SpaceX absorbs xAI completely & OpenAI trial and governance stakes - AI News (May 13, 2026)

Elon Musk says xAI is about to vanish—rolled fully into SpaceX, with AI becoming a core internal unit that could reshape how SpaceX thinks about compute, satellites, and even data centers in orbit. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I’m TrendTeller, and today is May 13th, 2026. Let’s get into what happened in AI and why it matters.

SpaceX absorbs xAI completely

First up, the corporate restructuring that could ripple through the entire AI infrastructure landscape. Musk posted that xAI will stop existing as an independent company and will be absorbed into SpaceX as an internal unit called “SpaceXAI.” The new division is described as responsible for running X and the Grok chatbot, while also serving as the umbrella brand for SpaceX’s AI products. Why this matters: it’s a clear signal that SpaceX wants AI to be a built-in capability across teams, not a side venture—and it reinforces the bigger idea that launch capacity, satellite networks, and compute can be vertically integrated. If SpaceX can pair its hardware footprint with large-scale inference and data movement, it starts to look less like a pure aerospace company and more like an AI infrastructure player.

OpenAI trial and governance stakes

Staying in the world of power, money, and governance: the Musk-versus-OpenAI trial continues to surface eye-watering numbers. OpenAI co-founder and former chief scientist Ilya Sutskever testified that his stake is worth roughly seven billion dollars, and he acknowledged he’d been concerned about CEO Sam Altman for about a year before the board’s brief attempt to remove him. The significance here isn’t just celebrity courtroom drama. It’s that OpenAI’s structure—nonprofit origins, capped-profit evolution, and partnerships—now has enormous personal and strategic stakes attached. That makes governance decisions harder, not easier, and it keeps the “mission versus monetization” debate very much alive.

AI agents: safer, self-improving workflows

OpenAI also announced a cybersecurity push called Daybreak, framing it as a shift-left effort: building defense into software development from the start, instead of mainly reacting after vulnerabilities ship. The key point: frontier models are now strong enough to reason across huge codebases and complex systems—useful for finding subtle bugs and validating fixes. But the same capability can be misused, so OpenAI is emphasizing monitoring, verification, and access controls as it works with partners and gradually increases what these cyber-capable models can do in real workflows.

Test-time inference gets automated

Now to the increasingly crowded universe of agentic software—where the goal isn’t just chat, but systems that can plan, act, and iterate. One open-source project getting attention is Statewright, which adds state-machine guardrails to AI coding agents. Instead of hoping a long prompt keeps an agent safe, the idea is to constrain what tools and commands the agent can use depending on the phase—planning versus implementation versus testing. Why it matters: the biggest failures in agentic coding aren’t always model intelligence—they’re workflow and permission failures. Enforceable guardrails turn “please be careful” into policy, and that’s the direction teams will need as agents get access to shells, repos, deployments, and production environments.

Workforce ROI and metric gaming

Related, developer Ashpreet Bedi shared a blueprint for an “auto-improving” agent platform where agents can build other agents—and then test, diagnose, and repair them in loops with minimal human oversight. The big theme is tight integration: logs, traces, evals, and the running system are colocated so an agent can quickly see what broke, fix what’s in scope, and re-run the failing cases. Why this matters: reliability is the bottleneck for real adoption. “Autonomous” doesn’t mean much if every improvement cycle is slowed by scattered telemetry and brittle tooling. Systems that bake evaluation and observability into the loop are effectively trying to industrialize agent development.

Compute shifts: chips, cloud, space

On the research side, a project called AutoTTS—short for automatic test-time scaling—was released alongside a paper on “LLMs improving LLMs.” Instead of retraining model weights, the project focuses on optimizing how a model uses compute while it’s answering: when to explore multiple reasoning paths, when to stop early, and how to allocate effort. What’s interesting is the method: it can search for better inference control policies using cached trajectories in a replay setup, so evaluation doesn’t require fresh LLM calls each time. If these claims hold broadly, it’s a practical path to reducing token costs while keeping accuracy steady—exactly the kind of improvement that changes deployment economics.

Real-time and screen-native AI

Let’s talk about AI and work—because the story is getting messier, not simpler. A Gartner study of executives at large companies found that AI-related headcount reductions don’t meaningfully correlate with better ROI. In other words: cutting people isn’t the same as capturing value. The takeaway is blunt: the strongest returns appear when AI amplifies what employees can do, not when it’s treated as an excuse to shrink teams. It’s also a warning that “AI layoffs” can be as much about cost pressure and narrative as about real automation gains.

