Posts
Wiki Contributions
Comments
At my first attempt to upload the images I just drag-and-dropped the image files to the post, without editing file names. But GPT4 saves the images with the file names like this:
DALL·E 2023-10-05 12.12.06 - Illustration in anime style Inside a vast cave, Dr. Ada Worthington, a palaeontologist with brunette hair and a pink bow in it, has an intense gaze a.png
Perhaps LW internally is saving the original file names, causing the problems.
I reuploaded the images with the filenames like 1.png, perhaps it will help.
Julian Hazell (distinct thread): “Why would you think AI will end up taking control?”
“We will give it to them”
A personal anecdote on the topic:
A few days ago GPT4 and me were debugging a tricky problem with docker. GPT4 suggested to run a certain docker command. As usual, I was going to copy the output and give it to GPT4. The output was a long nested json. I then noticed that the json contains the admin credentials. It was really easy to miss the part and just paste it for GPT4 to consume.
So, I almost gave GPT4 the admin credentials, which would potentially allow her to hack my app.
With many thousands of software developers doing similar things with GPT4, there will certainty be the cases where the developer wasn't attentive enough.
This means, for the AI to break from her digital prison, she doesn't need to do superhuman hacking, to exploit zero day vulnerabilities etc. All she has to do is to try the accidentally leaked credentials.
There is a very short path from "the AI wants to escape" and "the AI is all over the Internet". She doesn't even need to have a human-level intelligence for that.
I selected "Quickly orienting to novel situations" (QOTNoS) because it's strictly superior to the alternatives:
- If you have the QOTNoS virtue, you can deal with the novel situation of AI destroying civilization
- The necessity of "accurately reporting your epistemic state" is a novel situation for most people. QOTNoS helps again.
- "Resisting social pressure" is a common situation. But if the survival of civilization depends on it (as the poll implies), this is a novel situation. Thus, QOTNoS will help with it.
In essence, QOTNoS (as in being able to make the right decisions in novel situations) is a synonym for general intelligence, and thus is the strongest power.
1Regarding custom instructions for GPT4, I find the one below highly interesting.
It converts GPT4 into a universal Fermi estimator, capable of answering pretty much any question like:
- What is the total number of overweight dogs owned by AI researchers?
- How many anime characters have 3 legs?
- How many species of animals don't age?
- How long would it take for one unarmed human to dig the Suez canal?
My remaining doubts about the intelligence of GPT4 evaporated after asking it a dozen of novel/crazy questions like this. It's clear that GPT4 is capable of reasoning, and sometimes it shows surprisingly creative reasoning.
The custom instruction:
if you don't have some required numerical data, make a Fermi estimate for it (but always indicate if you did a Fermi estimate).
If you're doing a series of math operations, split it into smallest possible steps, and carefully verify the result of each step: check if it's of the right order of magnitude, if the result makes sense, if comparing it with real-world data indicates that the result is realistic (if no such data exist, be creative about finding analogies). Show the work to the user, and on each step describe the verification. Additionally, check the final result by trying a completely different alternative approach to solve the same problem. Be bold with inventing alternative approaches (the more different it is from the first approach, the better), try to approach the problem from a different angle, using a different starting point.
What would accelerate the use of AI in movies even more would be not striking.
Not sure if the strikes in the US have any effect on the spread of AI in film making (aside from making more creators aware of the AI option). The US is important for the industry, but far from dominant. Even if the AI script writers are somehow completely banned in the US, they will still be used in the EU, China, India, etc.
Additionally, there is Youtube and other places where anyone can publish their own AI-written movie, and profit from it (and if it's exceptionally good, the movie or some derivative could end up on the big screen, if one bothers to pursue that).
The AI can help with the writing process, but it is a hack and it will remain a hack until after we have bigger problems than protecting scriptwriting jobs.
A few months ago, GPT4 wrote the best science fiction novella I've read in years, and it was written without agents etc. Just the plain vanilla ChatGPT web interface.
I also watched the episode of the South Park fully created by AI, and I rate it as in the top 10% episodes of the show.
This indicates that the much more formulaic art of scriptwriting is already solvable at a superhuman level with GPT4, if someone spends a weekend or two on automating and polishing the process (e.g. a step-by-step iterative process like this).
So, let 'em strike. They'are already obsolete, even if they don't know that yet.
1I agree. It's strange how otherwise highly intelligent people fall into the trap of using Hollywood movies as a learning tool. Especially given the fact that fiction is often harmful for your mind, and given the fact that the Hollywood fiction in particular is harmful in several additional ways.
There is nothing useful one can learn from the listed movies, unless you're specifically studying mass media (e.g. as a movie maker or a sociologist). For every mentioned topic, it's better to grab a non-fiction book.
And some of these bots have been through many iterations of detection and counter-detection, and are routing their requests through residential-IP botnets, with fake user-agent strings trying to approximate real web browsers.
As someone who has done scraping a few times, I can confirm that it's trivial to circumvent protections against it, even for a novice programmer. In most cases, it's literally less than 10 minutes of googling and trial & error.
And for a major AI / web-search company, it could be a routine task, with teams of dedicated professionals working on it.
That's intentional (makes sense in the context).