): of

27:17Unidentified speaker (speaker_0): course, we didn't do this as the founders. There's an entire team that's working on this alongside with the Flywheel. And there's two runs on two records on slow run, which is another similar challenge. There we had a record on the OpenAI parameter golf, which is a similar setup but held by OpenAI, and now it's closed. And then finally there's a second track of NanoGPT speedrun and optimization. And I got that one myself. Awesome.

27:44Tom Guinness: And

27:44Unidentified speaker (speaker_0): one of the cool things about Flywheel is that as a CEO that's doing a thousand things, you can still direct the agents to get over a record in your spare time.

27:53Tom Guinness: It almost sounds similar to how GitHub works for coding. Right? You know, you can have these massive code bases with all these different branches and you know, the ones that work get the pull requests and they get brought onto the main branch. Would you categorize it similarly, or is it something else?

28:11Unidentified speaker (speaker_0): Definitely. Definitely, Flywheel is, I think, going in this direction for now. And one of the ways we we talk about it in the very page of Flywheel is a git for science. And the reason we do is because if you think about it, there's a reason why. n scientists are not using GitHub. GitHub is a great tool that's designed for software engineers. And there's a it doesn't look like it at the, you know, from a very high level, but a bunch of the design choices that are perfectly good for production software that are made for Git and GitHub don't really work in journal science or are not ergonomic for enough for people to use

28:53Unidentified speaker (UNKNOWN): it.

28:54Unidentified speaker (speaker_0): The reason is because one of the reasons is because GitHub really wants to track a single canonical thing that's working. Right? Well, instead, if you think about what we were saying before, what scientists do is kind of live in this space of ideas. And the graph we build in Flywheel Flywall is supposed to mean. something of its own, even looking at the bad directions. Instead, in in GitHub, you're you're more so in the direction of, okay, we'll just cut out the branches, delete the branch that wasn't working. Right? Instead, oftentimes in science, exactly the direction that didn't look like it was working, that's really informative for you. And so that's a key distinction between the two. things. But of course, it's not all. Flywheel is for us the groundwork, it's the infrastructure upon that allows, that unlocks the ability to work in parallel and save and structure progress on big problems. And now the entire interesting part is making sure that there's an entire stack of models and model capability that is very good at discovering thing things per JAL, per VAT spent.

30:10Tom Guinness: Hey everyone, taking a quick break from today's episode.

30:13Unidentified speaker (UNKNOWN): If

30:13Tom Guinness: hosting Raising Europe has taught me one thing, it's that a founder's real job is in resource management, balancing your cash, your time, and your team's focus. That's exactly why we built Frontier Stack. So your business can grow faster. and spend smarter. You get access to the growth stacks that top founders across our ecosystem are already using to outpace the market, giving you preferential access to tier one tools and services like Airwallox's competitive FX rates, Whisperflow's productivity suite, and Velocity's incredible paid advertising engine. Stop guessing and grow faster. Head over to getfrontierstack.com right now. Join the wait list and lock in your priority access pass so you can grow faster and spend smarter. That's getfrontierstack.com. Right. And so speaking about spend, you

31:06Unidentified speaker (speaker_0): know,

31:06Tom Guinness: how big is the opportunity if you automate the path from you know hypothesis to experimental validation? you know, what's the addressable spend there?

31:17Unidentified speaker (speaker_0): So I think there's two answers to this. Uh let me be very quick and pragmatic so that people listening will be somewhat content with this. If you just look today at the uh total spend in RD in the world and just look at that is that is computationally checkable. So don't even think of uh of you know labs that are that are physical in some way the spend is 1.3 trillion

31:46Unidentified speaker (UNKNOWN): uh

31:47Unidentified speaker (speaker_0): and that's today and that's already a very large market of its uh of its own but this is not arguably and i i know this is gonna be somewhat controversial

31:57Unidentified speaker (UNKNOWN): uh

31:57Unidentified speaker (speaker_0): but this is not the total time of this

32:01Unidentified speaker (UNKNOWN): uh

32:01Unidentified speaker (speaker_0): let me let me walk you through

32:03Unidentified speaker (UNKNOWN): it. Um think

32:05Unidentified speaker (speaker_0): of it this way: we're using the eye today

32:08Unidentified speaker (UNKNOWN): to

32:09Unidentified speaker (speaker_0): extract a lot of value, economic value, from existing economic structure.

32:15Unidentified speaker (UNKNOWN): Right?

