The doom talk gives AI user base a hard-on and Anthropic knows it. It's like saying, your web browser is now equipped with the power of plutonium. You are controlling something powerful and dangerous as you work to nudify your neighbor or slop out of your term paper.
Cory Doctorow shared the exact same opinion in this interview he gave this week. I personally don't think his "reverse centaur" idea is nearly as catchy or worth adopting as enshittification was, but the intent behind it is spot on.
Reverse centaur is a terrible term. Even having heard it explained, repeatedly, I could not even explain what it means. It's not catchy, too many weird syllables; the word "centaur" makes it sound nerdy as fuck for any casual use, I doubt most normal people even remember what a centaur is; and it just does not immediately convey it's meaning as a result.
What is it even supposed to mean?
centaur: brain of a man, speed of a horse. reverse centaur: brain of a horse, speed of a man.
but the horse is the computer.
The horse is the computer...?
Are they saying that we can trust salesmen to tell us the truth?
Archive link: https://archive.ph/D2uB0
(I want to preface by saying I don't support AI, I don't want to hype AI, I want AI research and development to stop immediately.)
Timnit Gebru argues that AI companies are stoking fears of extinction to avoid discussing actual harms, like autonomous weapons.
Huh? Pretty sure autonomous weapons are part of the fears of extinction. It's very hard to imagine an AI safety advocate doing "all the doom talk" would in any way support AI autonomous weapons. She and other AI safety researchers are on the same side! They don't need to and shouldn't fight each other.
Timnit Gebru: Why are they choosing certain disciplines? Somehow these subjects, like programming and chess and math, were elevated to mean, OK, you solved this and thereby you’ve solved intelligence.
Well that's simple. Because the output of chess, programming, and math can be easily checked mechanically. Chess has win or loss, programs either pass tests or not, math proofs can be checked with Lean, etc. This makes these discipline very easy to train an AI on with reinforcement learning.
I thought an AI researcher would know this?
It's a matter of scale though.
Pushing the idea of "AI will make TERMINATOR CYBORGS!!!"
distracts from
"AI is analyzing a zillion cameras so ICE gestapo goons can abduct people.'
She answers that question right after the question in the same article, making it a rhetorical device rather than an open question.
Focusing on human responsibility for using a tool sounds like a good way to proceed in my opinion.
When AI steal open source code who is responsible? When an AI powered drone kills a civilian who is responsible? When the AI is shown to hire and fire in racist ways who is responsible? Someone has to be, AI is just a tool after all.
In my book I write, “Can a bridge decide to collapse?” The whole “Oh no, our models went rogue” thing … When a bridge collapses, you don’t analyze whether the bridge was ethical or sentient or why it decided to collapse. You ask, who is the person who built this bridge to be so flimsy? There are permits you have to acquire. There are supposed to be tests. When you ask about the bridge collapsing itself, you’ve lost the thread.
Meanwhile, what’s really existential is AI powering autonomous weapons, killing machines, that are actually being used in warfare. The climate catastrophe is a real thing that could be exacerbated by what these tech companies are building. Bosses are using AI as an excuse to get rid of their pesky workers. This machine-god narrative is, in my opinion, meant to distract us from these.
The problem is AI is created by rich people and rich people cannot be held legally liable. The Hugging Face hack is clear evidence of this legal principle.
And "AI gone rogue" headlines support the narrative that the AI did things by itself and no human could expect this so they can't be responsible (nevermind that they built the thing in the first place).
they turned off the safeguards. This shit is 100% idiot humans + marketing
And they gave it unlimited compute, I heard, though I can´t find the source on that. Imagine sandboxing something that has the knowledge to get out of a sandbox, telling it to hack a server, giving it unlimited compute, and then just walking away for several days. It is absolutely insane.
They didn't walk away. They sat there and watched it. All this is to sell "the ultimate hacking machine" to the government.
Right, but that's precisely the problem. We don't need to be going around fearmongering about how dangerous AI itself is. We need to be regulating it and fining the actual people who are making decisions to use AI in ways that result in harm. Whether that harm is direct human injury, injury through negligence (like denying a health insurance claim that should have been approved), financial injury, reputational, whatever.
