GPT/LLM links, 5/15

Haidt and Schmidt predict AI twitter trolls; Ethan Mollick on LLMs as human; A Stanford paper on LLMs as not so emergent; Patrick Collison and Sam Altman; The Zvi refutes me;

Jonathan Haidt writes,

Eric had been thinking about the interactions of AI with social media, and once I shared my perspective on the social psychology of that interaction, it became clear to both of us that generative AI could make social media much, much worse. Given the concerns that Eric described above, it seemed likely that AI was going to super-empower bad actors by giving them each an army of assistants, and it was going to supercharge intergroup conflict by drowning us all in high-quality video evidence that the other side is worse than Hitler. 

If ever there were a time when voters would seem to be ready for a third party, this is it. But it won’t happen. Already, William Galston argued,

if a No Labels ticket receives even a tiny share of the vote in key states, Mr. Trump could end up back in the Oval Office. No Labels leaders have promised to end their campaign if it becomes clear that their ticket will be a spoiler. The sooner they reach this conclusion, the better.

So Trump is worse than you-know-who. And this comes from a relatively sane columnist, not an AI, writing in the WSJ, not on Twitter.

Haidt and Schmidt’s recommendations for social media:

1. Authenticate all users, including bots

2. Mark AI-generated audio and visual content

3. Require data transparency with users, government officials, and researchers

4. Clarify that platforms can sometimes be liable for the choices they make and the content they promote

5. Raise the age of “internet adulthood” to 16 and enforce it

To me, (4) and (5) sound terrible. (4) converges on censorship of anything that government courts might not approve. (5) keeps kids from escaping government schools. Of course, I haven’t read their article ($Atlantic), so maybe the details make their ideas more palatable.

Ethan Mollick writes,

We want our software to yield the same outcomes every time. If your bank’s software mostly works, but sometimes scolds you for wanting to withdraw money, sometimes steals your money and lies to you about it, and sometimes spontaneously manages your money to get you better returns, you would not be very happy. So, we ensure that software systems are reasonably reliable and predictable. Large Language Models are neither of those things, and will absolutely do different things every time. They have a tendency to forget their own abilities, to solve the same problem in different ways, and to hallucinate incorrect answers.

I think that this scares software engineers. They are more likely to be systemizers than empathizers. To explain these concepts, I turned to ChatGPT-4.

Simon Baron-Cohen is a prominent psychologist and researcher in the fields of developmental psychology and autism, and he is known for his work on understanding the differences in cognitive styles between individuals.

One of the key concepts in Baron-Cohen's work is the idea of systemizing and empathizing, which refers to two distinct cognitive styles that individuals may exhibit.

The concept of "systemizing" refers to an individual's ability to understand and analyze complex systems, such as mechanical systems, natural phenomena, or abstract concepts. Systemizers have a natural tendency to seek out patterns, predictability, and logical explanations for things. They may be highly skilled in fields such as engineering, mathematics, or computer science, and they may excel in tasks that require attention to detail and accuracy.

On the other hand, the concept of "empathizing" refers to an individual's ability to understand and connect with the emotions and mental states of others. Empathizers are highly attuned to the feelings and needs of others, and they may be skilled in fields such as social work, counseling, or psychology. They may be intuitive in their interactions with others and have a natural ability to empathize with and support others.

According to Baron-Cohen, individuals can exhibit varying degrees of systemizing and empathizing abilities, and these abilities may be influenced by a number of factors, including genetics, early childhood experiences, and socialization.

Baron-Cohen has used the concepts of systemizing and empathizing to study a range of phenomena, including autism spectrum disorder (ASD), which he has hypothesized is associated with an extreme form of systemizing and a relative weakness in empathizing abilities.

Overall, the concepts of systemizing and empathizing have been useful in helping researchers understand the cognitive and social differences between individuals, and they continue to be an important area of study in the field of psychology.

I highlighted the term “predictability.” I believe that a lot of the fear of LLMs comes from their unpredictability. It used to be that when you got unexpected results from software, you looked for a bug. Even programs that used a random number generator gave predictable output once you knew what number they started with.

The output of LLMs varies in ways that upset those who are way over on the systemizer scale. I think that accounts for why some of the cries of alarm about LLMs come from software engineers. Unpredictability of LLMs reminds software engineers of human beings, who can be confusing and scary to someone way at the systemizer end of the spectrum.

But a recent paper by Stanford computer science folks says that LLMs are tamer than many techies believe.

the perception of AI’s emergent abilities is based on the metrics that have been used. “The mirage of emergent abilities only exists because of the programmers' choice of metric,” Schaeffer says. “Once you investigate by changing the metrics, the mirage disappears.”

…What it means for the future is this: We don’t need to worry about accidentally stumbling onto artificial general intelligence (AGI). Yes, AGI may still have huge consequences for human society, Schaeffer says, “but if it emerges, we should be able to see it coming.”

Patrick Collison’s interview of Sam Altman is at least as good as what Tyler Cowen would expect. For instance, Patrick asks Sam what would move his probability of a doom scenario. Sam says

what you really want is to understand what's happening in the internals of the models

To me, that sounds like he does not like the unpredictability of the LLMs, either. He goes on to say

what the world needs is not more uh AI safety people who like post on Twitter and write long philosophical diatribes it needs more people who are like going to do the technical work to make these systems safe

Zvi Mowshowitz writes,

Dustin Muskovitz predicts that by 2030 ~everyone will have a personal AI agent to do their paperwork for taxes, government services, health care forms and so on. I’d expect this to move quicker than that, although the future is always unevenly distributed

Even if that is true, can you imagine the amount of paperwork and regulation will be induced when compliance becomes easier? It’s kind of like the software developer’s lament: “What’s the point of trying to build an idiot-proof system? They’ll just build a better idiot.”

The Zvi swats away lots of bad arguments from AI non-doomers, including mine. One of his points:

the AI being allowed to use ‘hire people to do arbitrary thing’ [is a way for an AI to add a capability]. In many ‘the AI won’t be a threat’ scenarios, we forget that this is an easy option for whatever the ‘missing stair’ is in the plan.

That means that if we are looking for a capability the AI won’t be able to obtain, it has to be a capability that requires millions of people. Like producing a pencil, an iPhone, or an AI chip? Without the capability to undertake specialization and trade, an AI that destroyed the human race would also self-destruct.

The Zvi also has a very terse summary of the Collison-Altman podcast. I found the ChatGPT Chrome extension for transcribing podcasts more helpful.

Substacks referenced above:

@

AI #11: In Search of a Moat

@

AI is not good software. It is pretty good people.

@

AI Will Soon Make Social Media Much More Harmful to Liberal Democracy, and to Children