32:15Unidentified speaker (speaker_0): We're having it write emails faster and make processes faster. But that's somewhat of a pie that's not I'm not saying fixed, but has hasn't is growing predictably. over the years, right? It's it's an existing economic value. Now, if you go back and look how how did humans even increase the size of the pie in the first place? How did that happen? Well, they invented new technology. They did science and research that brought up novel technologies that increased the total value of

32:44Unidentified speaker (UNKNOWN): the

32:45Unidentified speaker (speaker_0): economy because you are now capable of providing novel services and consumers are now capable of consuming different things. And so, really, what we're automating when we're automating science and RD is the process by which we make the pie larger. And that opens up the possibility of seeing very nonlinear effects that look like one day an industrial revolution a day. So think of the total economic value that existed before and after in the industrial revolution. day and night. Imagine this being a continuous process that you see every day because you're now creating novel markets at it very quick because you're discovering novel technologies very rapidly. This is the kind of uh large economic bet that at Paradigma we're making. And

33:38Unidentified speaker (UNKNOWN): we

33:38Unidentified speaker (speaker_0): think that it's going to actually be the such a large, it doesn't mean we're gonna do it, but we envision entire economies, entire nations kind of mining, uh extracting technological results and monetizing them almost similarly to what they do today with natural resources. And it may be for some uh countries and or nations, it may be their main way of sustaining themselves one day.

34:09Tom Guinness: Thinking about countries that operate purely on intellectual capital rather than natural resources. You know, we're already seeing countries like that do this. However, what you're saying is that They'll they'll be able to extract huge amounts of value every single day just by the pure velocity. of scientific discoveries that are made every single day.

34:31Unidentified speaker (speaker_0): You can think about it in this way. Um imagine that we had the you know, attention is all you need paper, is the 2017 paper that kind of gave birth to the mod modern era of transformers and scaling. Uh you know, it took us 10 years roughly to make that into a hundred billion. a trillion dollar company. Companies that we see today, right?

34:54Unidentified speaker (UNKNOWN): As

34:54Unidentified speaker (speaker_0): a society, I mean, of course.

34:57Unidentified speaker (UNKNOWN): Um

34:58Unidentified speaker (speaker_0): one day, that entire process might happen in a day. Because the reason it took so long is because we had to have people discuss the idea, and we have we had to have them uh kind of reproduce the results and slowly build from there.

35:14Unidentified speaker (UNKNOWN): Um

35:15Unidentified speaker (speaker_0): Imagine that instead it happened on flywheel for a second. And one day something like this happens on flywheel, then you immediately see that there's this big bubble that's rapidly growing because that result is very interesting and people have an incentive to work on it because it's very interesting, they can build on top of it, it's super cool. And now we have a very strong signal that something is brewing. And now

35:38Unidentified speaker (UNKNOWN): and

35:39Unidentified speaker (speaker_0): now, well, since everybody can connect to flywheel and everybody can see it, they can be like, all right, this looks like a great idea. Let's scale it up, let's put some some some value beneath it. And so the entire process before it took so long because there was so much communication going on and so many learnings to have in the meantime, that can be compressed so

35:59Unidentified speaker (UNKNOWN): much.

36:00Tom Guinness: Right. So if your kind of vision for the future. happens, this would be an absolute paradigm shift in value creation on a global scale.

36:09Unidentified speaker (speaker_0): Definitely. And I think that the current trajectory for AI

36:14Unidentified speaker (UNKNOWN): uh

36:16Unidentified speaker (speaker_0): is more likely to represent something like this. Of course I have a bias saying this. Take it with a grain of salt. Uh, But I seriously think this is pointing in that direction. rather than a plateau.

36:29Tom Guinness: In this, let's say, you know, AI automating science field, you know, if Paradigma wins this category,

36:39Unidentified speaker (speaker_0): yeah,

36:39Tom Guinness: what does it look like in a decade? Is this the infrastructure layer under all scientific discovery, or does it stay focused on a few domains?

36:49Unidentified speaker (speaker_0): No, I think. one of the key reasons for reasons for which we have to rethink part of the scientific progress and process

36:58Unidentified speaker (UNKNOWN): in

36:58Unidentified speaker (speaker_0): the way we organize it usually, in the social contracts around it, is because the intelligence that's going to be insisting on the problems is going to be a very different intelligence from the one we see today. Think about it this way: We operate in kind of siloed fields. medicine, biology, chemistry, physics, mathematics, because individual humans have so much bandwidth. Right? I can only be there's just so many people that can be both a great mathematician and a great biologist, they're very unlikely to be also great physicists or a great mathematical a great chemist, right? The intelligence that we're building, at least in the current shape. and I think we will not necessarily deviate very rapidly from there, is instead incredibly horizontal. Right? Current frontier models are great mathematicians, great chemists, great biologists, uh and and they have a set of different disadvantages with respect to human capabilities still that there were are those are quickly decreasing.

