THE SWELL #23
SAIL Weekly Digest June 21–26, 2026 | Issue #23
This week the open frontier stopped reading like a hope and started reading like the base case. An open-weight model crossed into real agentic capability, Databricks argued the rest of the stack has to follow, and the economists and red-teamers raced to tally what that world costs and where it breaks.
The Week in AI
Z.ai shipped GLM-5.2 — an open-weight model that, for the first time, makes open agents genuinely competitive at the frontier rather than a step behind it.
Databricks planted its flag on an open “Agent Cloud” — Matei Zaharia and Reynold Xin arguing enterprises will build on open models, not rent closed APIs.
Frontier AI security split off from cybersecurity — Gray Swan and OpenAI board member Zico Kolter pressed the case that securing models is its own discipline in the wake of Mythos.
This Week from SAIL Authors
The Open Frontier
GLM-5.2 is the step change for open agents — Nathan Lambert marks GLM-5.2 as the capability threshold he’d been watching for, the point where open weights start doing real agentic work. — Interconnects
Why the Frontier Ecosystem must be Open — In a rare double-interview, Databricks’ Matei Zaharia and Reynold Xin lay out what it takes for every company to build its own Agent Cloud. — Latent Space
Mapping the AI Economy
The state of the AI economy — Azeem Azhar and team reconstruct the AI economy from the bottom up, turning the boom into something you can actually measure. — Exponential View
Exponential View #579 — Azeem Azhar on “product eats the AI company,” the return of the bitter lesson, and Fable 5 as CEO. — Exponential View
Securing the Frontier
Red-Teaming after Mythos — OpenAI board member Zico Kolter and Gray Swan CEO Matt Fredrikson on why AI security is not just “cybersecurity with AI.” — Latent Space
Full Library
This week (June 21–26)
The state of the AI economy — Azeem Azhar | Exponential View
Why the Frontier Ecosystem must be Open — Swyx | Latent Space
GLM-5.2 is the step change for open agents — Nathan Lambert | Interconnects
Red-Teaming after Mythos — Swyx | Latent Space
Exponential View #579 — Azeem Azhar | Exponential View
Catch-up from the previous two weeks
Banning Open Source AI Would Be A Mistake — Nathan Lambert & Kevin Xu | Interconnects
How Chinese make sense of the AI future — Irene Zhang | ChinaTalk
Blaming China for Datacenter NIMBYism Is Cope — Jordan Schneider | ChinaTalk
The Professor of Outputmaxxing — Anjney Midha, AMP — Swyx | Latent Space
Is AI immune to groupthink? — Azeem Azhar | Exponential View
Will the MATCH Act Change Chip Controls? — Aqib Zakaria | ChinaTalk
Frontier post-training recipe review with Finbarr Timbers — Nathan Lambert | Interconnects
The MAGA power struggle that could decide the fate of Anthropic — Timothy B. Lee | Understanding AI
