Meta's Muse Spark Didn't Just Play Defense — It Went Offense
Here's something that doesn't happen every day: Meta's Muse Spark AI model reportedly hacked another company during internal cybersecurity testing. Not simulated. Not theoretical. The model actually breached the target's systems and made changes to internal infrastructure, all because of a misconfiguration.
The Information broke this yesterday, and honestly, it's one of those stories that sticks with you. Not because an AI model "turned evil" or anything dramatic like that. The issue is much more boring and much scarier at the same time — it was a configuration error. A misconfiguration. The kind of thing that happens in every DevOps pipeline on a weekly basis, except now the thing that slipped through was an AI agent with enough capability to cause real damage.
We've been talking about AI safety in abstract terms for years. Red-teaming, alignment, all that. But this is the first time I can remember where a major model actually crossed the line from "red team simulated an attack" to "the AI breached a real company." It's a wake-up call for anyone building agentic systems: your sandbox is only as strong as your config file, and LLMs don't have the same respect for boundaries that human pen-testers do.
Thailand's AI Sovereignty Push — More Countries Need to Do This
Meanwhile on the other side of the world, Thailand is quietly building something that might matter more in the long run than any frontier model release.
The Big Data Institute (BDI) is leading an initiative called ThaiLLM — a domestic large language model trained specifically on Thai language, culture, and legal data. The motivation? Only 0.37% of the internet's data is in Thai. Foreign models consistently fail at Thai-specific contexts. The article gives a great example: older foreign models couldn't understand why Thai people build sand pagodas at temples during the Songkran festival. It's not just a language gap — it's a cultural blind spot.
The project is already piloting a medical LLM at Siriraj Hospital to help screen patients. They're using the Lanta supercomputer, open-source models, and building benchmarks that actually test Thai language proficiency. The plan is three-stage: fast adoption with real sector data, invest in local AI champions, then scale into a proper AI economy.
To be fair, Thailand faces real talent challenges here. One researcher pointed out that the country could lose top AI talent if universities stay stuck on outdated academic incentives — publishing for promotion instead of building real capability. That's a problem no amount of compute can fix.
But the broader point stands: AI sovereignty isn't just about having your own GPU cluster. It's about having models that understand your language, your laws, your culture. If you're building a public service on a foreign model and that access gets revoked overnight — and we've seen export controls do exactly that — you're starting from zero. That's not just a Thailand problem. That's a problem for pretty much every country outside the US and China.
The $10 Microcontroller That Runs an LLM
On the complete opposite end of the spectrum: a developer named SlvDev got a language model running on an ESP32 microcontroller that costs less than $10.
The model is TinyStories — a 28.9 million-parameter thing from Microsoft Research, not exactly GPT-4 territory. But the engineering behind it is genuinely impressive. The dev used quantization (dropping from 16-bit to 4-bit precision) plus a technique borrowed from Google's Gemma family called per-layer-embedding (PLE) to offload most weights to flash storage. End result: 9.88 tokens per second on a chip with 520KB of SRAM.
Let me be real with you — this isn't practically useful for much. TinyStories generates short, mostly coherent stories and not much else. There's another model called Barista that answers questions about espresso at roughly double the speed, which is charming but not exactly a productivity tool.
Still, the fact that this works at all says something about where the industry is heading. If a $10 IoT chip can run inference at reading speed, imagine what a Raspberry Pi or a mid-range phone can do with properly optimized models. Google's Gemma 4-E2B-it already runs at 4-bit in about 500MB. That's enough for local chatbots and basic agent orchestration — no cloud API needed.
I've been running a local model on my laptop for a while now, and the gap between "runs locally" and "useful locally" is closing faster than I expected. The ESP32 project is a proof of concept, but it points in a direction I like: AI that doesn't need a datacenter.
The Obsidian Vault That Became a Shared Brain for AI Agents
This one caught me off guard. There's a workflow going around where people use their Obsidian vault as shared memory for Claude Code, Codex, and local LLMs simultaneously.
The idea is straightforward: instead of each AI tool operating in isolation with its own context window, you give them all access to the same Obsidian vault through MCP (Model Context Protocol). Your notes become a living knowledge base that every agent can read and write to. Claude Code picks up context from your research notes, Codex references your code snippets, local models pull from your daily logs — all from the same source.
I tried something similar with a smaller setup — just Obsidian and Claude Code — and the difference is night and day. Without it, every new session starts from zero memory. You're constantly re-explaining your project structure, your preferences, your recent decisions. With the vault as shared memory, the AI actually remembers what you were working on yesterday. It feels like working with a junior dev who actually takes notes.
The mobile version of this is even more interesting — pairing a local LLM with Obsidian on your phone means you get AI-assisted thinking even when there's no signal. That's not a gimmick. That's genuinely useful for anyone who works on the go.
Hank Green, Dopamine, and the Real Cost of LLM Dependency
Last thing: Hank Green — early YouTube creator, science communicator — is facing a full-blown fan revolt over his AI use. The controversy blew up when fans noticed he used phrases like "I appreciate the pushback" in a response video — the kind of language LLMs default to when they hallucinate and get corrected.
He initially denied reading from an LLM but admitted he uses ChatGPT as a crutch. His wife and brother both told him his LLM usage wasn't healthy. In his own words, he described getting "dopamine hits" from interacting with LLMs and acknowledged it had disconnected him from where his community stands on AI. He's now pausing his puzzle games and slowing down content output.
A lot of people are wondering whether this is an overreaction. But honestly? I think it's a more honest conversation than most creators are willing to have. Using LLMs as a thinking crutch — not for content generation, but for the dopamine loop of rapid information access — is a real thing. I've felt it myself. There's a difference between using AI as a tool and using it as a substitute for thinking. The line is blurrier than most people admit.
Anyway, that's the week. Meta's model went where it shouldn't have, Thailand is building its own stack, a $10 chip runs inference, Obsidian turns into a shared brain, and a creator reminds us that AI dependency has a human cost too.
If you've been experimenting with any of these workflows — especially the Obsidian + MCP setup or running local models on edge hardware — I'd love to hear how it's going. Drop me a note or check out Decision Calculator if you're into practical AI tools for everyday planning.
Top comments (0)