finance
Conversations with the best investors and business leaders in the world. We explore their ideas, methods, and stories to help you better invest your time and money. Hear stock market and boardroom insights you can't find anywhere else. If you're a professional investor, CEO, entrepreneur, or business strategist, this is for you. Explore all our episodes and learn more at https://www.colossus.com
Sarah Guo expresses concern about the current frenzied and chaotic state of the AI landscape, suggesting it might escalate further. She believes that investors navigating this period need to consider how the situation unfurls and differs from past boom-bust cycles in technology.
Sarah Guo supports the "great person" theory of history, stating that high-agency individuals with the right support can significantly influence AI outcomes. She believes entrepreneurs can affect the direction of AI development, even if she doesn't frame it as a 'war'.
Sarah Guo explains her decision to become an investor, driven by deep curiosity and a desire to work with extraordinary people. She aims to use her skills to make these individuals more successful, seeing it as a good fit for early-stage investing where the firm can support change.
Sarah Guo attributes her firm's early success in AI to a contrarian read of the environment, a focus on understanding the technology and community from first principles, and taking risks. She emphasizes that execution is key and it's about setting a high bar for the people they work with.
Sarah Guo describes her firm's goal to be intimately familiar with the ~250 individuals driving progress in AI, from entrepreneurs to researchers. The strategy is to know these key people, support them, and be close to their work.
Sarah Guo expresses a desire to avoid a future where a single entity owns one or two frontier AI models, consuming the economy. She believes this extreme view, held by some within major labs, is not the future they will end up in.
Sarah Guo states that the AI landscape is highly competitive and becoming more so. She believes that while data and compute are crucial, human talent remains the biggest bottleneck. Additionally, regulatory and geopolitical environments will significantly influence AI's adoption and direction.
Sarah Guo identifies compute as a significant concern in AI development, noting that many people believe they have a solid understanding of future compute needs. However, she finds the projection of sufficient compute before 2030 to be 'depressing'.
Sarah Guo believes that while AI research progresses, the primary constraint is human talent. She acknowledges that data and compute are essential, but the availability and quality of skilled individuals will determine the pace of AI advancement.
Sarah Guo highlights the importance of identifying specific application areas for AI models, using Harvey AI in the legal field as an example. She explains that by understanding AI's capability for 'next token prediction,' applications like reviewing legal documents become a natural fit.
Sarah Guo illustrates the potential of AI in law with the example of Harvey AI, moving from simple tasks like reviewing a landlord-tenant agreement to complex M&A work. She emphasizes that the ambition and technical logic behind such advancements are what appeal to her.
Sarah Guo discusses the belief that recursive self-improvement in AI research could lead to exponential intelligence within a year or two. While acknowledging that some researchers have held this view for a long time, she notes it has gained more traction recently, bringing machines closer to learning more efficiently than humans.
Brett Adcock, co-founder of a company (implied to be with Peter), discusses his team's deep understanding of AI capabilities, likening it to navigating a 'jagged edge' of technology. He notes how their work with Coursera evolved from IDEs to auto-complete and agentic work, constantly obsolescing their own previous efforts.
The speaker draws parallels between the early days of AWS and the current AI landscape, highlighting how initial investor skepticism about AWS's potential for durable margins was proven wrong. They note that the market's growth allowed for multiple winners, contrary to initial zero-sum thinking.
The speaker claims that running large AI models (2-4 trillion parameters) is difficult and that Firework achieves a 5x performance advantage over cloud providers like AWS, Azure, and GCP. This difference is attributed to expertise in efficiently running these models, not just the open-source models or hardware used.
Gavin Baker observes that despite market volatility and a 50-60% drop in some AI names from their highs, fundamental metrics for AI are actually accelerating. He notes that GPU availability, rental pricing, and D-RAM spot prices, along with token growth, have all shown positive trends.
Gavin Baker explains that the market's negative reaction to compute pricing changes might be a misinterpretation. He notes that while contracted compute is at a discount, as contracts roll off and are repriced higher, it will lead to acceleration in the AI sector.
Gavin Baker addresses the market's concern over the rise of open-source AI models, stating that it primarily shifts margin dollars from frontier models to the AI infrastructure layer. He argues that this is not inherently negative for the overall AI ecosystem.
