New article's up. Hugging Face says roughly half the Fortune 500 now runs open source AI models instead of renting through an API. Sounds like a trend to chase, but it's a scale question, running your own infrastructure only pays off once your usage is massive and constant enough to beat per-token pricing. Most founder-sized businesses aren't close to that line, and renting stays the right call. The number that's actually relevant: McKinsey found 32% of orgs skipped buying a software product entirely because they built it internally with AI instead. That one applies at almost any scale. Full read: neonaliens.com/intel/rent-vs-own-ai-founders.html Anyone here actually built an internal tool with AI instead of paying for a SaaS subscription? What'd you replace?

Anthropic disclosed that five infostealer strains, Vidar, LummaC2, StealC, RedLine, and Acreed, are stealing active Claude sessions directly off infected machines. Not phishing, not a fake login page, the malware just needs to already be on your computer from something unrelated. If you or your team use Claude, ChatGPT, or similar tools on personal or lightly managed machines, this is worth 10 minutes today: run an anti-malware scan, check your active sessions in Claude and sign out of anything unfamiliar, and turn on two-factor if you haven't. Full breakdown: neonaliens.com/intel/claude-infostealer-malware-founders.html Has anyone here actually run a malware scan on their main work machine recently, or is it one of those things everyone assumes is fine until it isn't?

New roundup is up. The big one: 16 state Attorneys General sent a joint letter to OpenAI demanding answers on the Hugging Face breach, and a court separately ruled on whether Anthropic poses a national security risk. The security incidents we've tracked all month just moved from "industry handles it quietly" to actual legal and regulatory accountability. Also this week: Nvidia's reportedly in talks to invest in Perplexity at a $30B+ valuation, and a new Pew survey found 34% of US adults now use AI chatbots for health questions, but only 29% feel comfortable sharing health data with them. That usage-vs-comfort gap is worth sitting with if your product touches sensitive customer data at all. Full roundup: neonaliensai.substack.com/p/this-week-in-ai If you're depending on a frontier AI provider for anything core to your product, has this kind of regulatory pressure been on your radar, or does it still feel far off?

New article breaks down when building your own AI model makes sense, and for almost everyone reading this, the honest answer is it doesn't. Thomson Reuters could justify it because they sit on decades of proprietary legal and financial content no general model was ever trained on. That's the real bar: a genuinely unique dataset at real scale, where the AI capability built on it is core to your product. The lesson that actually applies at founder scale isn't "build your own model," it's figuring out what your version of that proprietary asset is, whatever process, data, or knowledge makes your business different, and making sure your AI usage is actually built around leveraging that instead of using AI the same generic way any competitor could. Full read: neonaliens.com/intel/thomson-reuters-ai-model-founders.html What's the thing in your business that a generic AI model could never replicate?