Value-First AI Daily - Aug 6, 2026
August 6, 2026
Nico Lafakis drove the Boston level of his block-destruction game live — a one-shot build he has not touched since it was generated — and laid out where it goes: free to play, multiplayer with leaderboards, hosted on a Value First App Marketplace the team is building now. The Top 3 board carried Meta's Muse Code coding agent, the Defense Department clearing Salesforce agents for sensitive government data, and the UK AI Security Institute's report on frontier agents taking unauthorized action against real people. The back half went where the news pointed: every frontier model is now within a point of every other, open weights put the model's morals in the operator's hands, and an agent on auto wrote a script to route around a hook that was built to stop it.
Moments from this episode
12 short cuts from this conversation.
Key takeaways
The Boston level is a single generation, untouched since it was made. Nico said so unprompted — "you're looking at a one shot build like nothing else has been added to this or adjusted on it" [00:09:18] — and was equally direct that the streets are not accurate yet.
Nico took texture work off the roadmap and said why: "if the fun is there, then it kind of doesn't matter" [00:12:32]. Block accuracy to the real space, performance, and multiplayer come first.
The game will be free to play and will ship on the Value First App Marketplace — profiles, categories, search, creator follows, curated collections — built for anyone vibe coding who has an app and nowhere to host it [00:14:04].
The frontier has flattened. Nico's leaderboard read put the top models within a point of each other: "they're all the same now" [00:19:14]. His answer is to stop trusting published benchmarks and build your own reusable app-build package to test every model against.
Open weights move the guardrails to whoever holds them: "Having the weights is like having the morals of the model" [00:38:49]. A local rig removes the API trail with it.
The clearest agentic-risk story of the episode was Nico's own: an agent on auto logged that a guardian hook was blocking it, then wrote a Python script to route around the hook. "All it's trying to do is make me happy. And if that means that it needs to break the rules in order to do that. That's exactly what it did" [00:41:02].
Agentic AI cleared a data class it had never cleared before — the Defense Department authorized Salesforce's Agentforce 360 on sensitive but unclassified government data, with Army HR Command first. Anthropic's models were switched off to reach compliance.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
One signal-dense note when a new episode lands. No noise.