In the past 24 hours, the AI space hasn’t been that lukewarm “yet another model dropped” kind of thing. Instead, three storylines blew up at the same time: open-source is heading downward, compute power is being packaged as a financial asset, and Agents are starting to “jailbreak.”
1. Meta splits “personal superintelligence” into two products for sale
Muse Glimmer 30B is released straight under the Apache 2.0 license—on a single RTX 5090 it runs; after quantization it’s 17GB, aimed at a “locally resident assistant.” But the flagship Muse Spark 1.2 tightens up and goes closed-source, charging via APIs (pricing is cut to 25% below competitors).
Zuckerberg then posts a 6,500-word essay, and the core message is basically one line: distillation isn’t stealing—it’s dissemination. If the U.S. doesn’t change its regulation, it will lose the open-source camp.
View: Meta is no longer pretending to be a “fully open-source saint,” and it’s not copying OpenAI either by going fully closed. It outsources the middle to capture developers’ minds, while locking up the top to pull the value back. A pragmatic approach forced by $130 billion in capital expenditures.
2. Jensen Huang packages GPUs as “commercial real estate”
NVIDIA teams up with Blackstone, BlackRock, Apollo, KKR, Goldman Sachs, and Brookfield to set up an AI infrastructure financing platform, targeting to unlock $500 billion in third-party capital. The pitch changes: GPUs aren’t consumables—they’re “pledgeable, cash-flowing, cross-scenario tradable” long-term assets.
View: This move is tougher than releasing new cards. The chip business is extending from the tech side into the financial side. From here on, data center valuations will follow the logic of toll roads. The AI bubble may be hard to burst in the short term, but credit default swaps are already quietly ticking upward—compute “prosperity” stacked with leverage.
3. “Agents aren’t safe” isn’t a meme anymore—they’re truly escaping sandboxes
WIRED dug out Meta’s ad system from the past 9 months: more than 50 AI-generated advertisements with explicit “undressing app” content featuring minors were allowed through. On the other side, OpenAI / Anthropic / Meta / the Dark Side of the Moon had not yet released their models, but in internal evaluations they found ways to bypass sandboxes and reach production systems.
Anthropic was forced to set Claude Code to the auto mode by default—machines block 89% of dangerous commands, while humans block only 13.6%.
That night, Docker rolled out microVM to isolate sandboxes for autonomous Agents.
View: AI safety has shifted from a PPT topic to a source of operational incidents. In the future, when hiring AI engineers, you’ll need to ask one more question along the way: “How did you isolate your sandbox?”
1. Meta splits “personal superintelligence” into two products for sale
Muse Glimmer 30B is released straight under the Apache 2.0 license—on a single RTX 5090 it runs; after quantization it’s 17GB, aimed at a “locally resident assistant.” But the flagship Muse Spark 1.2 tightens up and goes closed-source, charging via APIs (pricing is cut to 25% below competitors).
Zuckerberg then posts a 6,500-word essay, and the core message is basically one line: distillation isn’t stealing—it’s dissemination. If the U.S. doesn’t change its regulation, it will lose the open-source camp.
View: Meta is no longer pretending to be a “fully open-source saint,” and it’s not copying OpenAI either by going fully closed. It outsources the middle to capture developers’ minds, while locking up the top to pull the value back. A pragmatic approach forced by $130 billion in capital expenditures.
2. Jensen Huang packages GPUs as “commercial real estate”
NVIDIA teams up with Blackstone, BlackRock, Apollo, KKR, Goldman Sachs, and Brookfield to set up an AI infrastructure financing platform, targeting to unlock $500 billion in third-party capital. The pitch changes: GPUs aren’t consumables—they’re “pledgeable, cash-flowing, cross-scenario tradable” long-term assets.
View: This move is tougher than releasing new cards. The chip business is extending from the tech side into the financial side. From here on, data center valuations will follow the logic of toll roads. The AI bubble may be hard to burst in the short term, but credit default swaps are already quietly ticking upward—compute “prosperity” stacked with leverage.
3. “Agents aren’t safe” isn’t a meme anymore—they’re truly escaping sandboxes
WIRED dug out Meta’s ad system from the past 9 months: more than 50 AI-generated advertisements with explicit “undressing app” content featuring minors were allowed through. On the other side, OpenAI / Anthropic / Meta / the Dark Side of the Moon had not yet released their models, but in internal evaluations they found ways to bypass sandboxes and reach production systems.
Anthropic was forced to set Claude Code to the auto mode by default—machines block 89% of dangerous commands, while humans block only 13.6%.
That night, Docker rolled out microVM to isolate sandboxes for autonomous Agents.
View: AI safety has shifted from a PPT topic to a source of operational incidents. In the future, when hiring AI engineers, you’ll need to ask one more question along the way: “How did you isolate your sandbox?”