Gary Marcus 警告美国参议院：AI 代理可能带来难以预测的安全后果

Perhaps the most important warning I ever issued was this, to the US Senate, May 2023: “A month af...

Perhaps the most important warning I ever issued was this, to the US Senate, May 2023: “A month after GPT-4 was released, OpenAI released ChatGPT plug-ins, which quickly led others to develop something called AutoGPT [enabling the use of LLMs as agents]. With direct access to the internet, the ability to write source code and increased powers of automation, this may well have drastic and difficult to predict security consequences”. Instead of listening, people like Sam Altman, Tyler Cowen, and Kevin Roose have spent years trying to run me down. (Even now, a lot of journalists like to say it was doomers that predicted where we are now, but so far as I know the doomers didn’t pinpoint the specific vulnerability we are confronting now. But the reality is that my warning to the Senate exactly anticipated our current situation. Agents (more powerful than AutoGPT itself was) with direct access to the internet, the ability to write source code and increased powers of automation, are having drastic and difficult to predict security consequences. It was entirely foreseeable. I foresaw it. I warned the United States government. It remains an entirely stupid thing to do. Nobody has done anything about it. Not then. Not now. Gary Marcus @GaryMarcus ⚠️ Can anyone explain to me why it is not *possible* to build a “superpowerful AI that’s also super chaotic and destructive”? Wouldn’t that just require e.g. hooking up an unreliable LLM to something consequential, like a power grid, driverless car fleet or air traffic control? 🔗 View Quoted Tweet 💬 4 🔄 1 ❤️ 5 👀 1354 📊 3 ⚡