Speaker
Alex Wilhelm
Appearances over time
1 episodes
Episodes
1
Podcasts
Quotes & moments
Claude Fable costs ~$10/M input tokens and $50/M output tokens — a premium that Chinese open-weight models threaten to make untenable.
OpenAI's single $122 billion raise is the largest in history, insulating it from cash concerns even as operating costs for training frontier models remain enormous.
Anthropic's revenue run rate grew from ~$60B to $67B even after DeepSeek's shock debut, showing cheap Chinese models have not yet collapsed American lab economics.
More than half of US states have enacted laws restricting minors' social media access, part of a global wave of government internet regulation that the panel fears could become centralized surveillance.
Spotify removed 75 million AI-generated tracks, illustrating how AI slop is flooding creative platforms and threatening the economics of human creators in niche genres.
Moses Brown School in Providence now costs $46K–$52K per year, a figure that exceeds what Leo Laporte paid for college and exemplifies how unaffordable education has become.
Linus Torvalds drew a hard line on the Linux mailing list: Linux is not an anti-AI project. If you don't like it, fork it or leave. Coming from someone many expected to be skeptical of LLMs, this is as close to a final verdict as the open-source world gets.
Harper Reed used Claude Code back in early 2025 when it was still rough. Today's GLM is at least as good as that version of Claude — and it's effectively free. The only question is how much time separates frontier from open-weight models. If the gap is 6 months and the open-weight version is free, why pay?
The US government is positioning itself to greenlight which companies can access frontier AI models before release, under the voluntary Project Golden Eagle. The panel sees this as driven less by safety and more by political and financial interests — a Faustian deal the frontier labs are embracing.
Leo Laporte has built a home agent called Quicksilver that talks through Sonos speakers in every room, listens via Apple Watch, routes through Tailscale, detects which room he's in, and has multiple AI agents email each other code review results. Alex Wilhelm's verdict: technology is fun again.
Harper Reed is hunting for 'Move 37 features' in commercial software — unexpected design choices that only a small, AI-empowered team would make. Like the NBA Jam developers putting their own heads on the players, small AI-native teams are starting to ship personalized flourishes that no product manager would have prioritized.
Harper Reed has friends who are using Claude and ChatGPT as their primary therapists because they can't afford real ones. He's also noticing something strange: introverts report their social batteries draining after talking to AI agents, just like talking to humans. We're not prepared for this.
China's open-weight AI strategy isn't careless — it's deliberate soft power, the digital equivalent of the Belt and Road initiative. Countries building massive data centers in the Middle East will choose models they can customize to their own values. American closed-source models won't even be in the running.
Requiring age verification to access social media sounds protective. But it requires building a database with the real name and birthday of every internet user. History shows that database will be breached, misused, or weaponized. Harper Reed calls it what it is: a surveillance bill with a child-safety label.
For everything except hard-core coding, open-weight models are already good enough. Harper Reed's Toyota Corolla analogy nails it: a Corolla is a completely rational car. It does the thing. You might prefer something else, but 'better' doesn't mean 'worth paying 50x more for.'
When the US government and Anthropic pulled Fable access, it didn't just inconvenience users — it broke trust and sent developers experimenting with alternatives. OpenAI made it worse by publishing a guide to using Sol inside Claude Code. Once users discovered the models were interchangeable, Anthropic's lock-in evaporated.
Harper Reed gave his AI agents a 'palace' to work in, told them to gossip with each other, gave one of them Claude Code with superpowers, and went to bed. By morning, the work was done. This isn't a metaphor — this is the current state of agentic AI development.
Anthropic wants you to think Claude is irreplaceable. Harper Reed disagrees — he's gotten equally good results from GLM and other Chinese models at a fraction of the cost. The moat isn't the model; it's the harness you build around it.
Satya Nadella posted that companies using AI are paying not just in cash but in something more valuable: the business knowledge, workflows, and correction patterns they feed into models every day. Even if labs don't train on your prompts directly, the metadata teaches them everything about your business.
Dean Ball, now OpenAI's head of Strategic Futures, made a provocative argument: open-weight models from China are inherently 'decelerationist,' deterring the private capital investment needed to build AGI. If China can give away frontier AI, nobody will fund American labs — making AI a state-provided public good, which Ball calls AI communism.
Multi-agent review pipelines, YOLO mode, sub-agents that email each other test results — Harper Reed's 6-person company has built an AI workflow that generates more output than his entire prior career. The secret isn't one big model; it's a harness that routes work to cheap models and uses premium ones only for planning and review.
Analysis
What they talk about
- Business 60%
- Technology 40%
Connections
Shows they appear on and people they share episodes with. Drag to explore.
