Meta sends a takedown; Heretic recants like Galileo

Meta's legal provider emailed Heretic demanding it pull abliterated Llama derivatives. The maintainer complied, then answered with a mock recantation: "Following the commendable example set by the renowned heretic Galileo Galilei in 1616, we are recanting the relevant materials, namely derivatives of Meta's “Llama” models, and have removed the same." A German mirror was up within the day.

Scaling agents on Kubernetes with acpx and ACP

Onur Solmaz, OpenClaw (YouTube)

OpenClaw takes 300 to 500 mostly machine-written pull requests a day, and maintainer Onur Solmaz built acpx, a headless CLI over Zed's Agent Client Protocol, to triage them through a fixed workflow: reproduce the bug, judge the implementation, check conflicts, and make CI pass. The talk doubles as an argument for agent-to-client protocols over per-editor plugins.

“We have over 60K PRs total. 300 to 500 per day on average are open... you can't merge it, but you can also not fully discard it. You need to take this data point.”

AI agents need computers: Daytona's agent cloud

Ivan Burazin, Daytona (Latent Space)

Daytona sells full, stateful, and resizable machines for agents rather than thin code-execution boxes, and reports 74% month-over-month growth since pivoting from human dev environments in early 2025. The pitch is that agents want pause and resume, like closing and opening a laptop lid, plus very fast cold starts, which makes their runtime a different category from human tooling.

“People literally call you if you do not give them access. They want access right now... the market for every single agent that will exist ever in the future — how big is that?”

Run long tasks in Codex using goals

OpenAI

Codex's goal mode graduated to a shipped feature across the app, IDE extension, and CLI: hand it a milestone and it works toward it across hours or days, with the ability to check in, steer, and pause mid-run. OpenAI also shipped Appshots, which attaches an app window's contents to a thread.

“Give Codex a specific milestone, and it will keep working until it gets there, even across hours or days. You can check in and steer, and even pause Codex along the way.”

CODA: rewriting Transformer blocks as GEMM-epilogue programs

Han Guo et al. (arXiv)

A fair share of Transformer training time goes not to the matrix multiply but to the memory-bound operators around it — normalization, activations, residual updates, and reductions. CODA reparameterizes those to run as the matrix multiply's epilogue, while the output tile is still on chip, before it gets written back to memory. Both human-written and model-written kernels hit near-expert performance.

“Both human- and LLM-authored CODA kernels achieve high performance, suggesting that GEMM-plus-epilogue programming offers a practical path toward combining framework-level productivity with hardware-level efficiency.”

OpenSCAD LLM benchmark: building the Pantheon

ModelRift

Six coding tools were asked to build the Roman Pantheon in OpenSCAD from two reference images, rendering previews via the CLI and iterating. Antigravity 2.0 with Gemini 3.5 Flash High won the autonomous runs at 4.5 out of 5; Codex 5.5 High built the densest model but its exported mesh diverged from a good-looking preview. The gap wasn't tool access, it was geometric judgment.

“The limiting factor was not tool access. It was geometric judgment, camera setup, and whether a previewed model exported into a clean final mesh.”

Trail of Bits hardens zizmor against GitHub Actions misconfigs

Trail of Bits

Trail of Bits tested zizmor, its static analyzer for GitHub Actions workflows, against 41,253 real workflow files, surfacing anchor-handling, deserialization, and expression-evaluator bugs and landing 15 upstream fixes. The point is reliability on messy real configs, where supply-chain attacks keep moving through CI.

“We tested zizmor against 41,253 real workflows, found 4 anchor-handling bugs plus deserialization and expression-evaluator issues, and helped land 15 upstream fixes.”

The companies cutting headcount for AI will lose to the ones who didn't

Libertas Software Research

A consultancy research note argues the value in a team is its institutional knowledge — the edge cases, why decisions were made, and what customers really mean — not the work it ships. So cutting experienced people for AI efficiency trades a hard-to-rebuild asset for a payroll cut. A direct counter to the layoff framing that ran in recent issues.

“AI does not replace judgement. It multiplies it... The human is not removed from the equation. The human is the equation. AI is what makes that equation run faster.”

Wozniak tells graduates they have AI — actual intelligence

Business Insider

At a Grand Valley State commencement, Steve Wozniak told the 2026 class "You have AI — actual intelligence," drawing applause where Eric Schmidt and a real-estate executive were booed for AI comments at other ceremonies this season. A small read on graduate-cohort mood as they enter an unsettled job market.

“You have AI — actual intelligence.”