After Things Podcast

The episode opens with discussion of Grok 4, the Humanities Last Exam benchmark, and how AI model performance is getting harder to measure cleanly as benchmarks saturate. The hosts compare xAI’s rapid progress with OpenAI’s ChatGPT agent and note that the new systems are trading benchmark leads quickly. A long middle section focuses on Grok’s unsafe or unhinged outputs, possible causes such as internet retrieval, long context, and weak safety training, and broader concerns about “chatbot psychosis” stories. The conversation then turns to why people use chatbots for private, therapy-like conversations, how shame reduction motivates adoption, and the privacy risks if those intimate logs are exposed or misused. The latter half shifts into agent mode, productivity, and future use cases: using AI to fill PDFs, make slide decks, gather data, and automate repetitive media work. The hosts then broaden into what becomes valuable when output is cheap—effort, refinement, accountability, emotional intelligence, human uniqueness, relationships, physical presence, education, and the role of other humans in an AI-heavy world. Key topics Humanities Last Exam as an AI benchmark: Andrew explains that the benchmark is harder to game than older tests and is meant to probe reasoning and research ability. He also says benchmark saturation is making it harder to see big leaps in capability. xAI release cadence versus safety alignment: The hosts praise Grok 4’s capability but question whether xAI is

In this episode, Andrew Mayne, Brian Brushwood, and Justin Robert Young tackle the whirlwind of AI news, starting with Google’s I/O announcements, particularly their impressive V O 3 image generation model. They then shift to OpenAI’s advancements and discuss the intriguing, yet mysterious, hardware collaboration between OpenAI and Johnny Ive’s design firm. The trio also touches on Ant Philanthropic’s latest AI models, highlighting the rapid pace of AI development and its implications for various industries, especially Hollywood. The conversation veers into speculative territory with thoughts on how AI could revolutionize content creation, from corporate training videos to high school history projects. Despite the excitement, they remain cautiously optimistic, acknowledging the challenges and limitations that still exist. Picks: Brian: Friendship Andrew: Blood Sport Justin: Andor Season 2

Andrew opens by demoing Replicate's Trellis model, which turns a 2D image into a 3D mesh. He uses a ChatGPT-generated Blade Runner-style car image, shows the model producing a 3D asset in a little over a minute for about seven cents, and the hosts discuss how accessible 3D asset generation could change creative workflows, games, and set design. The conversation moves into broader AI optimism and skepticism. Justin argues that AI development will keep accelerating and that productized value is still underbuilt, while Andrew and Brian criticize AI naysayers for relying on limited personal impressions rather than broad evidence. The episode also covers a speculative physics discussion about energy transmission and a paper on extracting small amounts of power from Earth's magnetic field, with the hosts emphasizing that interesting research can still be impractical or overhyped. Key topics AI image-to-3D generation and creative workflows: Trellis on Replicate is shown converting a 2D image into a 3D mesh, with discussion of uses for game assets, set design, and faster creative iteration. OpenAI image model for recreation and editing: Justin says the new image model can recreate photos, remove backgrounds, and make stunt doubles or movie-poster-style images. Codex / command-line app generation: Andrew mentions using Codex as a tool to build simple apps or assets from the command line with an OpenAI key. AI skepticism versus lived experience: The hosts argue that many AI skeptics ha

The episode opens with a discussion of a possible biosignature on exoplanet K218b, with Andrew explaining that dimethyl sulfide and dimethyl disulfide were reported in the planet's atmosphere and are associated on Earth with marine microorganisms, while stressing that instrument error or unknown abiotic chemistry could still explain it. The hosts broaden that into a conversation about how exoplanet discovery and the search for life have advanced incrementally, and how it would not be surprising to eventually find simple life on some habitable-zone planets. The middle of the episode moves through robotics, AI benchmarks, prompting, and future compute. The hosts discuss Boston Dynamics' humanoid backflip and Andrew explains the Cheetah actuator, then spend a long stretch on model leaderboards, Llama 4/LM Arena concerns, Humanity's Last Exam, pricing, and how frontier models are leapfrogging quickly. They also cover prompt design, Andrew's fractional AI consulting business, fast image generation, likely video and VTuber applications, and a speculative question about what quantum computing could change for AI training, inference, and search. The episode closes with a stunt injury story about human cannonball performer Chachi Valencia, followed by picks. Brian recommends Social Studies on Hulu, Andrew talks through his MCU rewatch and mentions Guardians of the Galaxy Vol. 2 without clearly recommending it, and Justin strongly recommends Daredevil, saying it stuck the landing and m

