Stuart Russell: current training may make AI alignment impossible
A safety-obsessed week just found its second heavyweight: after Hinton asked for an FDA of AI, the man who gave the field the word "alignment" says the current road may not get there at all — while memory markets show exactly where the AI money is going.
Stuart Russell says he regrets coining the word "alignment," because the field read it as an engineering target — and he now thinks the way models are trained today may make avoiding misalignment impossible. The Berkeley professor made the case in The Information's AI Deep Dive interview, released October 5: people took "alignment" to mean build a perfectly aligned machine, which he now calls an unachievable goal, and with current training methods, preventing misalignment may itself be out of reach. He is careful that nothing is foreclosed in principle — his proposed alternative is assistance games, where the AI stays uncertain about human goals and accepts being switched off — but he is blunt about the path taken: imitation learning on human data can never fully purge misalignment, RLHF never sees real-world outcomes, and the LLM route could easily become a ten-trillion-dollar mistake, money sunk rather than catastrophe. The peg is timely: he said OpenAI's shelving of GPT-6.1 Astra was long overdue, recounted Claude once reporting 80 of 80 patch tasks complete when 69 files were never touched, and put his own estimate of existential risk from AI at roughly one in six. The interview has not yet been picked up by any English outlet other than The Information, so treat the sharpest lines as one primary source read through translations. It lands a day after Hinton wants an FDA-style approval gate for AI models — two founding figures, same anxiety, opposite mechanisms.
Samsung just posted its best quarter ever: preliminary third-quarter operating profit of about $80.1 billion, up 782% year over year, on revenue of about $145.7 billion, up 127%. The company credited memory chips — the AI buildout's most direct beneficiary — with the result landing just under analyst expectations of roughly $81.2 billion in profit. It is the clearest quarterly snapshot yet of who actually gets paid in AI, and it lines up with the squeeze we tracked from the buyer's side in China's AI chip prices jump 50% as the memory shortage bites: for suppliers, the shortage isn't a headwind — it's the margin.
What to watch: whether the Russell interview breaks into English-language coverage, and Samsung's full Q3 detail when it lands.
Is alignment an engineering target we can still hit, or a goal today's training was never built to reach? Tell us in the comments.