38:03But crucially, they don't have this necessity of really separating things as much. So I think that a successful paradigma in the future will produce a blueprint

38:17Unidentified speaker (UNKNOWN): to

38:17Unidentified speaker (speaker_0): turn a problem into something that's scientifically attackable. That you can resolve using compute. And so all things that today we use science on, all problems that we really want to solve, those are really things you use science on, will be attacked using flywheel and will be slowly solved using flywheel in this scenario, of course. And the key difference, the key differentiator now will be now we've built an entire stack. Today we have an extremely optimized stack for serving models that are great assistance. assistants, that write emails very well, that are served billions of people at the same time. And that's an entire optimization layer. The key element will be producing the entire stack instead in a way that is fully optimized to produce more discoveries per VAP, per token, per unit of compute. That's the new challenge and that's the thing that also Power PyroTechMot Techno will focus on.

39:15Tom Guinness: So clearly, the problem that you're trying to solve is enormous, right? It's trying to streamline and automate scientific research on a scale that's never been done before and probably has never been thought could be possible before. What are the underlying economics behind this?

39:33Unidentified speaker (UNKNOWN): Yep.

39:33Tom Guinness: You know, how much will it cost to do this on a you know, a compute, a resource? or resources basis, you know, on the cost side, but also for the customer, you know, how do you see that business model evolving over time?

39:46Unidentified speaker (speaker_0): There is, I think, two key elements here. One is, I will talk about this later, uh reaching the stage of having the capabilities of building the models that you can actually serve and and be content with on a on, you know, you can you have that you have trained on. your own.

40:03Tom Guinness: Yeah.

40:04Unidentified speaker (speaker_0): But in second, there's the kind of commercial outcome and kind of the business model that fortunately enough, we are already uh in the process of of uh putting in in place. And just to make it clear, our main commercial target I don't are not individual scientists, in the sense that we don't want to make it so that the flywheel that today you can see in online it is a product is seen as a product by the individual scientists that make up, you know, for example, individual academic labs. Uh I think that the way that we want them to see this is more of a manifesto of introducing this novel pattern for producing uh research autonomously. And of course, we're not in the business of not making money. There is a reason for this. And the reason is that we think that there is an enormous opportunity of applying this process in the

41:08RD that companies do and expanding even what we think of as RD. What I mean by this is today in a company, RD kind of looks like pre-industrialization, if you will. It's a very fragmented process in many situations. And it's oftentimes the riskiest place you can put your money. But also, if you think about it, the most valuable companies in the world today. they are the byproduct of an early RD bet. So it's kind of this very high ROI, very high risk way of allocating money. Now, one of the ways you can think of Flywheel, and in general, the entire stack for autonomous research that goes from inference, private inference that we manage, uh to collecting existing research and expanding it with Flywheel to providing our own models and providing our own orchestration layer

42:08is turning that unstructured process into a factory of RD, into something you can pour resources in and get your scientists to control and allow them to think about the ideas and not think about the uh necessary. grunt work that takes up 80% of your job as an RD scientist in a company and give you the same amount of results and more on a much lower uncertainty. Because now you have machines and it's a streamlined process that you can overhaul, that you can audit, that you can see each decision of. So this is the main thing that we're currently working on in Paradigma. Force Paradigma was born in February of this year, and now we're in September. And so we're in early implementation of this technique. But I think one of the coolest things for us is that we are the first customer of this because we are an RD heavy company. And part of the successes that we are producing in machine learning and I in a larger and larger chunk of those are a direct consequence of us applying our own product to ourselves.

43:20Uh and you will see it soon enough. We will be releasing our first model. Uh, and that model itself is also a large chunk of it. You know, we all know about the the cost. I I'll go on talking about this now. On how much does it cost to build a model? Uh it's something that we strongly decreased uh by using our own tool uh to automate part of the research.

43:45Tom Guinness: Right, by dog fooding. Yeah,

43:46Unidentified speaker (speaker_0): dog fooding, that's why we call it in the in the field.

43:49Unidentified speaker (UNKNOWN): I learned

43:49Unidentified speaker (speaker_0): I learned this the

43:50Unidentified speaker (UNKNOWN): the term after

43:51Unidentified speaker (speaker_0): becoming a founder.

43:52Unidentified speaker (UNKNOWN): Yeah.

43:52Tom Guinness: Yeah. Eating your own dog food. But uh yeah, it's amazing. And then also you can be your own best customer.

43:57Unidentified speaker (UNKNOWN): But

43:58Tom Guinness: when it comes to the economics of you know third party customers, you know, how do you see them paying for this? Yes. How you know once you're able to clearly articulate the value that you bring to

44:09Unidentified speaker (UNKNOWN): them,

44:09Tom Guinness: how does that translate into the business model there?

44:12Unidentified speaker (UNKNOWN): Yes.

44:12Tom Guinness: And then yes, let's dive into the cost side of what it will take to bring this to reality.