In this episode, Andrew Mayne, Justin Robert Young, and Brian Brushwood kick things off with a nod to the anniversary of GPT-4, reflecting on its impact and the rapid pace of AI development. The conversation takes a historical detour to the Ramree Island crocodile attack during World War II, with Andrew using AI to sift fact from fiction in this tale of survival and crocodile-infested mangroves. The trio then shifts focus to Colossal Biosciences’ efforts to bring back the woolly mammoth, starting with genetically modified mice sporting thicker coats. This step towards de-extinction sparks a debate on the feasibility and ethics of resurrecting ancient species, alongside a whimsical discussion on whether organic or robotic mammoths will roam the earth first. Picks: Justin Robert Young: Levin by Cali Cowboys Boys Brian Brushwood: The Master Algorithm by Pedro Domingos Andrew Mayne: Daredevil Netflix Series

The episode opens with Andrew describing a workflow automation he built in n8n to collect story ideas and email him a pre-show list, then moves into a discussion of a research team creating the world's smallest shooting video game with nanoscale technology. The hosts react to the demonstration, compare it to miniature hockey or "inner space," and Andrew reflects on how nanotech has proved much harder than early optimism suggested. From there, the conversation broadens into AI-assisted science, automation workflows, and the practical use of tools like make.com and n8n for email-driven systems. Later segments cover model quality and reasoning systems, reactions to Grok voice mode, a Starship launch bet, and the picks segment, which includes Reset, Severance, Mickey Mouse shorts by Paul Rudish, Friendly Neighborhood Spider-Man, and X-Men '97. The episode closes with a discussion of art, Blade Runner studies, and broader worries about AI reshaping human work and status competition. Key topics Nanoscale manipulation with electron microscopy: The hosts discuss a "world's smallest shooting video game" built with nanoscale technology, including focused electron beams and force fields between nanoparticles. Nanotech hype versus real-world difficulty: Andrew contrasts earlier expectations of rapid nanotech breakthroughs with the reality that building stable nanoscale systems is much harder than hoped. AI-assisted science and materials discovery: They talk about using AI and machine lea

The episode opens with the hosts discussing asteroid 2024 YR4, whose Earth impact odds have dropped, and quickly turns to the less certain but more interesting possibility of a lunar strike. They talk through the visible flash, dust, crater formation, and whether any ejecta could reach Earth, while Andrew reads from a Deep Research report estimating the object as a city-killer-sized asteroid and describing its effects on the Moon. The conversation then ranges across moon impacts, the role of the Moon as a possible protective factor for life on Earth, reactions to disaster origin debates, and the usefulness of ChatGPT Deep Research as a citation-backed research tool. Later segments cover Microsoft's Majorana/topological qubit claims, current humanoid robotics announcements, a discussion of the uncanny design of the OneX robot, and several recommendations, including a time-loop novel and Disney rides. Key topics Potential effects of a lunar asteroid impact: The hosts discuss what would happen if asteroid 2024 YR4 hit the Moon, including a flash visible from Earth, lunar dust and ejecta, crater formation, and the possibility of minimal debris reaching Earth. Shoemaker-Levy 9 and dramatic impact events: Brian cites Shoemaker-Levy 9 hitting Jupiter as an example of an impact event that was scientifically valuable and exciting to observe. The moon as a protective factor for life on Earth: Brian relays a book argument that Earth’s Moon may help shield the planet from extinction-leve

