Podcast
Pretrained
Hosted by Pierce Freeman & Richard Diehl Martinez · EN
Everyone's talking about AI but most of it is hype, jargon, or someone trying to sell you something. Pierce Freeman and Richard Diehl Martinez are two Stanford friends who actually work in this space with nothing to sell you but good vibes. We'll make you smarter - and you might actually enjoy it.
33episodes
Episodes
Newest firstAll episodes
Claude Will Never Stop Texting You
Mar 2501:09:13Tap to summarizeAnthropic shipped Dispatch - a way to text Claude from your phone while it works on your laptop. Rich thinks it might be as big as the invention of the cell phone. Pierce thinks it's the nail in the coffin for work-life balance. They debate what happens when there's never an excuse to stop working and whether AI productivity is even real or just more code to maintain. Plus: OpenAI acquires Astral, GPT 5.4 Mini and the parameter golf challenge, Elon's TeraFab chip factory comes for TSMC, Nvidia's pivot from training to inference with Vera Rubin, and Sam Altman wants you to stare into an orb so your agent can use your credit card.
Transcribe →AI Spent the Night Improving Itself
Mar 1800:50:55Tap to summarizeKarpathy dropped a project where an AI agent spent 12 hours improving its own training loop. Then the guys get into OpenClaw spreading to China and what it means when agentic tools land in a culture that's already way ahead on adoption. Yann LeCun raises a billion dollars to prove every other AI lab is thinking about intelligence wrong. And finally, trouble at Grok: layoffs, SpaceX fixers parachuted in, and only 2 of the original 12 founders still standing. Is XAI a real frontier lab or just a very expensive piece of the Elon cinematic universe?
Transcribe →Anthropic's Beef with the Pentagon
Mar 700:41:07Tap to summarizeRichard and Pierce are back in a brand new studio, fresh off Pierce's snowless ski trip to Japan. This week: Anthropic picks a fight with the Pentagon over three words and ends up blacklisted as a national security risk. Then the guys break down speculative speculative decoding, Chinese labs have been running industrial-scale operations to scrape Claude's capabilities, and the team behind Qwen quietly starts falling apart, right as their best model yet drops.
Transcribe →The Neo Lab Arms Race
Feb 2001:14:26Tap to summarize$126 billion in AI funding has been raised across 30+ "neo-labs". Most of them will fail.Pierce and Rich map the landscape - from Thinking Machines' $12B valuation on vibes alone to Ilya's SSI. Plus: why OpenAI might be the Fairchild Semiconductor of AI, whether Stanford professors make good founders, and the Sauron's Eye problem of building in a frontier lab's blind spot.
Transcribe →Ads in the Chat Window
Feb 1901:12:48Tap to summarizeThe day of reckoning: OpenAI is putting ads in ChatGPT. Pierce and Rich unpack how it actually works, why the shopping use case might be the wrong starting point, and what happens when sub-agents try to sell things to each other. Also: OpenClaw's Peter Steinberger gets acquihired by OpenAI, Waymo builds a world model for self-driving, Mistral's Voxtral takes on Whisper, and the model release arms race hits a new peak.
Transcribe →Your Agents Need an Org Chart
Feb 1301:06:59Tap to summarizeRichard and Pierce go deep on the agent swarm era. Rich argues we're living through a paradigm shift in how we interact with AI: from copy-pasting ChatGPT to orchestrating hundreds of parallel agents. They cover Anthropic's opinionated swarm architecture, why corporate org structures might be the right mental model for agent design, the liability problem when no human reviews the code, thinking tokens as a tuning knob, and whether a 100,000-line AI-generated C compiler is impressive or irrelevant.
Transcribe →Cracking Open OpenClaw
Feb 1101:13:53Tap to summarizePierce and Richard dissect OpenClaw (née Claudebot, née MoltBot) - also known as the desktop agent everyone's installing on their Mac Minis. They cover the 400+ malicious packages that hit Claw Hub in one week, slop squatting, trust networks for open source, Kimi 2.5's agent swarm mode, and newspapers that think Waymo is controlled by joysticks in the Philippines.
Transcribe →DeepSeek's New LLM Architectures
Jan 3001:00:08Tap to summarizePierce and Richard break down DeepSeek's latest model architecture moves in Manifold-Constrained Hyper Connections and Engram memory. Are these conceptually sound? Will they hop the pond over to US frontier labs?
Transcribe →Claude Gets a New Constitution
Jan 2801:14:27Tap to summarizeRichard's finally back in San Francisco! The gang covers Claude's new constitution, reinforcement learning with AI instead of human feedback, why Apple went with Gemini for their new Siri revamp, homographic encryption, a $480m fundraising round for Humans& (note the ampersand), and much more.
Transcribe →Why Your Agent is Cheating
Jan 2101:00:59Tap to summarizePierce and Richard are back for the second listener mailbag. They break down what reward hacking really is and why models so often learn the wrong lesson, explain practical fine-tuning (from pre-training to prompting), unpack why LLMs use tokens instead of words, how context length is a hardware versus mathematic limitation, and much more.
Transcribe →