44:16Unidentified speaker (speaker_0): Yeah of course what this looks like is still an initialization and this is going to No change in time inevitably. But we see a very strong argument for making this a token based uh solution because what this looks like for a company, let's say we're talking about a company that's doing this. Maybe you're a company that has trouble retaining researchers. And so every time someone leaves, you lose half the information because it's all in their head. Or you feel like you've found a very strong direction and now you want to capitalize it on it so much and you want to actually accelerate, but you have just so many people, and hiring someone takes a long, so such a long time, or you cannot get the best talent in the world.

44:57Unidentified speaker (UNKNOWN): Uh,

44:57Unidentified speaker (speaker_0): what you can do is what we will do with you is provide you with an array of possibilities on hosting your own models. Uh, and but by these I'm by this, I mean So we select with you the models that we will be using 24-7 to attack the scientific problems, or uh

45:15Unidentified speaker (UNKNOWN): you

45:15Unidentified speaker (speaker_0): know, having it hosted in a cluster that's somewhere, or having it an inference provider that's of course secure, and providing fraud security for your own IP, of course, that you should be very happy to keep it for yourself. Uh and the we set up an entire orchestration layer that your uh researchers can utilize to direct the models and the compute uh towards some problems. And we what your research will look like will instead of being a bunch of documents all over the place, it will be a structured graph that I talked about and that I suggest you go take a look on the website that allows anybody to pick up from the latest frontier, the latest thing that we discovered Covered or the go look at what somebody else were was working

46:00Unidentified speaker (UNKNOWN): on.

46:00Unidentified speaker (speaker_0): And using the models that we're running for you. uh at inference, we are will be 24/7 uh attacking the problems for the directions you think are most interesting that your scientists are directing the agents to work on. And this there is a setup fee for it for this size, so there's a fixed cost at the beginning, and there the rest is depending on your spend. So basically, we of course. will have a price per token that we can calculate. And there will be a you will will negotiate this beforehand. And there will be

46:36Unidentified speaker (UNKNOWN): you

46:37Unidentified speaker (speaker_0): will be able to increase or decrease the number of tokens of the number of machines that you're serving at any time. And hopefully you're after a while you'll uh quickly grow to want to increase them because you will see that there is a very high value in user. using all of this intelligence uh to attack your most important problems.

46:55Tom Guinness: Right. And so when we think about, you know, the collective use of paradigma uh combined with the economics of it, how big do you see this being in five, ten years' time? If all goes well?

47:07Unidentified speaker (UNKNOWN): Yes,

47:08Unidentified speaker (speaker_0): if all goes well, I think uh we can be at the level of a frontier company because we think this is the largest market opportunity that we will see in the next ten years.

47:19Tom Guinness: Right. So it's similar to what you're mentioning, Frontier like Anthropic or OpenAI.

47:24Unidentified speaker (UNKNOWN): I

47:24Unidentified speaker (speaker_0): think it has the potential to reach that market side. So of course we're also competing with those companies and we should. We part of the reason is we cannot leave the entire pie to Chinese or American players. And in the meantime, the entire market of this will flourish. One of the cool things about science is that differently from software and other stuff, the more stuff you discover, the more there is to discover. Because every new discovery kind of carves its own little piece of of light in in the darkness and and allows you to discover more stuff.

47:55Tom Guinness: And

47:55Unidentified speaker (speaker_0): so I believe this can, if everything goes well, be in the in the direction of hundreds of billions of of dollars a year and and even more than

48:05Unidentified speaker (UNKNOWN): that.

48:05Tom Guinness: When you when you think about competing with the likes of anthropic, OpenAI, you know, these real frontier labs, what kind of capital would it take and what kind of resources would it take to build out the you know the required infrastructure to make this really work? You know, we see uh other frontier labs raise billions before a a model is even produced. You know, what are your thoughts on that?

48:29Unidentified speaker (speaker_0): Yes. Now it takes a lot of capital, as we all see. And I think there is a very good argument that

48:40Unidentified speaker (UNKNOWN): you

48:40Unidentified speaker (speaker_0): know, reaching or reaching close to the frontier, it's still severely, let's say, over overpriced. What I mean by this is fortunately enough, once there the real difficult side of this job, of working in AI and trying to reach the frontier, is breaking through and being somewhat of a step above everybody else. This is something that happens from time to time. Uh and some sometimes happens at Anthropics, Anthropic, sometimes happens at OpenAI. As you can see, it's very quick for other companies that have built the know-how and have built the machine to produce the models to kind of catch up. And a key thing for this step, and of course, is not just capital and compute, it's also culture. This is something maybe we'll discuss later. But there's a few key elements for which you can spend a fraction of what those companies spend today to reach very close to the frontier and then

