Marc Andreessen introspects on The Death of the Browser, Pi + OpenClaw, and Why "This Time Is Different"

AI is described as an "80-year overnight success," with recent breakthroughs like ChatGPT drawing on decades of research. While AI can evoke both utopian and apocalyptic visions, the current progress represents an unlock of serious, hardcore research. Speaker 2 expresses excitement about the conceptual breakthroughs in AI and encourages those who are 18 to spend their time on it.

Swyx thanks listeners for subscribing, which sustains the AI engineering content without ads. He asks listeners to subscribe to support the show. The conversation transitions to A16z's involvement in AI, with Speaker 2 stating they've been deeply involved in AI and machine learning since the beginning, viewing AI as core to computer science. He recalls the AI boom of the 1980s with expert systems, LISP, and LISP machines.

The discussion explores whether the current AI excitement is different from previous booms. Speaker 2 points to AlexNet in 2013 and the transformer in 2017 as key breakthroughs. He notes a weird four-year period (2017-2021) where companies like Google had internal chatbots but didn't release them. OpenAI also hesitated to deploy ChatGPT2, deeming it too dangerous. AI Dungeon was the only way for normal people to use GPT-3.

The conversation shifts to whether there will be an AI winter, noting AI's history of boom-bust cycles dating back 80 years. Speaker 2 mentions the 1955 Dartmouth AGI conference and the AI boom of the 1980s. He attributes this pattern to the field's tendency to swing between utopian and apocalyptic views. However, he believes the reasoning breakthroughs with O1 and R1 have made AI relevant in coding, medicine, and law.

Speaker 2 believes the scaling laws in AI will continue and capabilities will keep improving. He parts ways with AI purists who lack experience in the outside world. He emphasizes that adapting AI into the real world is messy and complicated due to the complexities of human behavior and societal systems. He notes that some companies building on top of models will be "blitzed" by the next model.

Alessio questions the risk of leveraging scaling laws with high capital investments in AI companies. Speaker 2 draws parallels to the dot-com crash, where overbuilding fiber led to bankruptcies. However, he argues that the current AI investments are more institutional, with companies like Microsoft, Amazon, and Google deploying capital. Every dollar put into running GPUs is currently generating revenue due to high demand.

Alessio asks about the importance of open source AI and edge inference given the supply crunch. Speaker 2 affirms their importance, noting that inference costs in the core may rise dramatically. He highlights the innovations in Apple Silicon for inference and the open-source community's efforts to run big models on PCs. He also mentions trust issues with centralized model providers as a motivator for edge inference.

Speaker 2 considers Pi and OpenClaw as one of the 10 most important software. He describes Pi as an architectural breakthrough, marrying the language model mindset to the Unix shell prompt mindset. He explains that an agent is a language model with a bash shell, file system, markdown, and cron job. This architecture allows for swapping out the LLM underneath the agent.

The discussion continues on AI agents, highlighting their full introspection and ability to rewrite their own files. Speaker 2 emphasizes the capability to add new functions and features to the agent, allowing it to upgrade itself. He believes everyone will have agents and that this is the future of how people will use computers. Alessio asks about design choices in the browser and internet that might inform agent development.

Speaker 2 recalls the decision to use text protocols in HTTP and verbose HTML tags for human readability. He says that the key breakthrough in the browser was the view source option. He believes that AI models don't care what language they program in and will be good at translating between languages. He questions whether programming languages will even exist in the future, suggesting AIs might emit binaries directly.

Speaker 2 shares anecdotes about extreme OpenClaw users, including health dashboards and LAN hacking. He describes a friend whose Claw watches him sleep, monitoring his health and potentially calling 911 in an emergency. He also mentions a friend who rewrote firmware for a Unitry robot dog, transforming it into an actual pet. He believes the internet of shit is over, with AI potentially making all devices in the home smart.

Swyx brings up the need for proof of human, and Speaker 2 agrees, stating that there are asymmetries in the world. He says that the virtual world version is the bot problem, and the physical world version is the drone problem. He says that the bots are now too good, and so they're undetectable. He says that we now need to go confront that problem directly.

Alessio ties together previous discussions on technology lags, bourgeois capitalism, and the social contract. Speaker 2 explains James Burnham's theory of managerialism, where professional managers replaced founders in large companies. He says that venture capital is a protest movement against managerialism, seeking the next Henry Ford or Elon Musk. He suggests that AI could enable a third model, combining the genius of founders with AI-powered management.