paper clips, parrots and safety vs. ethics

This past week, OpenAI CEO Sam Altman charmed a room full of politicians in Washington, D.C., over dinner, then testified for about nearly three hours about potential risks of artificial intelligence at a Senate hearing.

After the hearing, he summed up his stance on AI regulation, using terms that are not widely known among the general public.

“AGI safety is really important, and frontier models should be regulated,” Altman tweeted. “Regulatory capture is bad, and we shouldn’t mess with models below the threshold.”

In this case, “AGI” refers to “artificial general intelligence.” As a concept, it’s used to mean a significantly more advanced AI than is currently possible, one that can do most things as well or better than most humans, including improving itself.

“Frontier models” is a way to talk about the AI systems that are the most expensive to produce and which analyze the most data. Large language models, like OpenAI’s GPT-4, are frontier models, as compared to smaller AI models that perform specific tasks like identifying cats in photos.

Most people agree that there need to be laws governing AI as the pace of development accelerates.

“Machine learning, deep learning, for the past 10 years or so, it developed very rapidly. When ChatGPT came out, it developed in a way we never imagined, that it could go this fast,” said My Thai, a computer science professor at the University of Florida. “We’re afraid that we’re racing into a more powerful system that we don’t fully comprehend and anticipate what what it is it can do.”

But the language around this debate reveals two major camps among academics, politicians, and the technology industry. Some are more concerned about what they call “AI safety.” The other camp is worried about what they call “AI ethics.“

When Altman spoke to Congress, he mostly avoided jargon, but his tweet suggested he’s mostly concerned about AI safety — a stance shared by many industry leaders at companies like Altman-run OpenAI, Google DeepMind and well-capitalized startups. They worry about the possibility of building an unfriendly AGI with unimaginable powers. This camp believes we need urgent attention from governments to regulate development an prevent an untimely end to humanity — an effort similar to nuclear nonproliferation.

“It’s good to hear so many people starting to get serious about AGI safety,” DeepMind founder and current Inflection AI CEO Mustafa Suleyman tweeted on Friday. “We need to be very ambitious. The Manhattan Project cost 0.4% of U.S. GDP. Imagine what an equivalent programme for safety could achieve today.”

But much of the discussion in Congress and at the White House about regulation is through an AI ethics lens, which focuses on current harms.

From this perspective, governments should enforce transparency around how AI systems collect and use data, restrict its use in areas that are subject to anti-discrimination law like housing or employment, and explain how current AI technology falls short. The White House’s AI Bill of Rights proposal from late last year included many of these concerns.

This camp was represented at the congressional hearing by IBM Chief Privacy Officer Christina Montgomery, who told lawmakers believes each company working on these technologies should have an “AI ethics” point of contact.

“There must be clear guidance on AI end uses or categories of AI-supported activity that are inherently high-risk,” Montgomery told Congress.

See also: How to talk about AI like an insider

It’s not surprising the debate around AI has developed its own lingo. It started as a technical academic field.

Much of the software being discussed today is based on so-called large language models (LLMs), which use graphic processing units (GPUs) to predict statistically likely sentences, images, or music, a process called “inference.” Of course, AI models need to be built first, in a data analysis process called “training.”

But other terms, especially from AI safety proponents, are more cultural in nature, and often refer to shared references and in-jokes.

For example, AI safety people might say that they’re worried about turning into a paper clip. That refers to a thought experiment popularized by philosopher Nick Bostrom that posits that a super-powerful AI — a “superintelligence” — could be given a mission to make as many paper clips as possible, and logically decide to kill humans make paper clips out of their remains.

OpenAI’s logo is inspired by this tale, and the company has even made paper clips in the shape of its logo.

Another concept in AI safety is the “hard takeoff” or “fast takeoff,” which is a phrase that suggests if someone succeeds at building an AGI that it will already be too late to save humanity.

Sometimes, this idea is described in terms of an onomatopeia — “foom” — especially among critics of the concept.

“It’s like you believe in the ridiculous hard take-off ‘foom’ scenario, which makes it sound like you have zero understanding of how everything works,” tweeted Meta AI chief Yann LeCun, who is skeptical of AGI claims, in a recent debate on social media.

AI ethics has its own lingo, too.

When describing the limitations of the current LLM systems, which cannot understand meaning but merely produce human-seeming language, AI ethics people often compare them to “Stochastic Parrots.“

The analogy, coined by Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell in a paper written while some of the authors were at Google, emphasizes that while sophisticated AI models can produce realistic seeming text, the software doesn’t understand the concepts behind the language — like a parrot.

When these LLMs invent incorrect facts in responses, they’re “hallucinating.”

One topic IBM’s Montgomery pressed during the hearing was “explainability” in AI results. That means that when researchers and practitioners cannot point to the exact numbers and path of operations that larger AI models use to derive their output, this could hide some inherent biases in the LLMs.

“You have to have explainability around the algorithm,” said Adnan Masood, AI architect at UST-Global. “Previously, if you look at the classical algorithms, it tells you, ‘Why am I making that decision?’ Now with a larger model, they’re becoming this huge model, they’re a black box.”

Another important term is “guardrails,” which encompasses software and policies that Big Tech companies are currently building around AI models to ensure that they don’t leak data or produce disturbing content, which is often called “going off the rails.“

It can also refer to specific applications that protect AI software from going off topic, like Nvidia’s “NeMo Guardrails” product.

“Our AI ethics board plays a critical role in overseeing internal AI governance processes, creating reasonable guardrails to ensure we introduce technology into the world in a responsible and safe manner,” Montgomery said this week.

Sometimes these terms can have multiple meanings, as in the case of “emergent behavior.”

A recent paper from Microsoft Research called “sparks of artificial general intelligence” claimed to identify several “emergent behaviors” in OpenAI’s GPT-4, such as the ability to draw animals using a programming language for graphs.

But it can also describe what happens when simple changes are made at a very big scale — like the patterns birds make when flying in packs, or, in AI’s case, what happens when ChatGPT and similar products are being used by millions of people, such as widespread spam or disinformation.