Human Extinction, Hallucinations, PDF Tools, and The Alignment Problem
Human Extinction, Hallucinations, PDF Tools, and The Alignment Problem
This Week in AI
This week’s update:
Here’s this week’s news, product applications, and philosophical implications. So here we go:
🦴 Human Extinction
💭 Hallucinations
🇯🇵 Japan Goes All In
📄 PDF Tools
↔️ The Alignment Problem
Latest News and Updates
Experts are warning AI could lead to human extinction. Are we taking it seriously enough?
This week, we saw many people again voice the concern of major repercussions to AI:
On Tuesday, hundreds of top AI scientists, researchers, and others — including OpenAI chief executive Sam Altman and Google DeepMind chief executive Demis Hassabis — again voiced deep concern for the future of humanity, signing a one-sentence open letter to the public that aimed to put the risks the rapidly advancing technology carries with it in unmistakable terms.
“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war,”
I posted previously about how we are facing a moment similar to the creation of atomic weapons, with all the destructive capabilities. As humans, we nearly took ourselves to the brink with that one:
Tiktok failed to load.Enable 3rd party cookies or use another browser
OpenAI is pursuing a new way to fight A.I. ‘hallucinations’
I was generating a document this week, and after feeding in information about my background, ChatGPT generated a few paragraphs about me. Surprisingly, it included the phrases “despite not having an MBA” and “having 5 years of product experience.” I do, in fact, have an MBA and have 15 years of product experience. It was strange to see ChatGPT generate false information about me, when I literally had input correct information moments before.
This problem isn’t unusual. And OpenAI is working on ways to fight it:
AI hallucinations occur when models like OpenAI’s ChatGPT or Google’s Bard fabricate information entirely, behaving as if they are spouting facts. One example: In Google’s own February promotional video for Bard, the chatbot makes an untrue claim about the James Webb Space Telescope.
Japan Goes All In: Copyright Doesn’t Apply To AI Training
Japan has decided to go all in for AI.
In a surprising move, Japan’s government recently reaffirmed that it will not enforce copyrights on data used in AI training. The policy allows AI to use any data “regardless of whether it is for non-profit or commercial purposes, whether it is an act other than reproduction, or whether it is content obtained from illegal sites or otherwise.”
We’ll see if this sets a precedent for other countries as well. It certainly has the potential to unleash a lot of generative AI training.
It seems the Japanese government believes copyright worries, particularly those linked to anime and other visual media, have held back the nation’s progress in AI technology. In response, Japan is going all-in, opting for a no-copyright approach to remain competitive.
A majority of Americans have heard of ChatGPT, but few have tried it themselves
It’s easy to think that everyone has heard of ChatGPT, AI tools, etc. And that everyone is using them. But that’s just not the case. When we’re surrounded by these tools, and other people who are tech savvy, it’s easy to think everyone knows about them. We forget that most people don’t adopt technology as quickly as those of us at the forefront. According to Pew Research:
About six-in-ten U.S. adults (58%) are familiar with ChatGPT, though relatively few have tried it themselves, according to a Pew Research Center survey conducted in March. Among those who have tried ChatGPT, a majority report it has been at least somewhat useful.
However, few U.S. adults have themselves used ChatGPT for any purpose. Just 14% of all U.S. adults say they have used it for entertainment, to learn something new, or for their work. This lack of uptake is in line with a Pew Research Center survey from 2021 that found that Americans were more likely to express concerns than excitement about increased use of artificial intelligence in daily life.
Useful Tools & Resources
This week I explored tools to help summarize and analyze PDFs. I’ll be creating a library of these tools soon, so remember to check back for everything soon.
ChatPDF
This was one of my favorite tools. It analyzed the documents I uploaded and then prompted me with questions I could ask:
Tiktok failed to load.Enable 3rd party cookies or use another browser
PDF.AI
This was another really useful tool as well. I uploaded a research paper (the same as above) and was able to ask questions.
Tiktok failed to load.Enable 3rd party cookies or use another browser
ChatWithPDF
There are a number of other tools as well, including plugins for ChatGPT like ChatWithPDF. Unfortunately when I used this plugin, it very confidently gave me a lot of incorrect information for the PDF. So, as with everything, you need to have some knowledge of the subject to check the work.
Deep Dive - The Alignment Problem
I dove into the book The Alignment Problem by Brian Christian this week. If you missed it, definitely check out the podcast above and the post here.
The heart of the alignment problem is ensuring that our machines—the AI we create, the algorithms and models—understands our objectives and values. That it understands the intent behind the instructions, rather than just the text it may see.
As Quanta magazine puts it:
Computers frequently misconstrue what we want them to do, with unexpected and often amusing results. One machine learning researcher, for example, while investigating an image classification program’s suspiciously good results, discovered that it was basing classifications not on the image itself, but on how long it took to access the image file — the images from different classes were stored in databases with slightly different access times. Another enterprising programmer wanted his Roomba vacuum cleaner to stop bumping into furniture, so he connected the Roomba to a neural network that rewarded speed but punished the Roomba when the front bumper collided with something. The machine accommodated these objectives by always driving backward.
These examples are amusing. We’ve seen similar problems in video games when we put AI to work. We may give points for not dying, which leads to the character never moving. Or give points for possessing the ball in a soccer game, which leads to lots of possession and never to scoring goals.
What happens when there is more at stake?
But the community of AI alignment researchers sees a darker side to these anecdotes. In fact, they believe that the machines’ inability to discern what we really want them to do is an existential risk. To solve this problem, they believe, we must find ways to align AI systems with human preferences, goals and values.
Of course, it’s not enough to understand goals:
If you believe that intelligence is defined by the ability to achieve goals, that any goal could be “inserted” by humans into a superintelligent AI agent, and that such an agent would use its superintelligence to do anything to achieve that goal, then you will arrive at the same conclusion that Russell did: “All that is needed to assure catastrophe is a highly competent machine combined with humans who have an imperfect ability to specify human preferences completely and correctly.”
This is the classic paper clip problem many of us are familiar with. As New Scientist puts it:
This goes back to 2003, when Nick Bostrom, a philosopher at the University of Oxford, posed a thought experiment. Imagine a superintelligent AI has been set the goal of producing as many paper clips as possible. Bostrom suggested it could quickly decide that killing all humans was pivotal to its mission, both because they might switch it off and because they are full of atoms that could be converted into more paper clips.
The way to solve for this won’t be easy, but will require everyone working together and addressing the problem. As Paul Christiano, a researcher at OpenAI, put it on Effective Altruism:
“Rather than taking our initial teacher to be one human and learning from that one human to behave as well as a human, let’s start by learning from a group of humans to do more sophisticated stuff than one human could do.” That could give us an AI, which isn't just “human-level,” but is the level of a human group.