One signal-dense note when a new episode lands. No noise.
Value-First AI Daily - Jul 28, 2026
Hosts: Chris Carolan, Nico Lafakis
July 28, 2026
Chris Carolan opened Ep. 6 with a line-by-line walk of the operating-rules file every agent on the team loads, cutting it from 58,000 bytes to 14,000 and 34 sections to 17, and named why it had grown: mistakes were being fixed by documenting them rather than removing what caused them. The sealed Top 3 followed and became the body of the show, with Nico Lafakis taking open-weights policy and chip export controls, what you can and cannot see inside a frontier lab, and his case for worrying about superintelligence labs before AI labs. Nico's city game returned with a graphics-settings pass, and Chris closed on an unplanned walk through a 3D wayfinding map of the Value-First sites.
Key takeaways
A rules file grows because of how mistakes get fixed. Calling out an error produces a written record of the error, the date, and an instruction not to repeat it — and the next rule lands on top of that. It reads like diligence and behaves like tech debt.
Ask who a context file is actually written for. Chris's first drafts came back in human-readable prose because he had said he wanted to review it; once the file was given permission to be machine-readable only, most of the words were not doing anything.
Cutting a rules file has an immediate failure signal. "You feel it immediately... oh, that's why that was in there." That feedback is what makes the cut safe to attempt rather than something to avoid.
Whatever governs your subagents is a much smaller surface than you think. They receive the operating-rules file and their own profile — everything else you have built does not reach them.
Nico's argument on the Open Secure AI Alliance: the harder problem is not security tooling, it is that you cannot inspect what your own agents are doing on a provider's back end without instrumenting it yourself and taking on the liability.
Read the primary source before the coverage. Nvidia's release on the Safe Superintelligence stake names no dollar figure; the number in circulation came from reporting, not from either company.
Setting up an accountability check on camera is only half of it. This episode set one up and did not return to state the result — visible in the tape, and worth more as an example than a clean claim would have been.
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.