simonmesmith

Wholeheartedly agree. And I think the problem has gotten worse for a few reasons. One is that if you’re interested in a subject, you likely follow it closely online, so reading a book-length treatment feels like learning a lot of stuff you already know. Two is that publishers try to find authors with social media audiences who likely already published all their good ideas for free and now have to dredge for additional material that’s usually less interesting.

When I was a child and teenager I loved science fiction. When I became an adult I shunned fiction as a waste of time and read only nonfiction books. Then I started to find that most nonfiction books could be summarized in a blog post or Wikipedia entry, regurgitated well established ideas, or spewed unfounded bullshit. I’ve since flipped my perspective again. I think good fiction is worth reading at book length while most nonfiction is better consumed in blog posts, articles, journal papers, or via a ChatGPT tutoring session.

I will highlight one consistent exception. I love every book I’ve read by Yuval Noah Harari.

Probably the two most telling phrases:

“Almost all scenarios that limit warming to 1.5°C or 2°C require novel CDR.”

“Virtually all current CDR (99.9% or 2 GtCO₂ per year) comes from conventional management of land, primarily via afforestation and reforestation.

Only a tiny fraction (0.1% or 0.002 GtCO₂ per year) results from novel CDR methods which involve storing captured carbon in the lithosphere, ocean or products.”

So, long way to go if we think this is what we need.

Good ideas posted in response to this. Perhaps we need to try just rearranging images of specific things. Like, rearrange a bird and see if it can figure out that. Above I jumbled up everything and that’s different than what they did with the text, where they grouped by words.

It strikes me that this opens or at least expands an attack vector for prompt injection. Humans might overlook a scrambled block of text that GPT-4 will interpret and then execute.

Also, I was curious about whether this would also work with images. So I chopped an image of a bird (a night heron) on a boat in my own collection (so, not in GPT-4's training data) into squares and randomly rearranged them, then tried to get ChatGPT-4 to interpret it. It doesn't work. I got this:

--- CHATGPT RESPONSE:

The image you've uploaded appears to be a composite of several smaller images that are cut and arranged in a grid-like pattern. This collage-like arrangement makes it difficult to discern a clear, singular subject. However, I can identify fragments of what seems to be a bird—possibly a magpie or some type of corvid—due to the black and white plumage. There are also elements that look like parts of a silver car, like a side mirror, and segments of wooden structures that could be parts of a bench or a deck.

The way the images are scrambled resembles a puzzle, where each piece needs to be placed correctly to form a coherent whole. This type of image could be used as a challenge to reconstruct the original picture, requiring the viewer to mentally rotate and arrange the pieces.

---

Maybe a different approach to scrambling an image would work.

Just a heads-up that there isn’t much new in here if you’ve been following the saga on Twitter and have seen Brockman’s and Swisher’s tweets and are aware of the other recent departures. It’s mainly a summary of tweets. Not criticizing, just trying to save you a click if you’re following in real-time.

I’m interested to see this happen but “as soon as” approximately June 2024 doesn’t feel that soon. It’s about as far from today as today is from the launch of GPT-4. Think of everything that’s happened since then. And things only seem to be speeding up. Still, I welcome a Siri that’s useful beyond setting timers.

Having worked on this problem in biology, I think one of the challenges you’ll find is that the knowledge graph will be extremely context-dependent and biased towards highly probable nodes and edges.

For example, if you ask an LLM to create a graph of all proteins related to X disease, and show how they interact, it will oblige. (You can try this yourself easily in the OpenAI playground. Just ask it to send you back a list like X -> Y -> Z or whatever. Or an array of source/target/relation triplets.)

The challenge is that what you get will be very dependent on how you phrase your request. So you’ll never know if you’re getting a “complete” graph or just the most probable graph for the request you made. If you’re an expert in the domain, you’ll know, but if you’re an expert you might not need the graph in the first place.

Yeah, but I wonder if it’s even more powerful to say, “we asked the world’s most powerful AI and it recommended that we lay off 20% of our staff, while ensuring we treated them all fairly.”