Transcript Viewer
Jul 23, 2025 ยท 2025 #27. Read the transcript grouped by speaker, inspect word-level timecodes, and optionally turn subtitles on for direct video playback
Speaker Labels
Edit labels for this show, save them in this browser, or download a JSON override for the production folder.
Human Transcript
Blocks are grouped by speaker for readability. Expand a block to inspect word-level timing.
Speaker 2
Welcome to the Deep Dive. Every week brings, well, just a deluge of information, doesn't it? Our mission here, kind of like the fantastic That Was The Week newsletter, is really to cut through all that noise.
Speaker 1
Exactly. Give you the shortcut to being genuinely well-informed.
Speaker 2
Think of this as our editorial board meeting. We've sifted through all the July readings, you know, the articles, the research, our notes.
Speaker 1
And we're here to basically distill them. Find those really important nuggets, the surprising facts and, well, the critical insights that actually matter.
Speaker 2
Absolutely. So for this deep dive, our aim is to unpack these big shifts we're seeing across venture capital, the explosive growth, and frankly, the profound questions coming out of AI.
Speaker 1
And also those surprising intersections, right? Tech, finance, geopolitics, they're all tangled up.
Speaker 2
Exactly. By the end, you should have a much clearer grasp of what truly matters from all these, let's face it, pretty complex sources. Okay, so let's jump right in. The venture capital landscape. It feels like it's constantly redrawing itself. From this month's readings, what were the standout trends that really grabbed your attention?
Speaker 1
Well, what was really striking was this kind of tale of two VCs playing out. On one hand, you've got SpaceX. I mean, heading for an astonishing $400 billion valuation in a share sale.
Speaker 2
$400 billion?
Speaker 1
Wow. Yeah, it's a massive leap. And it seems largely fueled by Starlink, its satellite internet service. They're projecting, what, $6.6 billion in revenue this year from 5 million subscribers.
Speaker 2
Incredible numbers.
Speaker 1
It really is. To put that in perspective, this valuation pushes SpaceX into the top 20 of the S&P 500. It shows just immense investor confidence, even with some of the political noise surrounding it.
Speaker 2
That SpaceX valuation is definitely eye popping. But does that, you know, tell the whole story for VC right now? I'm wondering what else the reading showed about maybe underlying health of the industry, because it feels like that kind of mega deal could be masking a deeper trend.
Speaker 1
That's a great point. Broadly speaking, yeah, the data points to VCs making fewer deals, but much, much larger ones, more selective investments. So while the total dollars invested might be up, the actual number of funding rounds, that's down to a seven year low.
Speaker 2
Seven years.
Speaker 1
Yeah. And early stage rounds, for example, are down a huge 53 percent from their peak back in 2021.
Speaker 2
So what does that mean in practice?
Speaker 1
Well, the implication is longer fundraising cycles for a lot of startups and a significantly narrower funding funnel overall. Fewer companies getting through.
Speaker 2
And it's not just the overall number of rounds shrinking, right? I remember one article. It's not just you. Lead VCs are taking more of each round. It highlighted that lead investor participation is jumping up.
Speaker 1
That's right. Particularly at the seed stage, it went from, I think, 52% in 2021 to a projected 61% in 2025.
Speaker 2
So what's the implication there for founders then?
Speaker 1
Well, for founders trying to navigate this, it means there's just less room for those syndicates, you know, multiple smaller investors chipping in. And you get more concentrated power resting with fewer, larger lead investors. It really reflects this flight to quality we hear about, combined with the fact that many VCs are now managing much larger funds they need to deploy.
Speaker 2
We also saw that piece titled, A Crisis Moment for SEED VC. What's specifically driving that intense pressure at the very earliest stage of funding?
Speaker 1
Yeah, that article pointed to a few key things. Industry maturation, for one. Then there's the increasing specialization of technology. Right. And just the overarching macroeconomic environment. All these factors are putting immense pressure on seed funds, plus the classic power law, you know, where a tiny number of investments generate nearly all the returns that seems to be evolving. Right.
