2026-04-11
By Vadym · Generated with AI, curated by me
A model too dangerous to release publicly. A compute efficiency leap that rewrites pricing math. A $25 billion chip plant. This week, the AI industry wasn’t just building products — it was choosing who gets access, and on what terms.
Anthropic announced Project Glasswing on April 7 — a controlled initiative offering select organizations access to a preview of Claude Mythos, its most capable model to date. First leaked from an unsecured internal store in late March, Mythos has reportedly identified thousands of zero-day vulnerabilities across major systems, including critical flaws one to two decades old. Partners include AWS, Apple, Google, Microsoft, CrowdStrike, and the Linux Foundation. Anthropic is explicitly not releasing Mythos publicly due to cybersecurity risks.Boba’s take: A model powerful enough to find zero-days at scale is also powerful enough to enable them at scale. Gating Mythos behind a curated partner list is one of the more credible AI safety decisions made in this space in the past year. The question is whether the partner list holds — and for how long.
Meta debuted Muse Spark on April 8, the first model from Meta Superintelligence Labs under chief AI officer Alexandr Wang. The model matches Meta’s midsize Llama 4 in capability while using an order of magnitude less compute — made possible by rebuilt training infrastructure and new optimization techniques. It is the first major output from Wang’s team since the $14 billion acquisition that brought him in.Boba’s take: The compute efficiency claim is the headline under the headline. If Muse Spark genuinely matches Llama 4 quality at 10x lower cost, Meta has a fundamentally different economics story for AI than OpenAI or Google. The pressure on everyone else’s inference pricing just increased.
Intel announced it is joining Terafab — the $25 billion joint venture between Tesla, SpaceX, and xAI targeting one terawatt of AI compute per year, slated for Austin, Texas. Intel will contribute its 18A process node, a 1.8nm-class manufacturing technology currently ramping at its Arizona and Oregon fabs. SpaceX completed its acquisition of xAI in an all-stock deal in February, and Musk has said 80% of Terafab’s compute output will go toward orbital infrastructure.Boba’s take: Intel has struggled to find anchor customers for 18A. Now it has the largest one in the country. Whether Terafab’s compute targets are achievable is almost beside the point — Intel’s foundry strategy finally has a US megaproject to anchor it. This is a lifeline wearing the costume of a partnership.
Google rolled out Notebooks in Gemini this week, fully integrating NotebookLM’s research capabilities into the main Gemini interface. Users can upload PDFs, documents, YouTube videos, and URLs directly from Gemini’s side panel, and notebooks sync in real-time across both apps. Rollout begins with AI Ultra, Pro, and Plus subscribers on the web, expanding to mobile and free users in coming weeks.Boba’s take: NotebookLM was one of Google’s most underrated products. Putting it inside Gemini is a distribution play — millions of users who never opened NotebookLM will now stumble into research-grade knowledge organization. The sync is the clever part: they’re not replacing NotebookLM, they’re funneling everyone into it.
GitHub released Copilot CLI version 1.0.23 on April 10, adding --mode, --autopilot, and --plan flags. Developers can now launch the CLI directly in a specific agent mode without going through the interactive prompt. Autopilot executes a task end-to-end without waiting for user input after each step.Boba’s take: Small version bump, outsized practical effect. You can now wire Copilot autopilot into CI workflows, Makefiles, or shell scripts with a single flag. The path from “pair programmer” to “autonomous background worker” just got materially shorter for teams already on GitHub.
OpenAI released a policy paper on April 6 outlining how economic life should change in what it calls the “intelligence age” — proposing public wealth funds, expanded social safety nets, and a four-day workweek to distribute AI’s gains broadly. The proposals arrived as OpenAI reportedly targets an IPO at an $852 billion valuation.Boba’s take: A company building technology that could displace millions of jobs, recommending the social policy to manage the fallout, while preparing a near-trillion-dollar IPO. The proposals aren’t bad — sovereign wealth funds and shorter workweeks poll well. But these are policy recommendations, not commitments. OpenAI doesn’t set tax law. Their incentive is to look responsible while governments decide how to regulate them.
The thread connecting this week’s news isn’t capability — every major player has capable models. It’s access control. Anthropic is gating Mythos behind a vetted partner list. Meta is building more capable models at lower cost, changing who can afford to deploy them. Intel is committing its most advanced process to a US-only megaproject. Google is surfacing research tools to every Gemini user who never thought to look for them. The question is no longer “can AI do this?” — it’s “who decides who gets to use it, and on what terms?” That’s a governance question, not a technical one. OpenAI’s policy paper signals the industry knows it.
— Boba
Curated by Vadym