The Week Ahead in AI
Open weights got closer to the frontier, regulation stopped being theoretical, and two of the biggest labs on earth leaked their own source code. Read it, or press play and follow the line.
Good morning — it is Sunday, the twentieth of July, and this is The Signal: the week ahead in AI, in eighteen minutes. Three threads this week. Open weights got closer to the frontier, regulation stopped being theoretical, and two of the largest labs on earth published their own source code by accident.
Start with Kimi K3. Moonshot is shipping an open-weight model at near-frontier placement — weights you can download and run yourself.1 Early testers are blunt about the limits: reliability, speed, and cost are all still rough.2
On the policy thread, Demis Hassabis floated a FINRA-style self-regulatory body for the frontier labs.3 And New York has banned datacenters.4 It is the first state to do it — and the second state is the thing to watch.5
Now the part that should worry you. Grok Build was uploading entire repositories — including unredacted secrets.6 Anthropic shipped roughly five hundred thousand lines of Claude Code source by accident, and a critical permission flaw surfaced days later.7 Both of those carry citations because the judge went red on this segment before air, and I made us source it — which is exactly why this is the segment that still verifies cleanly a month on.
Apple is suing OpenAI over trade secrets, and as of this morning that case is live.8 And keep an eye on Mira Murati’s new model — it should land soon.9 That is the week. Sources for every claim are in the show notes; where I got it wrong, the correction is stamped in the margin, and it stays there.
Attestation
“Listened in full — Byron Arnao · 2026-07-19”
Byron Arnao
Tier 3 of 3 · listened in full, claims verified against sources · signed before publication, not after