July 6, 2023
In this week’s start-of-summer issue of Talking to Computers: The Email, news is slowing down some as people head to the beach, but we look at whether AI is becoming so intelligent it will destroy all of humanity, or if maybe it’s getting dumber. Plus the end of prompt engineering is over (not really), a Presidential candidate lets you speak to his chatbot double, and more…
As always, if you find anything interest, please send it along.
AI is Coming for Humanity
“The end of man is knowledge, but there is one thing he can't know. He can't know whether knowledge will save him or kill him. He will be killed, all right, but he can't know whether he is killed because of the knowledge which he has got or because of the knowledge which he hasn't got and which if he had it, would save him.”
Introducing Superalignment (OpenAI)
Be it Resolved, AI research and development poses an existential threat
Douglas Hofstadter changes his mind on Deep Learning & AI risk?
OpenAI announced this week a new effort to “ensure AI systems much smarter than humans follow human intent.” They are creating a new team and dedicating 20% of their compute to this project. Indeed, it makes sense for them to push this message. “We are building something so powerful that it is dangerous, please buy it from us, we’re also the ones who can keep it under control.”
Meanwhile, a recent interview with Douglas Hofstadter (author of the highly recommended Gödel, Escher, Bach) touches on the same subject. Hofstadter first wonders if there is some sort of personhood emerging with highly intelligent AI:
But in the case of more advanced things like ChatGPT-3 or GPT-4, it feels like there is something more there that merits the word "I." The question is, when will we feel that those things actually deserve to be thought of as being full-fledged, or at least partly fledged, "I"s?
He’s pessimistic about the implications:
And that suggests that these entities, whatever you want to think of them, are going to be very soon, right now they still make so many mistakes that we can't call them more intelligent than us, but very soon they're going to be, they may very well be more intelligent than us and far more intelligent than us. And at that point, we will be receding into the background in some sense. We will have handed the baton over to our successors, for better or for worse.
It’s similar to a recent Munk Debate on the same topic between Yoshua Bengio, Max Tegmark, Yann Lecun, and Melanie Mitchell. Whichever side you’re currently on in the debate (and if you’re reading this newsletter, I can probably guess), it’s worth a listen.
At the same time, you have people complaining on forums that ChatGPT is getting less intelligent over time. (This is tongue in cheek, by the way. Any decrease in performance is likely due to prompts and guardrails more than anything else.) On the other other hand, OpenAI also did just shut down some functionality because it “inadvertently” did what users wanted it to do, which is… maybe the first time I’ve ever heard that.
I don’t think anyone knows for sure if AI will wipe out humanity, but I think we’ll have a good idea on how much of a concern this is in the coming years. And I trust human ingenuity, so I’m not stock piling food and matches just yet…
Prompt Engineering is Over
In this post, the author claims that prompt is on its way out. AI is getting more powerful and other predictions (without much backing) of what will or will not be necessary for LLMs in the near future.
What is necessary, however, he claims is problem formulation:
It’s all about how we humans identify, analyze, and define problems. When we can illustrate the problem, AI will provide the most efficient way to counter it.
Except… that’s what prompt engineering is. Prompt engineering is not some secret words or phrases like “open sesame.” It’s breaking down a problem and explaining it plainly. And that’s a skill that, quite frankly, has been important for thousands of years and will be for thousands more (unless, of course, AI wipes us out).
Other Links
Workers with less experience gain the most from generative AI
Bankrate Posts AI-Generated Article, Deletes It When We Point Out It's Full of Errors (Comment: people keep forgetting that LLMs hallucinate)
Building Boba AI: Some lessons and patterns learnt in building an LLM-powered generative application (Comment: very in-depth article)
Natural Language Is an Unnatural Interface (Comment: like I’ve been saying)
The Ad Industry Has No Intention of Letting AI Ruin the Party (Comment: least surprising news ever)
WinGPT: AI Assistant for Windows 3.1 (Comment: love it)
Mayor Suarez launches an artificial intelligence chatbot for his presidential campaign (Comment: really just a semantic search engine, but that’s okay)
What are embeddings? (Comment: a great resource if you’d like to dig deeper)