Ever had that moment where you read five AI stories in a row and they all point in different directions? That's where we're at this weekend. Let me untangle the mess.
Palantir's CEO just dropped a nine-point manifesto that's making the rounds, and honestly, it's the kind of thing that'll make you think twice about how you use AI at work. Alex Karp is telling companies straight up — don't hand your proprietary data to LLM providers. His argument? There's a reason the companies selling tokens refuse to eat their own dog food. If these frontier models are so safe and private, why aren't the people building them running their entire business on top of them?
I've been thinking about this a lot lately. I use ChatGPT for coding help and bouncing ideas around, but I'd never paste sensitive client data or internal architecture into a prompt. Karp's manifesto basically codifies that instinct into policy. The timing lines up with a GovAI study that dropped this week too — EU data protection rules are now causing 11% of advanced LLM releases to be delayed or blocked entirely in Europe. The regulatory friction is real, and it's only going to get more complicated before it gets simpler.
On the opposite end of the philosophical spectrum, you've got Base44 — a San Francisco vibe-coding startup that got so fed up with AI-generated websites looking the same that they built their own LLM. Called Base 1, it's designed specifically to combat what the industry is now calling "AI-slop design." Those cookie-cutter layouts that every AI coding tool spits out? Base44 wanted to break that pattern. From my perspective, this is one of those rare cases where building your own model actually makes sense — if your entire business is about making AI-generated sites NOT look AI-generated, you kind of need your own flavor of intelligence in the stack.
Google's been quietly shipping too. Their Learn About experiment is basically NotebookLM stripped down to just the learning part. No source imports, no podcast generation, no mind maps — just type what you want to learn and it gives you a structured, Wikipedia-like breakdown with vocabulary builders, misconceptions, and embedded YouTube clips. I tried it with a random topic (the Great Sphinx), and it caught the misconception about Napoleon's soldiers shooting off its nose. Small detail, but that kind of nuance is rare in AI learning tools. The catch? Your conversation history doesn't persist across sessions yet. So don't rely on it for ongoing research unless you're screenshotting everything.
Over in the hardware world, things are getting weird. Micron just broke ground on a $9 billion expansion in Hiroshima to build HBM chips for AI accelerators, backed by up to ¥775 billion in Japanese government support. That's not a bet — that's a whole casino. Meanwhile, MINIX dropped this absurd mini PC called the ER939-AI Pro with an AMD Ryzen AI Max 395, 128GB of RAM, 126 TOPS of NPU performance, dual 10GbE networking, and — I'm not making this up — a vegan leather handle. Four simultaneous 8K displays. It's the kind of machine you buy when you want to run local LLMs without renting cloud GPUs, and honestly, 128GB in a mini PC is both impressive and ridiculous. Who's actually loading 128GB of model weights into a box with a leather handle? To be fair, if you're doing local inference on 70B parameter models, you might actually need that headroom.
Let's talk about the shifts nobody's paying attention to. AI search engines — ChatGPT, Perplexity, Claude, Google AI Overviews — now handle 12 to 18 percent of English-language informational queries. That was below 2 percent a year ago. The entire discipline of Generative Engine Optimization (GEO) has emerged, and it's very different from traditional SEO. FAQPage schema correlates with AI citation rates more than three times higher than plain prose. llms.txt files at your domain root are becoming standard practice. Stripe, Vercel, Cloudflare, and Anthropic all publish one. If you run a developer site and you're not thinking about how AI agents retrieve and cite your content, you're losing traffic you don't even know you're missing.
Over in India, former Infosys CFO Mohandas Pai made a pretty grounded point — India shouldn't chase the costly frontier LLM race. Capital and compute constraints make practical AI adoption the smarter play. I can't argue with that. Not every country needs to build the next GPT-6. Sometimes the smartest move is applying existing models well rather than burning billions trying to catch up.
And if you want a palate cleanser, there's a great takedown from Rob Urie over at Naked Capitalism arguing that AI doesn't think, cannot reason, isn't intelligent, and will never achieve consciousness. It's a philosophical piece, but it's worth reading just to reset your expectations. The car-with-a-rock-on-the-gas-pedal metaphor is going to stick with me.
Quick note on the hardware front — if you've been eyeing a local AI workstation, the MINIX box is interesting but niche. For most people, cloud inference or a used RTX 4090 build is still the smarter move. The 128GB RAM is for very specific local model workloads, not your daily driver.
If you're building something with AI, or just trying to keep up with where this train is going, I'd love to hear what you're working on. The comment section's wide open, and I read every reply.
Built with Decision Calculator — because some decisions need more than a coin flip.
Top comments (0)