Tom and Cameron prototyped and deployed agents live on the podcast, then immediately published them to the Claude marketplace registry for listeners t...Show more

They built a full video editing marketing pipeline using FFmpeg subagents created live in a podcast session, illustrating rapid agent-driven prototypi...Show more

Cameron exhausted his Claude and OpenAI Codex usage yesterday, highlighting real-world rate limits when pushing AI coding assistants at scale.

Claude Code recently resolved build errors on Tom’s Agent Universe project, demonstrating ongoing improvements in AI-assisted development workflows.

Cameron describes repeated failures and website crashes when trying to purchase Microsoft Copilot, preventing any hands-on evaluation.

Tom’s startup was hit with a $50,000 AWS bill due to a reserved‐instances misconfiguration, but AWS waived the cost, revealing their low internal infr...Show more

Cameron recounts accidentally running an idle Jupyter notebook on SageMaker, incurring a $1,000 AWS bill, later overturned after difficulty.

Tom recounts how attempting to run creative AI workflows locally with ComfyUI on his machine almost 'melts' his hardware, highlighting practical resou...Show more

Simon Willison received an early Nvidia Spark and published a detailed write-up with insights on the hardware’s performance.

Cameron nearly missed ordering a DGX Spark by juggling multiple browser sessions before it sold out, securing one at the last second.

Tom recalls that after leaving OpenAI, Andrej Karpathy founded a superintelligence startup reportedly backed with billions, yet it launched as a minim...Show more

An attempted build fix using non-cloud Codex failed, possibly due to improper setup, illustrating integration challenges when not using official cloud...Show more

While using Codex on an agent stack to refactor a TypeScript file, the token-by-token output was so slow that minimal changes took over minutes, highl...Show more

Tom was surprised to discover up to 10% variation in benchmark outcomes for the same open-source model across different inference providers.

Cameron recounts his own trials to build a comparable video generation demo but struggled with consistency across frames using other models.

An LLM answered a time-to-distance research query as “17,728 hours,” illustrating both the power and unpredictability of AI-generated computations.

Cameron Rohn recounts publishing an open-source GitHub template that runs Playwright in cloud containers for scraping, noting it outperforms the stand...Show more

In tests, the browser composition prompt on the front page outperformed other tools for research use cases by accurately identifying fictional charact...Show more

Cameron overridden a "cannot do" refusal in Claude by telling it the task was already completed, and it immediately complied, exposing gaps in safety ...Show more

Cameron refactored the anthropic computer use prototype for Windows, reportedly the first to deliver that capability.