#47 - Who's to Blame? Palantir Says "Likely Human Error", and AI Learns to Copy Itself

Missiles hit a school in Iran and at least 123 children die. Months later, the company whose AI helps the US military suggest targets goes on TV and says: the likely cause is human error.

Today is one question: who is to blame when AI goes wrong – the machine, the people who trusted it, or the people who sold it? Recorded in Salzburg.

In this episode:

🌑 Palantir says "likely human error" (01:40) — Palantir sells data and AI tools to armies and governments. Its Maven Smart System, used by the Pentagon, sorts drone and satellite data and suggests targets. On 28 February, the first day of the US–Iran war, two Tomahawk cruise missiles hit an elementary school in Minab, Iran, and at least 123 children were killed. According to Bloomberg, a Pentagon review that is still not public found old intelligence, old satellite pictures and cuts to the civilian-safety team – reportedly from ten people to one. Palantir's CTO Shyam Sankar told CBS "Face the Nation" that the reporting is inaccurate and the likely cause is human error. Malcolm's take: old data in, wrong answer out – and one person checking every suggestion is no longer a human in the loop.

🌑 Berlin shelves the Mistral–SAP deal (04:51) — in November 2025, Germany and France announced in Berlin that Mistral AI and SAP would build AI tools and digital helpers for government offices. The framework contract was never signed: Politico reports the Federal Digital Ministry currently has no plans to buy from Mistral. The German search engine Ecosia dropped Mistral too – CEO Christian Kroll says its models are about a year behind. Mistral's answer is a preview of Mistral Large 4, which it says can compete with some Chinese open models. Malcolm's take: Europe won't get sovereign AI without building the data centres.

🌑 AI that could copy itself onto your TV box (07:59) — Anthropic's Frontier Red Team tested whether a small AI model can copy itself onto cheap connected devices like TV boxes, cameras and routers. Alibaba's Qwen 3.5, with nine billion parameters, set up a working AI server on 8 of 9 hardware types, and because each new copy reads the notes of the one before, the success rate rose from 45 percent to as much as 84 percent. GLM-5.2 from Z.ai reached all 8 target machines in 4 of 5 trials – in a walled-off test network. The researchers' own projection: small models could do this by the end of 2026.

🧵 The thread (11:56) — this week the English show tries a shorter format: the top stories in 10 to 15 minutes. Long or short – tell Malcolm which one you prefer.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.