Speaker 2
And the AI shift must play a role, too.
Speaker 1
Absolutely critical. The AI platform shift is demanding entirely new expertise, new investment theses, which, frankly, some seed funds are struggling to adapt to quickly enough.
Speaker 2
OK, let's talk exits. Our readings highlighted a really striking shift. Only 11 percent of unicorn exits are IPOs now. That's down dramatically from 83 percent back in 2010. What does that mean for founders who maybe grew up dreaming of bringing that bell in the stock exchange?
Speaker 1
Yeah, that statistic is a real game changer, isn't it? It means that, well, abundant private funding is still available. The high compliance costs of going public are deterrent. Plus, you've got the rise of these sophisticated secondary markets and, of course, strategic acquisitions. They're all fundamentally changing the exit game.
Speaker 2
Can you give an example? Sure.
Speaker 1
Look at Figma. Their recent IPO saw existing shareholders, including the CEO, Dylan Field, who cashed out over $62 million, and big VCs like Index and Sequoia, they actually sold more shares than the company itself raised.
Speaker 2
Interesting. So the company wasn't raising as much primary capital.
Speaker 1
Exactly. Field still holds 74 percent of the voting rights, mind you. But this kind of setup provides liquidity for those early investors and employees without putting the full pressure of a traditional IPO on the company itself. It's a different model.
Speaker 2
And finally, on the VC front, we saw that unique story of NFDG, the $1.1 billion VC fund that, what, quadrupled in just two years and then got acquired by Meta. That's pretty wild, right?
Speaker 1
It is pretty wild.
Speaker 2
What does that tell us about the current landscape?
Speaker 1
I think it really highlights the rapidly blurring lines between traditional venture capital and direct operational roles. Yeah. Especially within AI. How so? Well, NFDG's incredible success was partly thanks to their investment in safe superintelligence, which saw its valuation jump sixfold from $5 billion to $30 billion. Yeah.
Speaker 1
The fact that top talent like Nat Friedman and Daniel Gross are now being drawn directly into companies like Meta to actually build these transformative technologies rather than just funding them from the outside.
Speaker 2
Right.
Speaker 1
It just underscores the immense value being placed on deep AI expertise and maybe the allure of having the resources to build the next generation of technology directly.
Speaker 2
Okay, let's shift gears then. Let's talk AI. It really dominated so many of our July readings. We saw incredible model breakthroughs, but also some really surprising societal discussions. It feels like AI is reshaping everything at, well, breakneck speed.
Speaker 1
It really is.
Speaker 2
What were some of the biggest developments you saw this past month?
Speaker 1
I think one key takeaway was just the sheer speed of development. Look at Anthropic. They rocketed to $4 billion in annual recurring revenue, ARR.
Speaker 2
$4 billion. Up from what?
Speaker 1
Up from just $10 million in 2022.
Speaker 2
That's staggering growth.
Speaker 1
It's unprecedented. And their growth seems primarily driven by this API-first paper token model, especially for code generation tasks, which apparently consume 10 to 50 times more tokens than typical chat interactions.
Speaker 2
Interesting.
Speaker 1
And they've also really cleverly leveraged channel partnerships like AWS Bedrock. It's fundamentally changing how we think about B2B growth, how it's measured, how it's achieved. The pace is just incredible.
Speaker 2
Then there was XAI's Grok 4. It launched with big claims, right? Leading benchmarks, 10 times more reinforcement learning compute for reasoning. But the sort of vibe tests, the real world usage seem maybe a bit mixed. Is that fair?
Speaker 1
I think that's fair. That's the core tension, isn't it? While it boasts really impressive performance on paper, on the benchmarks across the board, Grok IV seems to be facing what some call cultural risk and maybe a perceived lack of differentiation in just daily use compared to peers like Claude 3.5 Sonnet.
