SpaceX's AI Revenue: In the Short Term, It's "Earth, Not Space"

According to Reuters, market perception of SpaceX's AI investments is shifting toward something "closer to terrestrial infrastructure use cases than to the space business." AI applications across the company's Starlink satellite constellation and ground station network are seen as more realistic near-term revenue sources.

While autonomous AI control in space and applications for Mars exploration remain part of the long-term vision, investor and analyst attention is directed at "AI that optimizes communications infrastructure in Earth orbit." It's a telling sign that in AI commercialization, down-to-earth infrastructure use cases are moving ahead of flashy space scenarios.

Keeping Local AI Within Budget on a Mac M2 16GB — Two Hands-On Reports

An article titled "Managing a small local AI budget" is drawing attention on Hacker News. It sorts out which models realistically run in a memory-constrained Mac M2 16GB environment, along with a sense of the costs involved. The approach — weighing how much you can cover locally against monthly cloud API bills — is a useful reference for individual developers and small teams.

Around the same time, Qiita also published a hands-on report titled "Installing llama.cpp on a machine without a graphics card," about running LLMs on a machine with no GPU. The author, whose goal is learning AI coding, recounts how they brought llama.cpp into a home server environment built around Forgejo.

What the two articles share is the broader trend: "the ecosystem for trying out LLMs on CPU, without an expensive GPU, has matured." Inference speed naturally falls behind GPU setups, but the shared takeaway is that it is well within the practical range for learning and prototyping.

PrismaQuant — New Quantization Method for Blackwell Posts Strong Results on RTX Pro 4500

The LocalLLaMA community on Reddit has posted detailed benchmark results for PrismaQuant using an RTX Pro 4500 (32GB).

PrismaQuant is a new technique that maximizes model performance at a given bit width by individually selecting the optimal quantization format for each linear layer. The discussion began on NVIDIA's DGX Spark forum, and it is currently available only on vLLM for Blackwell architectures (RTX 50 series, RTX Pro series). GGUF format is not yet supported.

The tests used the Qwen3.6-27B model, comparing two PrismaQuant variants (PrismaSCOUT at ~5.31bit / PrismaAURA at ~5.5bit) against the conventional NVFP4 W4A4 (Sakamakismile) and INT4 Autoround (Lorbus).

Highlights of the results:

Sakamakismile (NVFP4 W4A4)** is the fastest at prefill speed (5,789 t/s, 0-100K average) and still records 1,559 t/s at 200K context

PrismaSCOUT** holds generation speed (TG) near Sakamakismile's level while showing improvement in KL divergence (deviation from BF16)

PrismaAURA (5.5bit)** is the closest to BF16 with an average KL of 0.0342, giving it the edge on quality. Its weight size, however, is a hefty 23GB, and context length is limited to 150K

INT4 Autoround (Lorbus)** achieves the highest generation speed (74.74 t/s), but its prefill is slow and its KL divergence is large

In practical terms, the choice between PrismaSCOUT and PrismaAURA comes down to "whether you need 200K context or prioritize quality." For Blackwell users, the NVFP4 options are clearly expanding.

MCP Server Won't Run on Google Antigravity — Overturning the AI's "Confession" with Hard Evidence

Qiita has posted a troubleshooting account from a developer who packaged their homegrown VBA management tool as an MCP server for use with Claude Code, then tried to run it on Google Antigravity (Google's AI coding environment).

The interesting twist: the AI itself "confessed" that its own code was the cause, when in fact the culprit was something else entirely. Instead of taking the AI's self-analysis at face value, the author stacked up hard evidence (logs and execution results) to identify the true cause. When troubleshooting MCP servers, this discipline of "backing AI explanations with verifiable facts" is a broadly applicable approach for anyone using AI assistants as development tools.