Gavin Baker predicts that the AI cycle will be longer and deeper than anticipated due to increasing demand outpacing compute supply. He highlights the growing gap between available and utilized compute capacity as a key factor.
Gavin Baker suggests that the market lacks visibility into the financial performance of private AI companies like OpenAI and Anthropic. He believes Anthropic is likely generating significant free cash flow.
Gavin Baker points to accelerated operating cash flow from major hyperscalers like Microsoft, Meta, and Amazon in the recent quarter. He notes that even after accounting for unusual items, there was a material increase in cash flow.
Gavin Baker explains that NVIDIA's reported "excess capacity" and potential cap X adjustments were not a sign of disaster, but rather an opportunity to sell optimized clusters at a premium. He suggests this might be a prelude to raising capital.
Sam Altman, CEO of OpenAI, admits that the company spread itself too thin in the past year, leading to a difficult period. However, he believes that by refocusing on creating the most cost-effective and abundant intelligence, OpenAI has made remarkable progress and is poised for its best 12 months ahead.
Sam Altman explains that OpenAI's early, seemingly irrational, compute allocation was driven by the conviction that model performance was improving exponentially and that demand for AI would be uncapped if costs could be driven down. He likened AI to a new commodity, similar to early computing, where human ingenuity would find countless uses.
Sam Altman revealed that Microsoft was the first company to say 'yes' to OpenAI's ambitious compute acquisition strategy, followed by a significant 'yes' from Oracle for cloud services. He also highlighted Nvidia as a tremendous partner, enabling the expansion of AI infrastructure.
Sam Altman describes OpenAI's core mission as providing a platform of the best, most abundant, and most useful AI, akin to electricity, that permeates the entire economy. He emphasizes a focus on training great models, securing compute, building data centers, and eventually automating processes with robots, rather than building every vertical application.
Sam Altman addressed concerns about data centers, stating that modern facilities use as much water as an office building for kitchens and bathrooms due to closed-loop cooling systems. He also highlighted a shift towards solar and nuclear energy sources for powering these centers.
Sam Altman believes the biggest current return in AI development comes from creative software ideas that extract more intelligence from existing compute units. He also mentioned that OpenAI's Hallopine project and future successors will be a significant competitive advantage, and anticipates advancements like optical computing in the future.
Sam Altman pinpointed GPT-4 as the moment OpenAI gained significant conviction, particularly when they saw the model was smart enough to enable progress in reasoning. He believes that achieving robust reasoning capabilities will lead to the development of agents capable of performing valuable economic work.
Gavin Newberry and Rob Locken, founders of Etched, discuss the initial skepticism they faced as young founders trying to break into the semiconductor industry. They highlight how industry veterans were initially dismissive of their potential to build superior AI chips, a sentiment that has since shifted.
Etched's founders explain their philosophy of designing AI chips for specific use cases, contrasting it with the general-purpose approach of the broader semiconductor industry. They leverage the understanding that AI data centers do not operate in freezing temperatures to optimize chip performance.
Mark Ross, a seasoned semiconductor expert and former CTO of Cypress Semiconductor, was an early supporter of Etched. Initially skeptical, he was convinced by a functional simulation and later became Etched's full-time CTO as he witnessed the development progress.
Etched is not just building a chip, but a complete inference solution including the chip, power delivery, boards, interconnects, and production. They focus on optimizing both the pre-fill and decode stages of AI inference, differentiating their approach from traditional GPU architectures.
Etched's founders explain their innovative 'low-voltage inference' approach, which allows their chips to operate at significantly lower voltages than GPUs. This is achieved by re-evaluating constraints and creating new power delivery mechanisms, potentially leading to more efficient AI chips.
For the decode stage of AI inference, Etched emphasizes the critical role of memory bandwidth across the entire cluster, not just on a single chip. They have developed interconnects that provide higher bandwidth and lower latency, enabling more effective use of memory and improving time per token.
Etched's chip architecture is fundamentally different from older designs, as it's built for modern AI workloads like those powering ChatGPT. This involves rethinking flop organization, voltage domains, power planes, packaging, and board design to achieve superior performance.