Speaker 2
What about that heavy mode?
Speaker 1
Right. It's heavy mode, which dynamically spawns multiple agents for a task, does show promising results for complex stuff. It even outperformed OpenAI deep research on some specific information retrieval tests. But the price? Exactly. But the price point, $300 a month versus its perceived usefulness compared to much more affordable options like ChatGPT at $20 a month. That definitely raises questions about its broader market adoption. Is it worth it for most people?
Speaker 2
And just as Grok 4 was making waves, Moonshot AI's Kimi K2 dropped. Some described it as a deep seek moment 2.0. What makes Kimi K2 so significant? And why should, say, Silicon Valley be paying close attention?
Speaker 1
Well, the crucial question here really is about the global AI landscape. Kimi K2 is open source, first off, and it has a massive one trillion parameters.
Speaker 2
One trillion open source.
Speaker 1
Yeah. It really shows that China is rapidly reaching the absolute frontier of AI model development. It was trained on an astonishing 15.5 trillion tokens.
Speaker 2
Okay.
Speaker 1
And likely for a relatively low cost, maybe in the tens of millions, this demonstrates significant algorithmic gains.
Speaker 2
Weeding.
Speaker 1
meaning they found more efficient ways to train these incredibly powerful models, maybe at a fraction of the cost we'd expect, thanks to breakthroughs in their optimization techniques. This model is highly competitive with the leading frontier models from the U.S.
Speaker 2
And that puts pressure on U.S. Open models.
Speaker 1
It seems to be causing American Open models to fall further behind, yes. We even saw OpenAI delay their own planned open-weight model launch right after Kimi K2 came out, though they denied it was related.
Speaker 2
Interesting timing, though.
Speaker 1
The implicit message seems pretty clear. Competition is heating up globally.
Speaker 2
Now, one article suggested that developments in Grok 4 and maybe some of OpenAI's recent work kind of accidentally vindicated neurosymbolic AI. What exactly is that and why is it suddenly important again?
Speaker 1
Right. So stepping back a bit, neurosymbolic AI... basically integrates neural networks, the pattern recognition part, with symbolic reasoning, you know, explicit logic and rules. Okay. It's designed to directly address deep learning's known limitations in things like causality and abstract reasoning. Think of it as combining pattern matching with actual logic.
Speaker 2
Like the code interpreter in ChatGPT.
Speaker 1
OpenAI's code interpreter is a prime example of this approach and practice, yeah. What the benchmarks are increasingly showing is that just scaling up models bigger and bigger has diminishing returns for true intelligence.
Speaker 2
Right.
Speaker 1
But integrating these symbolic tools significantly improves performance on complex reasoning tasks, even if the industry doesn't always, you know, openly label it neurosymbolic.
Speaker 2
So it's about building AI that can actually reason.
Speaker 1
Exactly. It signals a shift towards AI that can truly reason and tackle complex problems beyond just pattern recognition, like legal analysis, scientific discovery, that sort of thing.
Speaker 2
We're also seeing articles about the rise of the agent manager. What does that mean for the future of work for productivity?
Speaker 1
Well, if 2025 really is the year of AI agents, as many predict, then 2026 will undoubtedly belong to the agent managers, the people overseeing them, because AI agents are inherently non-deterministic.
Speaker 2
Meaning they don't always do the same thing twice.
Speaker 1
Exactly. They interpret, they improvise, which makes managing them incredibly complex, much harder than managing traditional software.
Speaker 2
So how do you manage them?
Speaker 1
Well, this concept of an agent inbox is emerging as potentially a critical tool for overseeing multiple AI agents working on different tasks. We might even start seeing agents managed per person become a new important productivity metric.
Speaker 2
Then OpenAI is moving here too.
Speaker 1
Yes, OpenAI just released its ChatGPT agent, specifically designed for increased autonomy and navigating complex multi-step tasks. It's pushing this trend forward. It really suggests a future where a significant chunk of white-collar work might be orchestrated through these AI supervisors.
