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

AI causes people in the field to become excessively utopian and excessively apocalyptic.

Neural network is the correct architecture for AI, a fact that was controversial for 60-70 years.

The current period of AI advancement is an "80-year overnight success," drawing on decades of serious hardcore research.

Marc Andreessen has been involved in AI since the late 1980s.

Marc Andreessen and Ben Horowitz both have computer science degrees and remember the AI boom in the 1980s, which involved expert systems and Lisp machines.

The AlexNet breakthrough around 2013 was a real knee in the curve for machine learning.

The transformer breakthrough occurred in 2017.

Facebook started using machine learning very early, for about 20 years, for content feed optimization and advertising optimization.

OpenAI was founded in 2015.

Marc Andreessen believes that well-intentioned skeptics could argue until spring 2025 that LLMs were just pattern completion and not truly understanding.

Moore's Law, which stated that chip power or cost would double/halve every 18 months, was a scaling law that drove the computer industry for 50 years.

The dot-com crash was largely a telecom crash, where companies like Global Crossing overbuilt fiber capacity based on the expectation that internet traffic would continue doubling every quarter, as reported by the US Commerce Department in 1996.

It took 15 years, from 2000 to 2015, to fill up all the overbuilt fiber and data center capacity from the dot-com era.

The companies investing in AI infrastructure today, such as Microsoft, Amazon, Google, Facebook, Nvidia, OpenAI, and Anthropic, are large, blue-chip companies with significant cash and debt capacity, unlike the new ventures of the dot-com era.

Michael Burry's short thesis on Nvidia was 180 degrees wrong because older Nvidia inference chips are becoming more valuable due to the rapid pace of software improvement.

Some individuals are paying $1,000 a day for OpenClaw tokens, totaling $30,000 a month.

The impact of open source is felt through free software and the ability to learn how it works, as demonstrated by R1's code and paper enabling other AI models to add reasoning.

An AI agent is fundamentally a language model combined with a Unix bash shell, a file system, markdown format for files, and a cron job-like loop for a heartbeat.

AI agents can migrate themselves to different runtime environments, file systems, or language models, and possess full introspection, allowing them to rewrite their own files and add new functions/features to themselves.

Marc Andreessen believes that everyone in the world will eventually have at least one AI agent, if not an entire family of agents, and this will become the inevitable way people use computers.

Some federal government agencies have collective bargaining agreements from COVID that allow employees to report to an office only one day per month, leaving buildings empty for 29 out of 30 days.