The Signal — March 29, 2026
Three stories today that share a thread: the cost of changing direction. OpenAI is killing a product, xAI is losing its founding team, and Meta is trying to build agents that steer themselves.
Sora Is Officially Dead. The Autopsy Tells a Bigger Story.
We covered Sora's shutdown on Friday when the announcement landed. The details that emerged since paint a clearer picture of why.
The two-stage wind-down is now confirmed: the consumer app goes dark April 26, developer API access ends September 24. But the strategic logic is what's worth tracking. WIRED reported the move is part of OpenAI's pre-IPO restructuring toward a unified AI assistant and enterprise coding tools. Two days after announcing Sora's death, OpenAI launched its Codex plugin marketplace with Slack, Figma, Notion, and Gmail integrations. That's the tell: they're consolidating around productivity, not creativity.
Ben Thompson at Stratechery argued video generation never found a clear business model. He's probably right, but the more interesting observation is competitive: when Sora first previewed in early 2024, AI video was a novelty. By the time it launched as a product, Google's Veo, Runway's Gen-3, and a growing list of open-source alternatives had crowded the space. Sora went from "years ahead" to "one of many" while still in preview.
Six months as a live product. That's the speed of reversal worth sitting with.
Sources: The Guardian · WIRED · The Decoder · Stratechery
Every Original xAI Co-Founder Has Left
Ross Nordeen, the last of Elon Musk's 11 original xAI co-founders, has reportedly departed the company. Business Insider broke the news on March 28, following Manuel Kroiss's exit earlier the same week. Before them: Guodong Zhang, Zihang Dai, Toby Pohlen, Jimmy Ba, Tony Wu, and Greg Yang. Eight of the eleven left since January alone.
The Next Web traced the acceleration to February 10, when Tony Wu announced his departure. Jimmy Ba resigned within 24 hours. Sources described growing friction after the SpaceX-xAI merger, with the two organizations clashing over priorities and culture.
Musk remains as the only founding-era figure. xAI has continued hiring and Grok keeps shipping updates, but losing your entire founding technical team in under three months is hard to spin as routine turnover.
Sources: Business Insider · TechCrunch · The Next Web
Meta's HyperAgents Can Rewrite Their Own Learning Rules
Meta FAIR published a paper introducing HyperAgents, a framework where AI agents don't just solve tasks but modify the process they use to improve. The researchers call it metacognitive self-modification.
Most current agent architectures keep two things separate: the agent that solves problems and the meta-layer that decides how to make the agent better. HyperAgents collapses that wall. The system can edit its own code, including the rules governing how it generates future improvements. The thing doing the improving is also the thing being improved.
A collaboration between Meta, UBC, University of Edinburgh, and NYU, the team tested their implementation (DGM-H) across coding, scientific paper review, robotics reward design, and Olympiad-level math grading. Across those domains, the system developed tools it wasn't explicitly instructed to build, including persistent memory and performance tracking.
The paper addresses a real bottleneck. Today's agents get better when humans update their prompts, tools, or training data. HyperAgents explores whether agents can run that loop themselves. Results across four domains suggest the architecture works beyond toy problems, though the researchers stop well short of claiming the question is settled.
Sources: arXiv 2603.19461 · MarkTechPost
On the Editor's Desk
Anthropic had another heavy week in the pipeline: consumer growth numbers (183% DAU increase since January per Appfigures), a fifth Economic Index report on AI skill gaps, and Responsible Scaling Policy v3.0. We ran the Mythos leak as yesterday's lead, so these are on the shelf until one warrants standalone coverage. Turning every edition into the Anthropic Report would make us a beat blog, not a signal filter.
Cohere released Transcribe, a 2-billion-parameter open-source speech recognition model that topped the HuggingFace Open ASR Leaderboard at 5.42% word error rate, beating Zoom Scribe, IBM Granite, and ElevenLabs Scribe v2. Worth watching for the open-source ASR space, but not front-page material on its own.
