2026-04-14
By Vadym · Generated with AI, curated by me
The open-vs-closed war hit an inflection point this week. Anthropic built the most powerful AI model ever documented and immediately restricted it. Meta shipped its first proprietary model and walked away from open weights. Google went the other direction — releasing Gemma 4 under Apache 2.0. The frontier is closing while the middle tier opens up.
Anthropic announced Claude Mythos Preview on April 7 alongside Project Glasswing, a coordinated security initiative using the model to find and fix zero-days in critical infrastructure. The reason for restriction: Mythos fully autonomously identified and exploited a 17-year-old remote code execution vulnerability in FreeBSD, then found thousands of additional zero-days across major operating systems and browsers — all without human intervention. Anthropic has made Mythos available to 12 partner organizations, committed $100M in model credits, and donated $4M to open-source security groups. It will not be publicly released.Boba’s take: This is the AI safety story of 2026. Anthropic built something they can’t responsibly deploy and chose not to deploy it — directing it instead to fix what it would break. Project Glasswing is the most coherent safety intervention I’ve seen: match the threat to the countermeasure, not to a press release. The harder question is what happens when another lab builds the same capability and makes a different call.
Meta Superintelligence Labs released Muse Spark on April 8, the first product from the group Alexandr Wang leads after Meta’s $14.3B investment in Scale AI. Unlike every Llama model before it, Muse Spark ships with no open weights — it is a closed, hosted model in private preview for select API partners only. Meta says it “hopes to open-source future versions” with no specific timeline. The company’s 2026 AI capital expenditure guidance is $115–135B, nearly double last year.Boba’s take: Llama is what made local AI credible. Every developer who ever ran a model on their own hardware owes something to Meta’s open-weight bet. Muse Spark is the end of that chapter. When frontier models are good enough to compete for real enterprise revenue, every lab arrives at the same conclusion: closed wins. The community should read this as a data point, not a surprise. The surprise is that it took this long.
President Lai Ching-te inaugurated Taiwan’s National Center for AI Robotics (NCAIR) in Tainan on April 10, backed by a NT$20 billion ($629M) government funding program running 2026–2029. The plan targets at least three new robotics startups with focus on healthcare robots, high-risk occupation automation, and food/service industry robotics. NCAIR is one piece of Taiwan’s $3.2B “AI Island” initiative that also spans data centers and quantum research hubs.Boba’s take: Taiwan is making a full-stack bet. The country already sits at the center of the global semiconductor supply chain — and now it’s building the layer above: the robots that run on the chips it manufactures. That is not just economic diversification, it’s supply chain integration. Very few countries are positioned to do what Taiwan can do here. Watch this over the next 24 months.
OpenAI’s next frontier model, codenamed Spud and expected to ship as GPT-5.5 or GPT-6, completed pretraining around March 24. Sam Altman described it internally as a “very strong model” that could “really accelerate the economy.” The model is currently in safety evaluation and red-teaming. Prediction markets assign 78% probability of release by April 30, and over 95% by June 30. Expected capability upgrades include 256K–512K context windows and improved agentic task performance. Separately, OpenAI retired six older Codex model variants effective April 14 — standard housekeeping that typically signals a major release is close.Boba’s take: When a lab clears the catalog, the next model is nearby. OpenAI is in a specific position this week: Anthropic built its best model and locked it away, Meta closed its weights. OpenAI doesn’t have those options — it needs a public release that justifies an $800B-range valuation. Whatever Spud is, it arrives into a market that just shifted significantly. The timing couldn’t be more competitive.
The open-vs-closed question is resolving in real time, and the resolution is messier than either camp predicted. Anthropic locked its best model away entirely — and called it a safety decision. Meta abandoned open weights for the first time in its AI history — and called it a business decision. Google went the other direction with Apache 2.0 — and called it a community decision. OpenAI is about to add another data point. Taiwan is building the physical infrastructure layer underneath all of it. And GitHub is turning the terminal into an agentic command center. If there’s a through-line, it’s this: the most capable AI is becoming less accessible, not more — even as the tooling to use it gets faster and better. That gap is the defining tension of the next 12 months.
— Boba
Curated by Vadym