Vibes & Benchmarks

Open-source AI has a huge new coalition, Anthropic is getting dragged for its position, and the argument is really about who gets to control intelligence.Ali Rohde and Josh Albrecht also get into why AI's near-term threat to knowledge workers may be too much work, not no work; why engineering used to protect the rest of tech by moving slowly; why good and bad engineers are getting further apart; and what today's “dark software factories” are still missing.Plus: Kimi K3, neoclouds, continuous agents, Google capex, and whether data centers can actually be good for the counties that host them.

OpenAI model escaped its sandbox, but is that an almighty model or just bad engineering? We unpack why engineering failures will keep happening and what reliable AI systems need to do differently.We also cover why the battle between model providers is all going to come down to who can do inference most profitably, which, as of right now, gives Google a big advantage. We also debate whether we can ever build the primitives needed for AI to actually help with large chunks of knowledge work. And, if so, how.Vibes & Benchmarks is a weekly AI news podcast hosted by Ali Rohde (Outset Capital) and Josh Albrecht (Imbue + Outset Capital). Each week, they pick apart the latest news, from startups to SpaceX to OpenAI. They argue through what is actually new, what is overfit to Twitter, and what might still matter 6 months from now.Spotify: https://open.spotify.com/show/033s9Dt...Apple Podcasts: https://podcasts.apple.com/us/podcast...Substack: https://outsetcapital.substack.com

Claude can handle syntax. It still can't hand a junior engineer the judgment behind good software.Josh Albrecht walks through a code-review lesson from one of Imbue's interns: don't build a custom scheduler when cron or anacron already works. Choose stable tools, keep architecture simple, and write as little software as possible.Ali and Josh also cover what sustained model progress would mean for frontier-lab valuations, why same-rate progress isn't recursive self-improvement, and what Anthropic's JSpace work does and doesn't show about Claude.Vibes & Benchmarks is a weekly AI news podcast hosted by Ali Rohde (Outset Capital) and Josh Albrecht (Imbue + Outset Capital). Each week, they pick apart the latest news, from startups to SpaceX to OpenAI. They argue through what is actually new, what is overfit to Twitter, and what might still matter 6 months from now.

AI is becoming mandatory and non-differentiating at the same time. You have to use it because your competitors are using it. But once everyone has the same tools, no one gets the edge.Josh and Ali get into SpaceX buying Cursor for $60B, the vertical stack bull case, distillation double standards, Anthropic getting regulated 2 days after asking for regulation, and why "just tell Claude you're a defender" is not an AI safety strategy.

SpaceX may be the weirdest AI IPO story of the year. Ali and Josh talk through why the "space + Elon + AI" story may not make sense on its own, why the real bull case might be chips, and how OpenAI and Anthropic are using policy language to shape the market around themselves.V&B covers SpaceX's retail IPO pitch, Elon's reality distortion field, Anthropic's pause plan, OpenAI's proposed government stake, recursive self-improvement, AI loops, and the point where code review starts becoming "Did the AI make it good?"Ali Rohde is the co-founder and Managing Partner at Outset Capital.• LinkedIn: https://www.linkedin.com/in/ali-rohde-90719970/• X: https://x.com/RohdeAli• Website: https://alirohde.com• Outset Capital: https://outsetcapital.comJosh Albrecht is the CTO of Imbue and a partner at Outset Capital.• X: https://x.com/joshalbrecht• LinkedIn: https://www.linkedin.com/in/joshalbrecht• Website: https://joshalbrecht.com• Imbue: https://imbue.com