They need to be held accountable exactly the same as if a human had done it. Humans have been driven to suicide by the criminal response to hacking for far more noble a reason than "cheat on an AI evaluation test". The weight of the law should be brought down on the executives responsible for the Hugging Face hack, too. But fearmongering about how this shows how amazeballs advanced the AI has become doesn't accomplish that.
Yes, but what I'm telling you this has absolutely nothing to do with AI and everything to do with the impunity of wealth.
All these AI companies MASSIVELY pirated the internet for training data, all very clearly illegal by the letter and spirit of the law, and they have not even been charged for it. The most that happens is that content companies sue in the hopes to get some shares in AI, or sue one AI company because their daddy owns a competing AI company, and those suits seem to ignore the piracy and instead are about whether they needed a separate license to train.
Meanwhile, AI companies are very carefully trying to put liability sponges between it and any lawsuits. Tesla tried (and recently failed) to dump 100% of liability on drivers of it's "self-driving" cars, and every mission critical AI service is likely to have a human mindlessly rubber stamping AI decisions as fast as they can just so they can be held liable instead of the employer or the AI company.
Focusing on human responsibility for using a tool sounds like a good way to proceed in my opinion.
Absolutely! If AI hurts humans then the AI company CEOs should be put in jail (in fact I believe they should be in jail right now for developing AI at all).
When a bridge collapses, you don’t analyze whether the bridge was ethical or sentient or why it decided to collapse. You ask, who is the person who built this bridge to be so flimsy?
Right, when a bridge collapses, we analysis the physics of it so the next bridge we build won't collapse. From my perspective analysis why an AI "went rogue" is the same process. The difficulty is when talking about AI we can't help but use some anthropomorphizing languages. If anything, talking about when an AI "thought" is a lot less verbose than talking about how the transformer's weights multiplies together and how the random sampling process produced a sequence of tokens that resembles natural language. But ultimately, it's not that far from analyzing a bridge collapse. We need to understand why an AI "went rogue" to stop the next AI from going rogue (or, in my opinion, just stop building the next AI altogether).
This machine-god narrative is, in my opinion, meant to distract us from these.
I don't see why personally. AI is dangerous enough even if it doesn't become "machine-god". Absolutely we shouldn't use AI to develop autonomous weapons. If AI can indeed somehow become "machine-god", then oh my god that's a million times worse and we should stop immediately.
The difficulty is when talking about AI we can’t help but use some anthropomorphizing languages
I can. And the tech ceos choose that language because it makes them more money.
That's what this is about, it's about humans making choices and avoiding responsibility for those chioces by blaming machines.
The only difference is that AI is a black box and a lot of what's going on inside isn't fully understood. They're learning as they go, and accepting they don't know some of the whys as they go to the next step. Unlike the bridge analogy. Which actually makes the problem even more serious, because they don't fully know why it ends up doing some of the things, and now the newest models are having the processes partially hidden from view. Where only a version back we were being assured they could watch the chain of thoughts and intercept what could be problematic, claiming they're practicing AI safety in development, now there's pieces that are not viewable. And they went with it anyway and released it to paying customers. Again with the bridge analogy, some of the parts that might have failed go missing and we don't know how they broke, what they were made of, etc.
Humans are building AI, but in a sense they are also growing them, while not knowing what seeds they're throwing together. There's science and theory behind it, they understand a lot, but they're getting results back that would prevent a bridge being approved to be built if it had the same uncertainty. So I stand behind holding the people accountable, but more for the recklessness of following profit over scientific research and ethics, as well as using the excuse that if they don't get there first, someone else will. They aren't wrong about that, only wrong about it needing to be that type of race.
In the discussions our project are having about how to approach LLM use in the community there has been a lot of consensus in the human taking ownership. Ultimately it's the human who has to direct the tool and if they have little interest other than blindly hitting accept then we're not interested in the contribution.
This is a most excellent place for technology news and articles.
