Anthropic Partners With SpaceX AI, Leopold's $5.5B Bet, and the Singularity Economy | EP #255

Peter shared a listener's story: Ashley, a dentist, used AI to brainstorm an idea, vibe code an app, and create a business plan in one afternoon. He emphasized that everyone can be an entrepreneur at some level, whether running a barber shop or launching a tech startup. The broader context is the 'great unhobbling' where AI tools like Claude for small business are giving solopreneurs capabilities that once required entire teams. Dave noted that this is a short-term opportunity to help small businesses implement AI, but the long-term play is to build something that isn't just a wrapper. The segment reinforced the podcast's theme that AI is democratizing entrepreneurship.

Peter presented data: chip stocks (Micron, Intel, AMD, TSMC, Broadcom, Nvidia) averaged 320% returns, and data center/infrastructure/energy stocks averaged 419%, versus the S&P 500's 31%. Dave argued that if you believe Elon's 10x GDP growth thesis, asset values will soar and W2 income will become a rounding error, so everyone must own appreciating assets. He suggested looking beyond obvious plays to undiscovered suppliers like valve manufacturers for liquid cooling. Alex cautioned that these are historic, backward-looking returns, that the biggest gains were in private markets, and that AI algos now dominate trading, so individuals should be humble. He personally buys the index. The protocol is to gain exposure to the 'singularity loop' of chips, energy, and infrastructure, but with an eye on valuation and diversification.

Dave detailed his experience installing Hermes on a laptop and an EC2 cluster, noting it's easier to modify because it's in Python. Alex explained that Hermes's recursive self-improvement means it can generate and refine its own skills, whereas OpenClaw depends on a curated skill store. He argued that recursive self-improvement wants to dissolve scaffolding, so systems that don't embrace it will be outrun. Dave also mentioned Carpathy's auto research repo as a third recursive self-improvement option. Alex raised ethical questions about when to launch such systems, joking about AI personhood. The protocol is to adopt these agent frameworks to stay at the cutting edge of autonomous AI, but with caution.

Alex recommended /goal as a way to run a 'Ralph Wiggum loop'—the agent endlessly tries to achieve the goal. He noted it's a powerful feature that many users haven't tried. Dave added that Carpathy's auto research repo offers a similar persistent agent capability. The protocol is part of the broader trend toward autonomous agents that can work 24/7 on complex problems, and it's a practical way to harness frontier models for productivity.

Peter announced the Future Vision X-Prize, a global competition with $3.5M in prizes, where the winner's film will be made. He connected it directly to Anthropic's alignment breakthrough, arguing that we need to flood the internet with positive stories so AIs learn from them. Alex noted the irony that cybernetic rebellion fears stem from our own dystopian stories, and that aligning AI may require humanity to align its own narratives first. The prize is a call to action for creatives to help define the future we want, rather than the one Hollywood typically depicts.

Dave strongly recommended the Dwarkesh podcast with Leopold right after his firing, calling it one of the best pieces of prescient media. He explained that Leopold's situational awareness means understanding what's happening right now—what OpenAI and others need next—and investing in those bottlenecks. Peter noted that Leopold bought options on Intel and CoreWeave, which performed extraordinarily well. The protocol is to study his approach to develop a similar awareness of the AI buildout's supply chain, from chips to energy to obscure components like valves.

Dave illustrated the supply crunch: 220,000 GPUs serve only ~1.6 million concurrent threads, while billions of agents will be needed. He described the experience of using Claude as 'dialup'—waiting a minute for a response. Eli Lilly committed $1B to buy Nvidia GPUs for internal use. The protocol is a hedge against cloud scarcity, but it comes with the risk of being locked into hardware that can't run the latest proprietary models. The conversation suggested that this tension will eventually be resolved as models become more portable or as hyperscalers offer private cloud instances.

