Happy Friday.
The current thing in tech and business continues to be the debate raging around Claude Mythos.
Today’s lineupUCLA Human Genetics Department Assistant Professor Alexander Young at 11:40 AM
Firefly Aerospace CEO Jason Kim at 12:00 PM
Grüns Founder & CEO Chad Janis at 12:15 PM
Antares Industries Co-Founder & CEO Jordan Bramble at 12:45 PM
Pie Founder Andy Dunn at 1:00 PM
Gusto Co-Founder & CEO Josh Reeves at 1:15 PM
by Brandon Gorrell
The world’s “first” “one-person $1B company,” Meta’s new frontier model Muse Spark, Anthropic’s release of Claude Mythos (and reaching $30B ARR), and Intel joining Elon’s Terafab project were a few of the stories that dominated the tech timeline this week. But there were several thoughtful longposts on X that people have been having productive discussions about as well. They’re worth reading in full, but here’s a brief rundown of their key points:
Nic Carter lays out his case for why Adam Back isn’t Satoshi. The TL;DR: Nic thinks Satoshi is dead. From his article:
If Satoshi were alive today, they would have an enormous responsibility to resolve the biggest lingering problem for Bitcoin – the roughly 1.7m BTC sitting naked in “pay to public key” (p2pk) outputs, waiting to be seized by a quantum computer… This could entail the unexpected resurfacing to the market of nine percent of Bitcoin supply thought forever lost. The price impact would be catastrophic.
Recent papers from Caltech and Google suggest that the ability for quantum computing to break Bitcoin may come sooner than expected. About 9% of BTC supply, or $120 billion worth, belong to Satoshi and other wallets that are presumed to be lost. These BTC use outdated p2pk addresses which are uniquely vulnerable to quantum (and their users won’t fork to a quantum-resistant BTC when we get a fix).
“If Satoshi was alive, they would feel an overwhelming sense of duty to resolve the p2pk coins problem, either by rotating their coins to safer addresses, or burning them,” Nic wrote.
Dwarkesh sums up his conversation with Dylan Patel about why energy won’t be the main bottleneck to scaling AI compute. In the conversation, Dylan is somewhat handwavy about the United States’ ability to harness more energy for AI. From Dwarkesh’s write-up of the discussion:
You can do jet engines bolted to the ground. Ship engines. Diesel recips from auto manufacturers with declining volumes. Fuel cells. Each category alone delivers tens of gigawatts by end of decade. Combined, hundreds… Fundamentally, there’s a lot of different ways to bring power online over the next few years. Building more logic and memory is far more difficult and centralized, so that’s where Dylan thinks the bottleneck will be.
The take proved to be pretty contrarian as far as the discourse was concerned. Jigar Shah, the Director of the U.S. Energy Department under Biden, responded by basically saying Dylan and Dwarkesh don’t understand how the power grid works. A read through the post’s quote tweets provides a really good survey of the different interpretations of which parts of the supply chain are ultimately going to be the biggest constraints on AI — there are a bunch of thoughtful takes to read through.
Andrej Karpathy says he’s seeing a growing gap in the public’s understanding of AI capability. Everyday users of pro-tier models (e.g. $200/mo) who use them for deep research, writing, search, and advice, have been somewhat ‘left behind’ in Karpathy’s interpretation of things. While the models have continued to get better, they haven’t improved as much in these everyday tasks, so AI progress is simply less obvious to this class of users.
The second group of people Karpathy identifies are those who use frontier agentic models professionally. Andrej writes:
This group of people is subject to the highest amount of “AI Psychosis” because the recent improvements in these domains as of this year have been nothing short of staggering.
Karpathy posits a few straightforward but interesting reasons the models have improved ‘peakily’, with agentic coding advancing the most rapidly: “1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them.”
We send this email every weekday, right as our show goes live. Subscribe to get it in your inbox.
HeadlinesWSJ: Inside the Deadly Civil War That Tore Apart a Group of Chimpanzees in Uganda
The Information: SpaceX Posted Nearly $5 Billion Loss Last Year from AI Spending
WSJ: A Fiery Re-Entry Awaits the Artemis Astronauts
David Sacks jumps into the Mythos debate
OpenAI says CEO Sam Altman’s house was targeted with a Molotov cocktail
Marginal Revolution: AI, Unemployment and Work
Will Manidis builds website to track data center bans and moratoriums
Eric Seufert: The App Store discoverability squeeze
CoreWeave Announces Multi-Year Agreement With Anthropic
FT: xAI sues Colorado over first state AI anti-discrimination law
WSJ: Meta Banks on AI to Clear the Smoke of Social-Media Lawsuits
The Cut: Would You Buy Yerba Mate From Barron Trump?
Bloomberg: Anthropic Model Scare Sparks Urgent Bessent, Powell Warning to Bank CEOs
WSJ: How VC Money and Israel Outrage Derailed a Hot Hollywood Startup
Axios: What AI CEOs still don’t get about Washington
WSJ: Why Ferrari Is Unlike Any Other Luxury Brand, According to the Hosts of ‘Acquired’
Bloomberg: US Government Pitches Gamers a Career as Air Traffic Controllers
Posts of the Day