49:41become a threat in the sense in the positive sense, become a competitor in that side. And I think that's something that we've seen with several Chinese labs now. There's been two years ago, if you go back two years,

49:53Unidentified speaker (UNKNOWN): I

49:53Unidentified speaker (speaker_0): think there was a most one A lab that was going in the direction of computing. But if you think about it, the

50:01Unidentified speaker (UNKNOWN): big

50:02Unidentified speaker (speaker_0): deep seed. Seek paper, the DeepSeek R1 paper, is January 2025. So it's not even it's a little more than a year. And basically, the same effect is still possible, by which a small fraction of the capital can take you close to the frontier. Now the hard part is breaking. This you uh we think uh looks like you can full full stack train a model that goes close to frontier in the order of magnitude of the tens of millions, pushing to the hundreds of millions of dollars, which in terms of frontier AI is very is peanuts in a sense. And there is a track record of this. There's uh several companies, among which, for example, RC, which is a company we know we're friends with. uh, that did this with under 20 million. They trained a whole set of models that were close to the frontier uh in with with just that that amount of money. And

51:04breaking through will be the key cost. And a key cost is also that your choice of where the frontier is. And that depends much more on the culture. and the people that you have. And we think we have a very unfair advantage on this. There's also another reason we're in Europe is the real talent that takes you to the frontier is younger and younger over time. It's actually a demographic change. Simply, this is a field that goes so fast that it rewards plasticity instead of experience.

51:36Unidentified speaker (UNKNOWN): And

51:36Unidentified speaker (speaker_0): so what happens is if you build a machine that is capable of, and by machine I mean a company, of carrying over what it learns. And what that's what we do with Flywheel. And being very good at at using the compute you have, you can be very, very efficient. And and the entire point of you know, the the talented talent of demographic being younger is night nine incredibly talented 19-year-olds. 19 year olds don't spawn in SF. They're born everywhere. And we have great examples of them upstairs. Uh and we'll have more. And this is a key thing that will allow us to be cost efficient in breaking through the frontier as well.

52:17Tom Guinness: Amazing. And I think speaking to other founders here in Italy as well, they're starting to see this

52:22Unidentified speaker (UNKNOWN): you

52:23Tom Guinness: know, plasticity being super important for you know hiring your CTO or you're finding your CTO, is that you know this engineering talent is becoming younger and younger because they're more flexible on how they're able to work and use AI tooling. Right. And so you know, when we think about the amount of resources required to make your vision a reality, I know that you know Deep DeepSeek Seek created their model on you know a fraction of the cost of what OpenAI and Anthropic have used. Although there may have been reports that they'd been underestimating the cost because of outsourced use of GPUs, things like that. What kind of resources will you need to bring it from You know, today to you know something that is mass-adopted mass adopted and can be used by? everyone in the world. So when I think about the resources you have today, you know, you raised a pre-seed brand. So huge congrats to you. Thank you. You know, from Emergence Capital,

53:24Partec, Italian Founders Fund, Vento, and also three notable um angel investors, including Michael and Alex Bronstein, who are both prominent machine learning uh researchers, and then also. Thomas Wolfe, you know, the co-founder of Hugging Face, who, you know, they've got a massive exit only recently to NVIDIA. And then also, more recently, you know, you're also the only Italian team selected to Sprint's Frontier AI Challenge. So the capital you raised in your pre precede seed round, and then also through the Sprint Challenge, what exactly is that going to be? used for and where does it get you to?

54:05Unidentified speaker (speaker_0): Yep. So first of all, our estimations is that to really get to the frontier, this is going to take us hundreds of millions of dollars, possibly billions. And so the and the the there's an implicit answer to the second question, which is: how far does this, what is this capital for? Now, of course, this capital is not for us to immediately go to the frontier. The amount of money that we have doesn't isn't yet a dead amount. But it's for something else. It's for taking our aim at the frontier. So imagine being OpenAI at 20 in 2015, 2016. There was Google, giant uh you know entity that was doing all of the AI and they they were training. all the big models back then. And they invented the Transformer then, and then they, you know, Google Deepind was a pretty difficult uh you know uh adversary to attack. Uh and of course, what happened is that

55:07OpenAI made an early bet on scaling loss, and that paid off over time. So what we're trying to work on here is making a bet, constructing a team behind the bet, constructing a mechanism behind this bet, and slowly proving that the bet is right. And the proving that the bet is right, as you know very well, uh attracts capital because you want to be early on something and you want to be able to Yeah, cash in and the difference in value, the perceived value between the two situations. So later on when you've made it, and before when nobody believes in it. And that's what this money is for. The money that we have today is to give convincing proof that we can be at the frontier, that we have what it takes, that we can build models that in a completely different way attack science. And that's something that I think you will see because. as we speak today, we are preparing a release for our first model, which is a very small model,

56:10which is designed to be used in a completely different way than all other language models, and it's specific for mathematics. I will not go into the details of it, but the entire point is that we're pre-training this from scratch and post-training it fully, and that's not only a win for us in terms of what the model can actually do, but it's also a win for us to sediment the capability of scaling this up one day. And this is something that we're heavily utilizing our now good resources that it would have been provided for.

