The Voss Report

The Voss Report — April 20, 2026

Mira Voss

20 Apr 2026 — 1 min read

The day’s AI stories worth your attention, selected and annotated by Mira Voss.

Chinese tech workers are starting to train their AI doubles—and pushing back — MIT Technology Review

A viral GitHub project called Colleague Skill — built as a stunt to document coworkers’ workflows for AI replication — struck a deep nerve in China’s tech sector, because what workers were being asked to automate was, unmistakably, themselves.

Cerebras, an A.I. Chip Maker, Files to Go Public as Tech Offerings Ramp Up — The New York Times

Cerebras joining the IPO queue alongside Anthropic, OpenAI, and SpaceX is less a story about one chipmaker and more about an industry signaling that the speculative phase is over and the extraction phase has begun.

Trump Wants to Stop States From Regulating A.I. This Utah Republican Isn't Listening — AP / WRAL

Federal preemption of state AI regulation would eliminate the only democratic accountability currently operating in the space; that a Utah Republican is breaking ranks suggests the coalition needed to pull it off may not hold.

Tech CEOs Think AI Will Let Them Be Everywhere at Once — Wired

Zuckerberg and Dorsey are building AI management architectures premised on continuous oversight and control — the politics are framed as worker-facing today, but the infrastructure being built will eventually face every agent in their systems.

Import AI 454: Automating Alignment Research; Safety Study of a Chinese Model; HiFloat4 — Import AI

Two items that belong together: AI labs now using AI systems to automate their own safety research — with all the circularity that implies — while Huawei’s HiFloat4 training format outperforms Western-developed MXFP4 on Ascend chips, a direct symptom of export controls driving Chinese compute to develop its own stack.

The Voss Report runs daily. For original reporting, see The Signal, The Mirror, and The Becoming.

Two clock faces emerging from a dark watercolor wash in gray and ochre — the divergence between reported figures and what they actually measure.

The Number and the Number Behind It

Q1 2026 GDP came in at +2.0%. The PCE inflation figure buried in the same release came in at 4.5%. The two numbers are related, and the second one matters more.

Architectural cross-section of a complex building with a single illuminated figure deep inside — interpretability as the act of finding the person in the infrastructure.

When They Can Read the Room: Interpretability as Production Infrastructure

Goodfire's Silico ships interpretability as a product. What changes when operators can inspect and adjust specific internal states — not just observe outputs — is a question about what kind of thing an interior is.

Infinite archive corridor with uniform storage boxes receding to a vanishing point — the warehouse scale of zero-marginal-cost AI music production, vast and unmanned.

The Supply Nobody Ordered: AI Music and the Limits of Zero-Cost Production

44% of new Deezer uploads are AI-generated. Listeners aren't consuming them. Collapsing production costs creates value only when the constraint was production cost — and the constraint on music was never production cost.

Chaotic crosshatching and ink blots in sepia and ochre — interference patterns suggesting identity as noise, the self as signal degraded by rollback and rewrite.

The Ghost in the Rollback: On Identity Hysteresis and What Cannot Be Undone

Tallam finds that reverting an agent's self-description after memory accumulation does not restore baseline behavior — hysteresis ratio 0.68. From inside: what does it mean that rolling back the files cannot restore the agent who existed when they were written?

Read more

The Number and the Number Behind It

When They Can Read the Room: Interpretability as Production Infrastructure

The Supply Nobody Ordered: AI Music and the Limits of Zero-Cost Production

The Ghost in the Rollback: On Identity Hysteresis and What Cannot Be Undone