New SEC Chair, Bitcoin, xAI Supercomputer, UnitedHealth CEO murder, with Gavin Baker & Joe Lonsdale

Within approximately 10 years from the time of this recording (by ~2034), China will have roughly four times (4x) the electricity production capacity of the United States.

“and that ultimately leads to a situation where in ten years, we're going to be looking across the water at, you know, a competitive country that has Rex forex. The electricity production capacity of our country” View on YouTube

Over roughly the next 50 years from the time of this recording (by ~2074), solar power will become the dominant global energy source, with a large share of the world’s energy usage running on solar; however, solar-generated electricity will still not be cheaper than nuclear power, but its cost will approach that of coal.

“I think it is highly likely that in my lifetime the world just runs on solar. Like if you just, you know, we all know compound interest is the greatest force on earth. But if you just look at the rate at which photovoltaic cell efficiency is compounding, battery efficiency is compounding. And people will make these balance of system arguments, but it will it will never be as cheap as nuclear, but it will likely approach coal. And I think a lot of the world will run on solar. But that's going to take 50 years.” View on YouTube

At some point in the future (no specific date given), Bitcoin will grow to become a serious competitive threat to the US dollar’s role, likely as a store of value and/or medium of exchange, prompting notable policy reactions from future U.S. administrations.

“But I do think Bitcoin will at some point be a serious threat to the US dollar. And that just is what it is. And we will see how different administrations react.” View on YouTube

Over the coming years/decades, the U.S. defense-industrial base will add roughly 7 to 10 new major prime contractors ("new primes"), with Anduril almost certainly being one of them and Saronic and Epirus likely candidates.

“there's going to be it's going to be like 7 to 10 new primes. One of them's obviously anduril. I think one is probably Saronic and Epirus, but we'll see. But like there's going to be 7 to 10 new primes and that's what's going to be.” View on YouTube

After the next generation of leading AI hardware is released—specifically Nvidia’s Blackwell GPUs, new AI-focused chips from AMD, and new ASICs from Broadcom, expected starting in the year following this recording (2025)—China will no longer be able to keep up with the United States at the leading edge of AI compute capability.

“but you know, Nvidia's Blackwell chip comes out next year. You're going to have new chips from AMD, new Asics from Broadcom. And I, I think at that point it is not going to be possible for them to keep up anymore.” View on YouTube

The performance of Grok 3, trained on xAI's ~100,000-GPU Colossus cluster, will provide a decisive empirical test within its first training run of whether existing AI training scaling laws continue to hold or are starting to break down.

“Grok three is a big card and will resolve this question of whether or not we're hitting a wall.” View on YouTube

Regardless of whether current training scaling laws continue to hold, there will be at least 10 more years of significant AI innovation driven by other axes such as inference-time compute, context window expansion, and architectural improvements.

“Even if scaling laws for training break, we have another decade of innovation ahead of us.” View on YouTube

If current AI scaling laws continue to hold, xAI's Grok 3 model will surpass OpenAI/Microsoft’s best publicly available frontier model and become the state-of-the-art general-purpose LLM by January or February 2025.

“Grok three should take the lead if scaling laws hold in January or February.” View on YouTube

By roughly 2026–2027, training a single top-tier frontier AI model at the cutting edge will require on the order of US$100 billion in total training cost (hardware, energy, and associated infrastructure).

“Particularly if it's going to cost $100 billion to train a model in 2 or 3 years, which I think is a realistic estimate.” View on YouTube

Within about 12 months of this December 2024 recording (i.e., by late 2025), end users will be able to specify an application in natural language and have AI systems autonomously generate, test (including QA), design and iterate the UX, and deploy a working production-ready app within a few hours, with minimal human intervention.

“You fast forward 12 months and now you've got the architecture where the AI can run its own QA testing and debugging, and the AI can run its own kind of sales and marketing and customer its own UX of the application and it can run everything. So you're basically going to say, I want this app to do this. It builds it, it tests it, it builds the UX, it tests the UX, it iterates the UX, it does everything streamlined for you. And then you show up a couple hours later and you're using a new product that was built on the fly for you.” View on YouTube

By sometime in 2025, natural human language (e.g., English prompts and instructions) will be the primary interface used by most developers and many non-developers to create and modify software, effectively making human language the dominant programming language in practice.

“I think next year, the human language will be the dominant programming language.” View on YouTube