AI Week Feb 12th: Does ChatGPT understand what it's saying?
Hi! In this week's AI week: Does ChatGPT understand what it's saying? (1000 words, about a 7-minute read)
What elephant?
But first, we need to talk about Sam Altman's ask for $7 trillion.
SEVEN TRILLLLLLIONN DOLLARS
The Wall Street Journal reported that ChatGPT maker OpenAI's CEO, Sam Altman, wants to raise 5-7 trillion dollars to build chip foundries for AI chips.
Both training AI models and using them requires special, expensive AI chips. (BTW, this also uses a lot of power and a lot of water.) And ChatGPT isn't even very smart. But Altman's OpenAI wants to develop artificial general intelligence (AGI), capable of human-level reasoning or beyond. Reasoning by analogy (which ChatGPT is pretty good at), if ChatGPT-level "intelligence" takes a lot of NVidia chips, power and water, developing AGI will need thousands of times as many chips (and power and water--which is a whole other issue).
To put this in perspective, $7 trillion is a mere $700 billion more than the US's annual budget. It's more than three times Canada's annual GDP. Or, as the Register notes, "it's enough cash to gobble up Nvidia, TSMC, Broadcom, ASML, Samsung, AMD, Intel, Qualcomm, and every other chipmaker, designer, intellectual property holder, and equipment vendor of consequence in their entirety – and still have trillions left over."
Who among us doesn't want the GDP of several countries devoted to realizing their dreams? As Gary Marcus asks, what could possibly go wrong?
P. S.: Altman snarked back on Twitter/X to all the blowback, Xeeting "you can grind to help secure our collective future or you can write substacks about why we are going fail [sic]"
Do LLMs understand what they're saying?
It's pretty clear to me that image-generating AI doesn't "understand" the images it makes, or we wouldn't get images like the one above. But we humans are extremely language-oriented, and we have cognitive biases toward thinking that anything that talks must have a mind. (And for good reason: for our species' entire evolution, this was true.) There's an ongoing argument over whether or not LLMs like ChatGPT understand what they're saying, for any value of the word "understand."
The two sides of the debate can be summed up as:
1. Stochastic parrots
Large language models (LLMs) are "stochastic parrots", as Drs. Emily Bender and Timnit Gebru put it in this 2021 paper. They aren't understanding in any sense of the word. They generate text that is coherent, but isn't actually communicating anything.
[A LLM] is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.
2. Not stochastic parrots
The 2021 "stochastic parrots" paper has become kind of a touchstone for researchers to formulate "oh no it isn't" hypotheses against. These hypotheses have been evolving with continued research into LLMs' capabilities. I can't do all the positions, hypotheses and research justice, but here are some highlights from the past few years.
2022: It's alive!
In 2022, Google researcher Blake Lemoine was fired after going public with his claims that Google's chatbot, LaMDA, was sentient. (The consensus of basically everyone else was, and is, that LaMDA is not sentient.)
2023: OK, it's not alive. But it's intelligent!
LLMs like ChatGPT seem very intelligent at first blush, as they're able to write code, play chess, do well on the SAT exam, etc. However, a number of tests with GPT-3.5 and GPT-4 found that their abilities are more consistent with memorization and pattern matching. For example, LLMs trained on "A is B's parent" fail to learn "B is A's child".
https://towardsdatascience.com/is-chatgpt-intelligent-a-scientific-review-0362eadb25f92024: OK, it's not intelligent. But maybe it understands what it's saying.
Some research suggests that LLMs can "learn" meaning from the texts they're trained on.
Other recent research finds that LLMs aren't just mimicking their inputs, but combining them in a way that hints at understanding--specifically, in a way that is consistent with their having learned competence in linguistic skills like "using irony" and "understanding the word because", and combining them in ways unlikely to be present in the training data.
It's a really clever argument; I can't help thinking, though, that the elephant in the room is literally the elephant in the room at the top of this newsletter.
Quanta Magazine
Far from being “stochastic parrots,” the biggest large language models seem to learn enough skills to understand the words they’re processing.
Longread: But what about ToddlerAI?
What if we trained an AI the way humans learn language: would it understand words with reference to the real world, as humans do? NYU researchers tried this, training a neural network by exposing it to 61 hours of audio and video from a toddler's point of view, including things like a parent naming and describing object. Like a toddler, the AI was able to learn some simple words and images.
Yann LeCun, head of AI research at Meta, argues that superhuman AI must first learn to learn more like humans, and that research must understand how children learn languages and understand the world with far less energy and data consumption than today's AI systems.He is trying to create systems that can learn how the world works, much like baby animals.
NYU researchers develop AI that mimics a toddler's language learning journey
Researchers at New York University have developed an AI system that might be able to learn language like a toddler. The AI used video recordings from a child's perspective to understand basic aspects of language development.
BONUS: AI helping to fight climate change
This week's cool AI application: AI could help fight climate change in several ways, not the least of which is convincing climate skeptics that climate change is real simply by chatting with them.