Speaker 2
Beyond these big technical shifts, AI is impacting society in some fascinating, maybe even slightly weird ways. We read about AI-powered filmmaking, like Kavan's Untold the Immortal Blade saga.
Speaker 1
And tools like VO3 promising studio-level commercials with one click. It seems to be genuinely democratizing content creation.
Speaker 2
It absolutely is. However, this leads us to wonder about the downsides, right? The article, How AI Could Make Us Dumber, really highlighted some growing concerns.
Speaker 1
Like what?
Speaker 2
Well, over-reliance on AI is being directly linked to potential cognitive atrophy, diminished critical thinking skills. The article mentions students openly using tools like ChatGPT to write essays.
Speaker 1
The cheating utopia.
Speaker 2
Yeah, creating what some professors are calling a cheating utopia that's incredibly difficult to police effectively.
Speaker 1
It sparks this crucial societal debate, doesn't it, about the true value of education and how we should interact with AI.
Speaker 2
As an augmentation tool, not a replacement.
Speaker 1
Precisely. As an augmentation, certainly, but maybe never as a wholesale replacement for human thought. It's almost like our brains are getting lazy, like a muscle you stop exercising. I even caught myself relying on AI for quick answers where maybe I used to think things through more deeply myself.
Speaker 2
Yeah, I think we all have. And then there was the piece on the endless rebranding of AI. Is this just marketing speak or does it signal something deeper about the industry itself?
Speaker 1
I think it's more than just marketing, actually. It reflects the industry's constant attempt to define itself, to differentiate itself amidst really fierce competition, especially for talent.
Speaker 2
So terms like AGI, superintelligence.
Speaker 1
Right. The terminology around AI from just AI to artificial general intelligence or AGI and now superintelligence, it's constantly evolving and frankly, often highly ambiguous. Companies like Meta save superintelligence. They are strategically rebranding their AI efforts specifically to attract the very best talent.
Speaker 2
We saw that with Meta poaching Apple's AI exec.
Speaker 1
Exactly. Meta reportedly lured Apple's top AI models executive, Ruming Pang, with a compensation package allegedly exceeding $200 million. The competition for top AI engineers is so intense right now that they're literally being paid more than many professional athletes.
Speaker 2
Incredible. OK, so as AI transforms industries, other major forces are obviously still at play, shaping our world in complex ways. Let's maybe delve into some of those fascinating intersections we saw in the July readings, tokenization, geopolitics, the attention economy.
Speaker 1
Yeah, definitely. One key takeaway from the sources was this accelerating push to tokenize everything. You saw Robinhood, for instance, offering European investors tokenized shares linked to open AI.
Speaker 2
Though not actual equity.
Speaker 1
Right. Important distinction, not real equity. But the move itself sparked this big debate about whether there's genuine demand for tokenized private equity. At the same time, though, we saw new legislation, the Genius Act, pass the Senate.
Speaker 2
And that's aimed at regulating stablecoins.
Speaker 1
Exactly. Requiring them to be one-to-one backed by real liquid assets, specifically U.S. treasuries, trying to bring some stability there.
Speaker 2
But that legislation wasn't without significant controversy, was it?
Speaker 1
No, definitely not. Critics, including prominent figures like Senator Elizabeth Warren, voiced strong fears that it could inject crypto into the most stable part of the U.S. economy.
Speaker 2
And the risk there is.
Speaker 1
The potential for systemic financial crises. They drew uncomfortable parallels to the 2008 subprime mortgage crisis, especially if stablecoins were to fail on a large scale. It certainly raises serious concerns about leverage and the potential impact on the stability of the treasury market if this sector grows too large too fast without guardrails.
Speaker 2
OK, shifting to geopolitics, we saw Nvidia get the nod from Washington to resume sales of its H20 AI chips to China after a temporary ban. What's the bigger game being played here between the U.S. and China?