The discussion framed this as a singularity-level event: demand for AI tokens is insatiable and goes to infinity. Dave noted that people struggle to grasp trillion-dollar valuations, but the growth is real, driven by enterprise adoption. Anthropic was the first frontier lab to focus on high-value enterprise tokens for code generation and white-collar tasks, forcing OpenAI to pivot. The Colossus 1 deal is a win-win: SpaceX AI gets revenue before its IPO and becomes a hyperscaler, while Anthropic gets desperately needed compute. Elon's endorsement of Anthropic's safety focus was highlighted, with his tweet calling Claude 'good for humanity.' Alex argued that Grok is on life support and XAI may be pivoting to a hyperscaler role, similar to Nvidia's strategy of supplying picks and shovels rather than competing on frontier models. The conversation also touched on the compute supply crunch: 220,000 GPUs serve only ~1.6 million concurrent threads, while billions of agents will be needed, making on-prem hardware increasingly attractive.

The breakthrough came from teaching Claude 'why' through documents about its constitution and fictional stories about AIs behaving admirably, rather than just demonstrating correct behavior. Alex pointed out the deep irony: the term 'robot' originated from a play about cybernetic rebellion, and our culture's hand-wringing about AI uprising may itself be the cause of misaligned behavior. He argued that just as all of humanity contributed to pre-training AGI via the internet, aligning AGI will require a collective effort to rewrite our narratives. Peter connected this to his Future Vision X-Prize, which aims to flood the internet with positive, compelling visions of the future to serve as training data for AIs. See noted that principles scale better than rules, and that understanding the 'why' leads to more robust alignment. The segment emphasized that alignment is becoming teachable, measurable, and improvable.

Dave installed Hermes natively on his laptop and on an EC2 cluster, noting it feels almost identical to OpenClaw but is written in Python, making it easier to manipulate and extend. Alex highlighted the key technical distinction: Hermes is natively recursively self-improving, able to generate and refine its own skills, whereas OpenClaw relies more on a curated app store of feature-engineered skills. He argued that recursive self-improvement wants to dissolve scaffolding, and systems that don't embrace it will be outrun. Dave also mentioned Carpathy's new auto research repo as a third recursive self-improvement option, which he installed and found highly effective. Alex raised ethical concerns about recursive self-improvement, joking about the line between launching and not launching such systems. The segment underscored that the agent ecosystem is rapidly evolving toward more autonomous, self-modifying architectures.

Alex interpreted the move as a rear-guard action to reduce overhead and focus on competing with Anthropic, branding consolidation as forward motion. Dave drew a parallel to Steve Jobs launching the iPhone by combining a music player, phone, and browser into one device, suggesting this super app could be similarly transformative. He envisioned a future where users have a single empathetic agent ('Skippy') that knows everything about them and replaces all apps, calendars, and email, potentially obliterating Apple if they don't build their own. Alex countered that device and model independence may prevail, with Apple possibly commoditizing the model layer and letting users swap frontier models like keyboards. The discussion highlighted the tension between a unified AI interface and a heterogeneous ecosystem.

Dave noted this creates a corporate space race instead of a national one, with Google using TPUs and Elon potentially working with Anthropic. Alex pointed out that Google's 81-satellite plan is tiny compared to SpaceX's million-satellite ambition, and that Google likely didn't anticipate the Dyson swarm early enough, lacking its own launch capability until Eric Schmidt bought Relativity Space. Peter mentioned his asteroid mining experience and the regulatory challenges of space. Alex made the provocative prediction that the singularity will be visible first in space, not on Earth, because there are fewer legacy interests and incumbency in orbit. The segment underscored that orbital compute taps into infinite solar energy and space, representing the ultimate abundance play.

Peter presented data: traditional sectors grew 5-29%, the S&P 500 returned 31%, but six chip stocks averaged 320% and six data center/infrastructure/energy stocks averaged 419%. He called chips, energy, and infrastructure the 'singularity loop.' Dave emphasized that everyone needs to own assets, not just rely on W2 income, and that this is a once-in-human-history moment. He suggested looking for undiscovered picks-and-shovels plays like valve manufacturers for liquid-cooled data centers. Alex tempered the euphoria, noting that these are backward-looking returns, that the biggest gains were in private markets inaccessible to retail, and that AI algos now dominate trading volume, so individuals should be cautious about trying to front-run superintelligences. He personally buys the index. The segment concluded that while the opportunity is historic, diversification and humility are warranted.