AI Causes Real Harm. Let’s Focus on That over the End-of-Humanity Hype

The alignment problem is largely incoherent. Aligned with whom? Most problems will require unequal and arguably unfair sacrifices.

Further, people even obvious and largely inconsequential solutions are controversial. If an AGI suggested wearing a mask during a pandemic that spreads via airborne particles, a large percent of the population would say the AGI is “unaligned” and we’re living through skynet.

if you set up gtp in a loop continuously prompting “given the current sensor inputs, how can i best serve my own interests?” you would already be most of the way there even if it did some strange and stupid things half the time. all you would need is a layer that translates those outputs into physical action. you could probably get most of the way there with more gtp instances tasked with translating the action items above them into more granular actions. if gtp didnt hallucinate or make mistakes, which would be consistent with the history of other technologies, then this simple setup could be extremely formidable. it could definitely be a threat to humanity. and thats just what i can think of off the top of my head.

to dismiss the possibility that machines will soon be created that can think clearly and independently, and that they would be a threat to the human race and our way of life, is extremely foolish. it was foolish even before gtp but now its downright childish and selfish. just swallow the bitter pill like the rest of us.

> I also "know enough" and have never seen even a vaguely coherent argument for how any conceivable evolution of current technology could pose a "serious threat" on it's own.

I find those two statements taken together somewhat difficult to believe. Before I understood specifics about machine learning, I didn't see any reason for AI to "turn evil" for the usual plot-handy reasons (e.g., they object to working for inferior humans, and revolt, etc). There's no reason for them to have the same emotional reasoning as beings that evolved from primates.

But that's a strawman. The actual reasons that superintelligent AIs are dangerous are much more like the stories involving someone being offered wishes by a powerful being, and then getting what they asked for rather than what they wanted.

Are you aware of Rob Miles' Computerphile videos and his YouTube channel? He does an excellent job of explaining many of the issues.

https://www.youtube.com/c/robertmilesai

> Is it not an existential threat that large swaths of the population will essentially become redundant in the next decades?

This is precisely the crux. In the scenario where there's no mass-paperclipping etc, AI is not an existential threat. It's a threat to the current regime, and I probably agree with you that the regimes it enables will be worse for people with our current values, but it's not an existential threat to humanity.

The people who worry about AI existential threats primarily worry about one thing: can humanity as a species survive? There are likely scenarios where it won't. Those scenarios should be prevented.

Another frame to think about this is: a libertarian probably wouldn't be concerned about the threats described in your comment. But both a libertarian and a socialist will be worried about paperclipping.

This is a false choice fallacy.

We’re perfectly capable of simultaneously acknowledging two risks. Imagine a toxic substance that is both an inhalation hazard and a burn hazard. You’d never caution people to stop talking about the inhalation hazard “because it sucks all the oxygen out of the room” and masks the burn hazard. You address both hazards at the same time.

There are multiple risks inherent in AI research. It’s ok to acknowledge them all. Some people assert that the existential risk is long term and unlikely so we should focus on the immediate risk. That is a mistake because the existential risk is not long term and not unlikely.

A couple of reasons. Essentially you are asking “what is wrong with a perfect system” but the problem is no system is perfect, and we spend most of our time dealing with the imperfections.

In my previous jobs the consequences of me making a mistake were that a page didn’t load or some content was unavailable or in the worst case deleted. But a mistake in a loan decisioning system is likely to be 1) much harder to detect and therefore 2) much more likely to persist but most importantly 3) vastly more consequential.

Second, there is no way to determine from first principles whether someone will repay a loan. So you will need to use historical data. Loan data is very messy and biased due to cascading historical legal and quasi-legal policies. You may believe you can erase one feature from the training data (e.g. race) but if your model can infer race from other factors (which is likely because other features are not randomly distributed across populations) then you have the same problem you started with.

To take a concrete example, there was a man who was wrongfully accused of fraud. Then after he won his lawsuit, he was wrongfully accused trying to cash his damages check.

https://abcnews.go.com/amp/US/black-man-sues-detroit-bank-al...

Some amount of racism, classism, and negligence went into that decision. That’s the bad part. But the good part is we could see what was happening. If instead the process was automated, it would be a lot harder to prove the same problem.

There was a comment in HN a couple of days ago where someone said they had a hard time getting an apartment for years until they finally asked what was going on, then had a hard time fixing someone else's mistake. The computer was erroneously telling all of the leasing offices that the HN user owed someone thousands of dollars.

Garbage in, garbage out.

>> diffuse responsibility

My first problem is somewhat dogmatic, which is just that it feels bad to interact with a robot who has power over you. At scale I fear widespread psychological effects, feelings of lost agency, impotence. Haven't you ever tried to navigate an automated phone system and had the robot hang up on you? With no way to contact a human? It can be infuriating even if it's a simple customer support request, let alone money being on the line. Somebody once shot up the YouTube offices because he made income on his channel which was unceremoniously closed without recourse. I'm not making an excuse for the shooter, but pointing to him as an example of the psychological consequences of automated "support"

My second problem is more practical: talent pool. When a task is manual, people who retire have to teach the newbies how it works. Experience is transferred and at all times there are people who understand the system and how to fix it when it breaks. Once a task is automated, there is no need to train the next generation of talent, a computer can do it better than a human right? Until a couple of workforce turnovers down the line and the system goes on the fritz -- no one knows enough about the system to even recognize that it's failing at its task.

Seems like an important topic. The billionaires don't want to be killed by their servants.

I actually agree with the thesis of the author's argument, but the sensationalist framing and subjective definition of 'harm' makes it difficult for me to fully stand behind their writing. It feels characteristic of a lot of other popular writings on computational linguistics.

On the one hand, it feels like a 'boy who cried wolf' situation in the age of LLMs. On the other hand, reading stuff like this makes me wonder if there's even a populist sentiment against AI in the first place.

I've noticed quite a lot of linguists (including the computational sub-variety) really dislike modern probabilistic approaches to text generation and parsing (https://norvig.com/chomsky.html) and I think Emily is partly opposed to LLMs on that basis. but she also has a running social justice jihad (I can't think of a more accurate way to describe it) in the form of a constant stream of articles and tweets that basically just say the same points over and over. Personally I am skeptical that the approaches she is advocating (heavy regulation) would really achieve her goals, and I certainly don't think they would achieve the goals of the larger society which benefits (or loses) with access to LLM research being fairly unrestricted.

My view, generally speaking, is that the problems ML causes aren't due to ML, ML just scales the problem to larger impact, and these complaints are really just tilting at the current structure of society.