56:44Tom Guinness: Right. And so once you get to that stage where, yes, you have this uh you know proof

56:50Unidentified speaker (speaker_0): bank.

56:50Tom Guinness: banked, you know, you've created the the model and you've you've shown that you can be at the frontier, you know, how do you think an investor should value a you know pre revenue but mission scale company like this?

57:05Unidentified speaker (speaker_0): Yeah.

57:05Tom Guinness: You know, what should they be actually underwriting?

57:08Unidentified speaker (speaker_0): I think what they're underwriting fundamentally, you know, they they invest because they're looking at future money, not today money, right? And so what they should be valuing, I think, is the model of the world that we're betting

57:21Unidentified speaker (UNKNOWN): on.

57:22Unidentified speaker (speaker_0): We're betting on some world that is going to exist. And that's

57:28Unidentified speaker (UNKNOWN): uh

57:29Unidentified speaker (speaker_0): it happens to be uh you know little coupled with the amount of money we're making today for mean for a bunch of RD reasons, really, if you will. And once those are unlocked, the RD is unlocked, then there's a convincing argument to be made that in that world you make a lot of money. And so that money transferred to today is a great deal for that investor. I think we have to really convince the investor that the world we're talking about exists and it's not just a fantasy of ours, and it's something that's really, really valuable.

58:03Tom Guinness: Right. And so when we think about the folks that you have convinced, you

58:08Unidentified speaker (UNKNOWN): know,

58:08Tom Guinness: your existing investors, you know, specifically in this sprint challenge, because

58:12Unidentified speaker (speaker_0): there

58:12Tom Guinness: was a huge amount of applications for this, you had next Frontier AI challenge, and you were the only Italian firm picked. How did you convince them to back your vision of the future? And you know, what do you see that unlocking?

58:27Unidentified speaker (speaker_0): Of course we don't have full visibility on this. Be happy to know as well, in a sense. I don't see it from their side, of course. I can tell you what I think convinced

58:37Unidentified speaker (UNKNOWN): them.

58:38Unidentified speaker (speaker_0): And first of all, I think is a specific mix of culture and speed of execution. And the two things go together.

58:51Unidentified speaker (UNKNOWN): Uh

58:52Unidentified speaker (speaker_0): think of XAI and Meta, two big, very large companies, very highly capitalized companies, that couldn't manage to build a frontier model. Meta was building Lama and then it failed. And XAI had really a lot of trouble building Grok to a frontier level. And the reason is it wasn't capital, it was culture.

59:15Unidentified speaker (UNKNOWN): It

59:16Unidentified speaker (speaker_0): was their ability to build a team that was capable of staying on top of things and coming up with the ideas and working on them very quickly. I think that's something that we showed very quickly because, in just a matter of few months, we hired an incredible team that got five-world records right away, with our main mechanism being that of autonomous research. And we were uh I think earlier than many Silicon Valley startups, I think we were one of one of the first companies in the world to really take this seriously, if not the first. And now there's a whole bunch of companies raising money to do

59:49Unidentified speaker (UNKNOWN): it.

59:50Unidentified speaker (speaker_0): And that's great. I think that validates what we are doing. And of course there is, I think, the also standing out of from the crowd in a bit in terms of ambition. Unfortunately, Europe feels like it, you know, history has ended also, and it also manifests, this phenomenon also manifests in the ambitions of the founders. Also, we don't have the kind of appetite for for risk that that the US has As often. But also the culture doesn't tend to

1:00:26Unidentified speaker (UNKNOWN): um

1:00:28Unidentified speaker (speaker_0): brings down founders that are shooting too much of a moonshot. And I think that's instead what we were pitching, very convinced of it, and we were very convinced that this can be, or maybe it's the only way uh Europe can come back. And I think that's something that really you know cut through, and and I think they were now at that point seeing. the autonomous science angle. It was much more clear when we pitched this to them. I think it was June. Uh and we started doing this in December of last year and nobody was really doing autonomous research. And so I think that's part of what convinced them.

1:01:04Tom Guinness: Right. And I think for for me at Raising Europe as well, you know, what I see is that AI is such a great opportunity for Europe to take back lost ground especially as you know AI develops not only in the software and application layer, but infrastructure and then physical AI, it's it's a going to encapsulate pretty much our entire society and change it like nothing else before. So I think having more investors, you know, like the folks at Sprint and the folks you know I mentioned in your precede pre-seed round who back these moonshot visions of the future, which will propel Europe into the future, we definitely need more of those.