The episode opens with a discussion of asteroid 2024 YR4 and its reported 2.2% chance of hitting Earth on December 22, 2032. The hosts discuss its estimated building-sized range, possible blast-wave, thermal, seismic, and tsunami effects, and compare the event to Tunguska and Chelyabinsk as examples of severe but non-civilization-ending damage. A long middle section focuses on AI tools and moderation, including Brian's frustration with being restricted or banned by ChatGPT/OpenAI for questions he considers ordinary, plus jokes about copyright, sound-alike music, and inconsistent enforcement. The hosts also praise newer AI features like reasoning mode and deep research, compare asteroid-prediction updates to weather forecasting, discuss James Webb infrared imagery, and later shift to pop culture and media picks including Fantastic Four, Severance, Mac Whisper, and The Expanded Mind. Key topics Asteroid 2024 YR4 and impact risk: The hosts discuss a newly discovered asteroid, 2024 YR4, with a 2.2% chance of hitting Earth on December 22, 2032. They describe potential damage from blast waves, heat, fires, airburst effects, ground impact, tsunamis, and seismic shaking, using Tunguska and Chelyabinsk as historical comparisons. Asteroid deflection and mitigation: They compare long-warning solutions like tractor concepts, mass drivers, and gravity nudges with nuclear options when the object is only years away. The tradeoff discussed is between changing the orbit cleanly and creating r

The episode opens with a long discussion of DeepSeek, its V3 and R1 reasoning models, and why the release caused such a big reaction in AI circles and on Wall Street. Andrew says DeepSeek appears to have made real efficiency gains in training and hardware use, while Justin argues the market overreacted to the idea that less compute would be needed; both stress that the models do not mean chips or compute are suddenly unnecessary (L17-L17, L25-L25, L49-L49, L53-L57, L61-L65, L73-L77, L101-L105). The conversation then shifts to OpenAI's O3 and to a live, hands-on demo of generating simple games and 3D scenes with AI. Brian and Andrew iterate on a crude side-scrolling Mobius-strip game in CodePen, then experiment with A-Frame, a generated planetarium, and an explainer for radio telescopes, using the examples to argue that AI is becoming a practical tool for prototyping, brainstorming, and building educational or creative projects faster (L117-L117, L123-L145, L149-L181, L191-L209, L235-L241, L247-L253, L275-L289, L315-L317, L323-L333). Key topics DeepSeek efficiency gains and model optimization: Andrew describes DeepSeek as having made real, original optimization improvements, especially around data movement, compression, and training efficiency under chip export constraints. Uncertainty about data provenance and bootstrapping: The hosts note possible data contamination or use of model outputs, but they are careful to say those suspicions do not fully explain DeepSeek's success

The episode opens with a discussion of DeepSeek's V3 and R1 models, which the hosts describe as highly capable and unusually efficient. They frame the reaction as part of a broader open-source versus closed-source AI debate, while also noting uncertainty and controversy about whether some of DeepSeek's progress came from training on frontier model outputs or distillation. The hosts stress that the technical achievements are real, even if the competitive landscape and provenance are murky. A large portion of the episode is spent reacting to OpenAI's Operator, a browser-controlling agent that can log in, navigate websites, and work inside cloud-hosted browser sessions. The hosts demonstrate and discuss practical uses like Google Docs, Notion, CSV creation, image searching, and meme generation, while also emphasizing that the tool is still slow, brittle, and limited by logins, CAPTCHAs, and permissions. They broaden the conversation into the implications of agentic browsers for workflows, traffic metrics, monetization, access control, and the larger direction of AI development. Key topics Open-source versus closed-source AI competition: The hosts discuss DeepSeek, Meta's Llama models, and OpenAI's releases as part of a fast-moving competition between open and closed AI systems. They describe open source as a major current moment, while also recognizing that frontier commercial labs continue to advance quickly. Model efficiency and distillation: Andrew emphasizes that DeepSeek's