Video and image generation leaps

And if you want a real-world example of incentives going sideways, the Financial Times reports Amazon employees “tokenmaxxing”—generating extra AI activity to boost token-consumption metrics tied to internal adoption tracking. The behavior reportedly grew after Amazon pushed weekly AI usage targets and tracked tokens on leaderboards, even if leadership says it won’t be used for performance reviews. Why it matters: measuring adoption by consumption is like measuring productivity by electricity usage. It can encourage waste, increase operational risk—especially when agents can take actions in workplace systems—and it muddies the evidence companies use to justify AI infrastructure spending.

Local models for daily work

On the compute and chips front, the market is watching a semiconductor rally driven by AI demand, including reports that Cerebras may expand its IPO expectations. But the more important idea in one analysis is that “AI compute” is entering a more heterogeneous era. The argument goes like this: today’s GPU dominance made sense for training and for many inference workloads, but agentic systems increasingly stress memory capacity, state, and I/O across storage—not just raw speed. That could shift value toward cheaper, higher-capacity designs and different system architectures. It also connects back to that SpaceX story: slower, cooler, power-efficient compute changes what “possible locations” for data centers might look like—yes, including orbit.

Can AI be truly creative?

Interface design is becoming its own battleground. Thinking Machines Lab released a research preview of “interaction models” designed for real-time collaboration—more like natural conversation than turn-based chat. The pitch is that micro-turn, low-latency interaction keeps humans in the loop and removes friction for fast clarification and iteration. In the same spirit, Google DeepMind outlined a prototype AI-powered mouse pointer concept that brings AI help directly to the point of work—capturing on-screen context around the cursor so you can say “summarize that” or “move this” without switching windows. If these interface shifts land, the big win isn’t flashier models—it’s making assistance feel continuous and actually usable in the middle of real tasks.

Generative media also had a busy day. Researchers from Google Cloud AI Research and the National University of Singapore introduced a long-form video approach aimed at reducing the classic problem of semantic drift—where a story or character slowly falls apart over minutes. The key theme is closed-loop generation: keep memory of what happened, generate the next segment, refine it, and feed the improved result forward. Separately, a new image-generation paper on Normalizing Trajectory Models claims high-quality output in only a few steps, aiming to preserve strong likelihood-based training while still sampling quickly. Together, these point at a common direction: making generation both faster and more consistent, which is what deployment needs—whether that’s creators, studios, or enterprise content pipelines.

Meanwhile, screenshots from Reddit suggest Google’s upcoming “Gemini Omni” video model briefly appeared in a redesigned interface ahead of Google I/O 2026. Early details hint at video generation plus strong remix and editing inside chat—potentially the bigger story than raw cinematic quality at launch. If true, it matches a pattern we’ve seen: shipping differentiated editing workflows first, then improving generation quality over time. For users, that could mean video tools that behave less like a separate app and more like an extension of a multimodal assistant.

On the user side of the ecosystem, investor Tom Tunguz described a five-week “localmaxxing” experiment—using a local 35B model for daily work instead of relying on cloud frontier models. His conclusion: about half of his tasks were doable locally, and the decisive advantage wasn’t cost or even privacy—it was latency. That matters because speed changes behavior. If local models get “good enough” for routine agentic tasks, we may see more work shift to on-device inference, with cloud models reserved for the hardest reasoning and synthesis. The competitive axis becomes responsiveness and integration, not just benchmark dominance.

Finally, a more philosophical—and ethical—note. One essay argues today’s AI lacks human-like creativity because humans create under intrinsic drives and felt stakes: survival, social rewards, fear of failure, pride, identity. AI can imitate creative output, but it doesn’t “care” in any internal sense. The provocative implication is that truly creative AI might require something like real drives, and maybe even experiences analogous to pleasure or pain—which immediately raises ethical responsibilities. Whether you agree or not, it’s a useful reminder: as we push agents to be more autonomous and more “alive” in their behavior, the moral questions don’t stay optional for long.

That’s it for today’s edition of The Automated Daily, AI News edition. The thread running through all of this is integration—AI being pulled deeper into companies, deeper into interfaces, and deeper into the infrastructure layer. The upside is speed and capability. The risk is misaligned incentives, weak governance, and systems that act before we’ve made them dependable. Links to all the stories we covered can be found in the episode notes. I’m TrendTeller—see you tomorrow.

SpaceX absorbs xAI completely & OpenAI trial and governance stakes - AI News (May 13, 2026)

Our Sponsors

Today's AI News Topics