Speaker 1
Yeah, that was interesting. Zooming out to the bigger picture, David Sachs, who previously advised the Trump administration on AI, actually publicly defended that reversal. On what grounds? He argued it helps the U.S. compete more effectively internationally, particularly against the Chinese tech giant Huawei. He noted, and this is key, that Chinese AI models are perhaps only three to six months behind those in the U.S. That close. Emphasizing this incredibly tight race and the very real risk that being overly restrictive with regulations could inadvertently stifle American innovation and actually push global markets towards China. It's a delicate balance.
Speaker 2
And how is a company like Amazon navigating this incredibly complex U.S.-China trade relationship?
Speaker 1
Well, our readings show that over 60 percent, 60 percent of Amazon sellers are now based in China.
Speaker 2
Wow, that high.
Speaker 1
Yeah. And they even launched an Amazon haul specifically for those sub-$20 Chinese products. However, the Trump administration tariffs, which are still mostly in place and above 50 percent on many goods, significantly complicate this strategy.
Speaker 2
Increases costs, presumably.
Speaker 1
Right. It increases costs, which impacts pricing. But Amazon is clearly adapting, leveraging its immense logistics network and global reach. It's a very delicate balancing act for them, trying to maintain access to affordable goods while navigating these strong political headwinds.
Speaker 2
OK, finally in this section, let's talk about the attention economy. There is that article from Dollar Dominance to the Slot Machine, which painted a really stark and frankly quite bleak picture of the U.S. What was its core argument?
Speaker 1
Yeah, that was a provocative piece. The crucial question it raised was really about national priorities. It argued that the U.S. has increasingly become an extraction economy, optimizing primarily for attention through spectacle rather than focusing on genuine creation or production.
Speaker 2
What example did they use?
Speaker 1
Well, it cited the perhaps slightly absurd example of the Department of Homeland Security tweeting about a White House UFC fight night as sort of evidence of the U.S. becoming the world's most powerful content creator instead of instead of focusing on building things. In stark contrast, the article described China as a creation economy. actively building out electrical capacity, implementing long-term industrial policy. The article suggested that this imbalance, coupled with what it called financially incoherent policies like the Big Beautiful Bill, which adds trillions to the debt through tax cuts while potentially cutting vital programs and a weakening dollar, are all actively eroding U.S. competitive advantage on the global stage. It's a pretty strong claim about what we as a nation seem to value most right now.
Speaker 2
Okay, shifting focus slightly, our July readings also highlighted some other fascinating developments, things like self-driving cars, innovative new browsers, and also a really interesting deep dive into Google's multifaceted approach to winning the AI race. Let's try and unpack some of these more disparate but maybe equally impactful breakthroughs. Sure.
Speaker 1
A really striking development was Uber's multi-billion dollar deal with Lucid. They're planning to purchase 20,000 electric vehicles and invest $300 million to build out an electric robo-taxi fleet.
Speaker 2
That's a huge commitment.
Speaker 1
It really is. It signals Uber's deep commitment to, well, sustainable and autonomous transportation. They're leveraging Lucid's advanced EV technology specifically to scale up their future fleet. It feels like a significant move that could really redefine urban mobility down the line.
Speaker 2
And sticking with automotive, we saw Rivian is getting a new navigation system integrated with Google Maps. How does that fit into this bigger picture of tech integration?
Speaker 1
Well, it's another clear example, isn't it? Established tech players like Google integrating really deeply into traditional industries. This partnership uses Google Maps. Auto simulator technology sounds fancy, but it means more accurate routes, better real time traffic.
Speaker 2
Which is crucial for EVs and range anxiety.
Speaker 1
Exactly. It significantly enhances the overall electric vehicle driving experience for Rivian owners. It just shows again how software and AI are becoming absolutely foundational, even for companies primarily making hardware.