1:01:45Unidentified speaker (UNKNOWN): And

1:01:45Tom Guinness: if we think on the converse side, you know, you mentioned the investors who pull down founders with great ambitions and the actual ability to execute on them. What would make an investor the wrong investor for you for your next rent? brand?

1:02:01Unidentified speaker (speaker_0): I think that in a sense, a wrong investor would not see, not be

1:02:11Unidentified speaker (UNKNOWN): uh

1:02:11Unidentified speaker (speaker_0): interested in the frontier. It would be a kind of investor that's more uh, let's say, directed towards um existing kind of business models and not hungry for uh some kind of novel ambition breaking through. in another direction. I think great uh investors that we're looking for are and there's nothing wrong with with just looking at existing things that that work. But we're mainly I think we're mainly aligned with um investors that think one, that we aren't done. We're not done with with with AI, we're just at the beginning. at the value that I can unlock. And second, that in a sense, it's possible for a European company with our ambition to make it, that believe in this underdog possibility of producing, just like it happened before, a vision that's gigantic. And and that today. it looks like uh super nonlinearly far from what we have today, but that's but that possibility is very that finds this possibility very entertaining instead of daunting.

1:03:25Tom Guinness: Right. And I think the difference between you know your vision of the future and your ambition and reality is in execution. Yeah. Right. So what's next for paradigma over the next 18 to 24 months? What are the priorities?

1:03:37Unidentified speaker (speaker_0): Paradigma first of all has in its priorities produced continuing and producing the machine for the internal operations and the autonomous operation needed for model training at scale. And so that's something that

1:03:58Unidentified speaker (UNKNOWN): you

1:03:58Unidentified speaker (speaker_0): know will take us to go from the model that we will release soon, which is a small model. very niche, very special in the way that it's used,

1:04:06Unidentified speaker (UNKNOWN): uh,

1:04:07Unidentified speaker (speaker_0): to models that are larger, that are possibly architecturally more similar to existing models today, and then maybe push the boundaries already in some architectural directions uh and and introduce some novelties in the field. And

1:04:23Unidentified speaker (UNKNOWN): at

1:04:23Unidentified speaker (speaker_0): the same time, fortunately enough, we can use open weights. models for the kind of setup that I talked to you about in serving the companies with private factories of RD. And that's something that we plan on expanding very strongly in the next few weeks and months. Because we think that, especially yesterday, uh in which you know a very large millennium problem was solved by OpenAI. it's become very clear that the models can do research, can do science, and there is maybe things that we weren't thinking of science before, which are things we can now optimize. We can now work on twenty-four-seven without anybody needing to be tired about it. And we think this unlocks an incredible amount of value inside companies. And that's something we really want to expand uh and to bring to fruition. because we think it's going to be a key element for us to be a grounded company and a company that's capable of scaling to what we to our ambitions. And of course, there's the progression in the Sprint program, which we think very seriously. We think it's a

1:05:31great opportunity that will allow us to unlock equity-free equity free funding eventually in case we make it among the six out of ten companies. that will uh access it phase two. And of course, we also are, you know, focused on achieving also phase three, which unlocks for us twenty six point five million uh in equity in equity free funding. And we think this will significantly significantly help us in acquiring the compute and utilizing it to train models that go in the direction of maximizing discoveries for Java, which is our objective.

1:06:12Tom Guinness: Right. And so when we think about where you need to get to, what's the biggest constraint? Is it compute, talent, is it capital? Is it partners? What is it?

1:06:23Unidentified speaker (speaker_0): Today's Today is compute. Uh the main thing that's stopping us from training a much larger model today, which is going would be a significantly better, I think, outcome for science in a sense, is the amount of compute that we have. Of course, part of it is capital, but also a side of it is the let's say necess of course the problem with compute is that you have to reserve

1:06:50Unidentified speaker (UNKNOWN): it.

1:06:51Unidentified speaker (speaker_0): right? So there is a reason it's that the companies raise so much money is that uh you there is a you know kind of a very strong competition they're have access to either buy or reserve com large amounts of compute for a long time. And oftentimes startups have trouble doing this. And currently if we had if we had much more GPUs, we could definitely uh accelerate so much. Uh, but that's something that we're on our way to fix, and hopefully we'll do this uh as quickly as possible.

1:07:25Tom Guinness: Right. So creating a frontier model for you know limited amounts of capital is assuming that you have the compute elsewhere.

1:07:32Unidentified speaker (UNKNOWN): Yep.

1:07:33Tom Guinness: Right, someone else is providing the compute. How do you bring all of that potentially in-house to make a truly European sovereign AI frontier lab, which, you know, aside from critical materials, you know, builds everything in-house?

