Fine-tuning your next big idea with AI? The moment you press enter, you might be giving it away.
Ep. 8 of AI Uncut: "free" AI, data ownership & whether AI labs are becoming the new incubators.
With INXM, operational data can stay in-house. 🎧👇
Is "Fast Thinking" really faster thinking, or just paying to skip the line?
That loading animation doesn't mean the model works harder. Often your prompt is just stuck in a queue.
What does €200/month actually buy? Nobody knows. AI Uncut Ep. 7 👇
Most companies still pick an AI model the way they pick a smartphone: one leaderboard number, and the rest is marketing.
Check the rooms that match your job: knowledge work (real files + tools), coding (merge-ready), science (verifiable), general (no inventing).
1/ "Superintelligence will wipe out humanity." Silicon Valley is calling for regulation. I'm deeply sceptical. Is this real, or the next Y2K? 🧵
A little disclaimer: This visual was generated by AI. The opinions are very much my own.
4/ I'm not anti-LLM. They're fantastic at certain tasks. But they're wildly inefficient at automation, and they were oversold.
1/ Satya Nadella just wrote about insider risk in the superintelligence era. If you're deploying AI, read it. His point: the safest system isn't the one where we trust the model most. It's the one that lets us trust it the least. 🧵
2/ With traditional software, we could trace every behavior back to a code path. With frontier models, that's gone. Yet we're giving agents our most sensitive data and letting them take mission-critical actions.
3/ Responsibility can't be outsourced. Too many companies and governments lean on a model provider's assurances. Those don't remove our accountability for what AI does in our name.
4/ The good news: security already has the playbook. Identity, least privilege, logging, containment. And one old rule: a program must never be able to bypass the mechanisms that enforce its permissions.
5/ That's how we build at INXM: deterministic environments where agents act, with every permission enforced outside the model. Guardrails that hold even when the model misbehaves. Full piece 👇
