Value-First AI Daily - Aug 24, 2026
August 24, 2026 · 46:56
Episode 19 went nowhere near the plan. Nico Lafakis came back from a My Chemical Romance show having made one Claude request all weekend, Chris Carolan ran the show's own command live on a model he switched to that morning, and the board opened on Thursday's item before he caught it on air.
Moments from this episode
Key takeaways
A harness is the software wrapped around a model, and it is now enough of a variable on its own that Nvidia moved Claude Opus 5 from 30% to 100% on ARC-AGI-3 without changing the model, per the day's sealed board sourced to TechCrunch.
A perfect score is worth exactly as much as the independence of whoever ran it, and on this one the team that built the harness did the scoring in-house with the added compute and token cost undisclosed.
The other half of that argument is Nico's and it holds: ARC-AGI is a legitimate third-party test, so what got beaten is an outside puzzle set, scored by an interested party, by two models working together rather than one model getting smarter.
A promotional rate is a discount on spend you already have rather than a new cost basis, so OpenAI's GPT-5.6 Sol output cut to $20 per million tokens from $30 through November 21 is a window to run deferred work in, not a number to plan on.
Getting real work out of AI starts with what the human brings to it, context, information and guidance, which is why Nico's frame for the whole category is an executive assistant rather than a replacement.
An agent's instruction is only as good as the human following it: the board was read off a browser tab opened before the board was sealed, and Thursday's item went out live before anyone caught it.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.