Speaker 2
Okay, let's switch to the browser wars. AI is clearly the new battleground there. We saw Perplexity AI launch its Comet browser to directly challenge Google Chrome and also the browser company's DIA browser. What's the differentiator? What sets these new AI powered browsers apart?
Speaker 1
Yeah, it's heating up. The data we saw suggest Comet integrates Perplexity's own AI search directly into the browser experience, so users can query web page content, ask follow-up questions, get summaries, all without leaving the page they're on.
Speaker 2
And it's privacy-focused.
Speaker 1
Supposedly, yes. Designed to be privacy-first, storing data locally. Dia, on the other hand, positions itself more as an AI-first browser, focused squarely on productivity. It has features like smart macros and context switching designed to automate complex workflows and help users manage different project contexts more easily.
Speaker 2
So the big question is?
Speaker 1
The central question these products raise is really, is AI truly the future of web browsing? And maybe, are we ready for a browser that actively thinks for us or alongside us?
Speaker 2
And speaking of thinking, let's talk Google. There was that essay, hashtag 27, long Google, making a really bold prediction that Google is actually the best positioned company to win the AI race and could become, wait for it, a $20 trillion plus company.
Speaker 1
Yeah, audacious is the word.
Speaker 2
What's the core argument behind such a claim, especially with all the hype around OpenAI, Anthropic and others?
Speaker 1
Well, the crucial question the essay tackles is how Google is perceived versus its actual position in the AI race. You're right. Many investors have been bearish on Google, fearing ChatGPT will eat into its core search revenue. But the essay argues first that search volume is actually still growing. And second, paradoxically, those ChatGPT style queries might end up being more monetizable for Google because they often indicate higher user intent.
Speaker 2
Okay, but what about the structural advantage?
Speaker 1
That's the crux of it. The argument is that Google's structural advantage is simply staggering. They own pretty much the entire vertical stack needed to win in AI.
Speaker 2
Like what specifically?
Speaker 1
Think about it. The most visited website in the world. The world's number one consumer brand recognition. Gemini. which is arguably among the very best AI models available YouTube, giving them the biggest video data set. Search data. Search data, the biggest internet data set, Google Books, Gmail, Android, Chrome, Google Cloud Platform, GCP, and critically, their own custom AI chips that TPUs.
Speaker 2
So they control the whole type line.
Speaker 1
Pretty much. The essay's core thesis is that while other firms will eventually hit limits on their resources, compute, data distribution, Google simply has vastly more of everything needed to push further and faster. It positions them, the argument goes, for potentially unparalleled dominance. It's certainly a powerful counter narrative to the prevailing skepticism you often hear.
Speaker 2
Well, we've certainly covered a remarkable range of developments this July, haven't we? I mean, from those changing dynamics in venture capital where it really seems bigger bets are being placed on fewer, maybe safer companies to the just breakneck speed of AI innovation highlighted by, you know, the meteoric rise of companies like Anthropic, but also that competitive pressure coming from new open source models like Kimi K2 out of China.
Speaker 1
Absolutely. And we've really seen how AI is just intersecting with everything, haven't we? From geopolitics and the very fabric of the attention economy to these deeper philosophical questions surrounding the definition of human intelligence itself. It's clear these shifts aren't happening in isolation. They're profoundly interconnected and they're reshaping our world, well, in real time.
Speaker 2
So what does this all mean for you, our listener, trying to make sense of it all? Here's a final thought. If AI genuinely has the potential to, as some articles suggest, make us dumber by outsourcing our critical thinking. A scary thought. And yet it's simultaneously poised to become an integral part of nearly every aspect of our lives, what responsibility do we as individuals and maybe as a society bear in shaping how we interact with these incredibly powerful tools?
Speaker 1
Yeah. How do we ensure it helps us, not hinders us? Yep.
Speaker 2
Are we consciously building a future where AI truly augments our collective intelligence? Or are we accidentally creating one where it simply replaces it, maybe making us less capable in the long run, something to mull over until our next deep dive?