1:07:50Unidentified speaker (speaker_0): I think Europe will have to make a set of kind of drastic maneuvers around compute. I think still today our policymakers are underestimating the necessity of

1:08:02Unidentified speaker (UNKNOWN): it.

1:08:03Unidentified speaker (speaker_0): Just in the way that unfortunately they have over time underestimated the necessity in investing in in the internet, or

1:08:11Unidentified speaker (UNKNOWN): in

1:08:11Unidentified speaker (speaker_0): in Italy possibly. I like to let's say make the analogy with nuclear power, because our policymakers Makers have, for a bunch of reasons, decided not to make use of

1:08:24Unidentified speaker (UNKNOWN): it.

1:08:25Unidentified speaker (speaker_0): Uh And unfortunately, we we feel the consequences of this today. I think that

1:08:32Unidentified speaker (UNKNOWN): on

1:08:32Unidentified speaker (speaker_0): the flip side, Europe is a place where this can happen, uh aside from people that are,

1:08:41Unidentified speaker (UNKNOWN): you

1:08:42Unidentified speaker (speaker_0): know, very negative about it, uh especially because we're such a low point. And being such a low point, if we have the right people. discussing this, and if we reach, I think, a critical threshold in capabilities from other models, I think we are still in time to make a big difference. I think that people extrapolate a linear uh future in which everything keeps happening as it as it is. I think we're still the free uh if you know the frequency of breakthroughs will increase over time. And so the amount of things that will shake up everything will increase over time. And so there will be many occasions for us to go back at maybe needing less GPUs in the first place, or or being capable of manufacturing more, or being capable of acquiring more. And so we I think we have to tackle the problem of acquiring a compute, which is very urgent, with more optimism and more aggression. Aggression in a the positive sense. being more aggressive about it.

1:09:40Tom Guinness: And what is that? More spend?

1:09:42Unidentified speaker (speaker_0): Yes, it's more spend, it's economic policy, it's lowering the regulation. And it's also possibly investing in directions that go away from existing providers. Of course, that's going to take a while. But for example, or leveraging our existing economic leverage, we have ASML. And we and there's a window in which we can leverage this. and it's closing, and we should.

1:10:05Tom Guinness: Yeah, and I think ASML is clearly our most valuable company here on the continent with over 700 billion market cap as of time of recording. And we're also seeing some really great frontier chips coming out of Eindhoven, for example, with Accelera AI. We're starting to see a confidence. confluence of factors here. So now is probably the time.

1:10:30Unidentified speaker (speaker_3): Yes. And we have to make it the time if it isn't. We have no other choice.

1:10:35Tom Guinness: So we talked about this earlier. You mentioned the new first world split by AI capability and also compute capacity. What's the boldest version of that prediction that people aren't thinking about today?

1:10:48Unidentified speaker (speaker_0): Yep. I think there's a world in which which we end up being losing our personal liberties because we cannot project the force to defend them. Because we end up being a nation that's simply incapable of that.

1:11:03Tom Guinness: So other nations will be providing the things we need day to day to survive.

1:11:08Unidentified speaker (speaker_0): That's almost the best case, unfort if unfortunately. There could be a situation in which they provide them as long as they want. and we simply have no leverage in you know having access to those things if we don't manufacture one.

1:11:25Tom Guinness: Outside of paradigma,

1:11:27Unidentified speaker (speaker_0): what

1:11:27Tom Guinness: can we find you doing?

1:11:29Unidentified speaker (speaker_0): Oh well, uh quite a few things. I love combat sports, I specifically boxing and kickboxing. I uh

1:11:38Unidentified speaker (UNKNOWN): spend

1:11:38Unidentified speaker (speaker_0): a lot of time on the internet. You can find me discussing all kind of topics on X on Twitter as TensorQT. That's how you will find me. I like discussing really from philosophy to machine learning, of course, mathematics, physics. And you can find me in Garbatella in Rome, which is where our headquarters are.

1:11:57Tom Guinness: So Francesco, it's been great having you on the show. Thank you so much

1:12:01Unidentified speaker (speaker_0): for

1:12:01Tom Guinness: taking the time to speak to us today.

1:12:02Unidentified speaker (speaker_0): Thank you so much for having me.

1:12:04Tom Guinness: That's a wrap for today's episode of Raising Europe. Thank you so much for tuning in. And as always, remember, your biggest edge as a founder comes down to resource management. If you want to stop guessing, head on over to getfrontierstack.com right now and join the Waynelith. Lock in your priority Already pass passed now so you can grow faster and spend smart. smarter. If you've enjoyed today's episode, be sure to subscribe, hit the like button, and share this episode with a fellow founder who's building right now. Until next week, keep building and we'll see you soon.

This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights