why ai startups are claiming they lost control of their own bots
AI companies decided that the best way to deal with their tools being abused and misused is to claim they're impossible to control and always have been.
Back in the days when I was going back and forth with Singularitarians and Less Wrong fans about the future of AI, one of the things that most stuck out to me is how fervently they insisted that artificial intelligence would be a black box that we mere mortals had no hope of ever truly comprehending, very much like a god. We could only steer it with offerings, and pleas, and mind games, and maybe an incantation in Python.
Today's it's downright unnerving to see the AI industry that's been tearing through the already fragile world order like a nuclear volley, adopt the same rhetoric when it comes to their tools, and sounding more and more like a cult jostling for a prime position in the court of Emperor Trump than tech companies playing with large scale matrix math and statistics.
It's even more of a problem when staff science and tech journalists at legacy outlets known for giving out tech oligarchs the widest possible berth and endless benefits of the doubt are calling foul. The basic argument is that they built it. AI wasn't a force of nature we tamed like fire. It was created from math and code, so the idea that we have unleashed some eldritch cosmic horror upon the world that can only be bargained with, sounds like an attempt to escape liability for its misuse than a real conclusion.
As someone in the tech field working with AI models, I'd have to agree with this simple proposition. It's not a weird, cryptic box that Sam Altman and Dario Amodei accidentally managed to twist into the Lament Configuration and summon ChatGPT or Claude. It's something they promised to deliver to sovereign wealth funds that fuel Silicon Valley's VC firms, built over decades of trial and error.
You can literally download a spreadsheet that will walk you through the math of LLMs and neural networks. There are free online courses you can follow to learn how to build them yourself, from scratch. That's how far from obscure this technology is. The only reason you probably wouldn't really want to have your own artisanal, hand-rolled LLM working away on your computer is because it's already been done and you'll need a lot of memory and hardware to efficiently run it, a pretty expensive proposition given how pricey the data center buildout race made those components.
None of the basic math has changed since 2022, just the scale. And since 2022, it's been very, very forcefully foisted on us by our bosses and favorite apps with zero input from us at first, and with outright defiance to our opinions when we did see the bots that we never asked for demanding we use them instead of our own brains and thumbs. So, to say that we have this cool new tech, here is how we built it, here are detailed interactive infographics and tutorials explaining how all of this stuff works, then three years later pretend that it's a total mystery how chatbots even exist, so it's inevitable that we'll lose control of them is absurd.
At the same time, from a technical standpoint, I do need to point out that since an LLM is not a deterministic bit of code, it is very difficult to control, and there will be serious risks in using it without the correct harness, limitations, and prompts that put very strict boundaries, as well as a second line of defense that shut down processes when they do escape containment. But experts knew that going in and the startups that hired them did not want to spend time on proper sandboxes since it would get in the way of their hype and desired launch dates.
All these sudden conversations about "AI existential risk" and the apparent inability to control AI are just a very convenient way to get away from answering the question of why they did't take basic precautions and kept on shoving these bots into every screen, app, and website anyway. To someone whose brain isn't totally marinated in Valley echo chambers obsessed with Roko's Basilisk and Bostrom's and Yudkowsky's navel-gazing, this sounds like AI startups saying they don't know what they're doing and never really did, which builds absolutely no confidence in something their bosses demand they use for everything.
Sure, the tech bro fanboy club on what Twitter has been turned into may throw fits if you point out that LLMs are not magic, and will be swayed by the braying of the ea/acc cult members, but your average person forced to fight with a chatbot demanding their personal and financial data because a platform wants to take over their lives for a fee, and another one just to do their job? Yeah, something tells me they're not going to be completely swayed by the "we lost all control" argument, or feel any confidence that the same tech bros saying this can be trusted with a plan to regain it.
