The biggest news this week was at long last the announcement of Google’s Gemini. Be sure to check that out. Note that what is being rolled out now is only Gemini Pro, the Gemini Ultra model that could rival GPT-4 is not yet available. It does not seem I am doing a good job cutting down on included material fast enough to keep pace. A lot is happening, but a lot will likely be happening for a long time. If your time is limited, remember to focus on the sections relevant to your interests.
Hey Zvi, do you plan on doing a recap of the year (in AI)? I think it would be pretty interesting to see your perspective on how things evolved (especially considering you've been making these round-ups for a while now...).
> I would think of this less as ‘catching them using ChatGPT’ and more ‘catching them submitting a badly written assignment.’
There's more to the post than that. Further in the tweet thread, Thinkwert says he that GPT's bad writing looks genuinely different. When a student has an empty argument, they are usually halting and confused and contradictory, but ChatGPT is smooth and confident but disconnected from the core discussion. Some humans can do that, too, of course, so it's not absolute proof, but he says that it's enough evidence that he can often use it to coax a confession when the student is guilty.
There's also discussion about how to improve your system prompts to coax ChatGPT to write better. IMHO, if you can use ChatGPT intelligently to help you write a better paper than you could write on your own, that's fine. The important is to make sure that students learn about good writing rather than just to prove that they didn't "cheat".
So that generate-video-from-image sample is clearly struggling in a few key areas—it’s obvious anywhere the fabric has to move substantially, the proportions are a bit janky, and the faces are definitely obvious—but I wonder how much of that is the wireframe movement template not being detailed enough. It’s been an internet eon now but we already had a wave of memes turning still portraits into singers and, as you said, this kind of thing is one TikTok dance away from rocketing into the stratosphere. The animated/anime ones are especially bad and would need so much work to clean up as to make it not worth it (at present) but I strongly suspect the human portraits in the hands of an experienced rigger/animator could do some wild stuff without anyone noticing.
Basically, I suspect legal challenges to AI use in entertainment will hit porn before anything else.
> The [Belrose and Pope] post still puts existential risk from AI, despite all this, at ~1%. Which I will note that I do agree would be an acceptable risk, given our alternatives, if that was accurate.
Have you given much thought to your rough upper limit for acceptable risk? Belrose and Pope don't seem to mention a time-frame. I find it hard to think about risk without one, so I'm going to assume we're discussing this century.
Has anyone else in the upper left quadrant (from section 16) given their own rough upper limit?
"Totally agree with all this analysis, and yet, if media is and previously wasn’t fully controlled by people committed to preventing gains to humanity, that has some bearing on whether AGI can be expected any time soonish."
(2) You probably meant "strangle", not "strange" in this line (though it's funnier as written):
"Hawley’s upcoming bill about Section 230 is a no good, very bad bill that will not only strange generative AI in its tracks[...]"
"Surely if the user typed ‘Tony Danza hates puppies’ then that would not allow a third party to sue ChatGPT in the absence of Section 230, that’s obvious nonsense."
You can sue anyone for anything (any many frequently do). The special magic of S230 is that it creates a procedural process to get a certain class of cases dismissed early. Nothing in S230 changes the definition of libel, it just states explicitly who can be legitimately held accountable for libel.
Otherwise defending (even if you win) such lawsuits can be ruinously expensive.
If we are indeed in a simulation of the most interesting historical epoch, the run-up to ASI, could it be pivotal that the AI regulatory effort by the United States is led by such manifestly dim and incompetent people? This seems like it could be a key variable, but I can't tell which way it might cut in terms of p/doom.
Hey Zvi, do you plan on doing a recap of the year (in AI)? I think it would be pretty interesting to see your perspective on how things evolved (especially considering you've been making these round-ups for a while now...).
> I would think of this less as ‘catching them using ChatGPT’ and more ‘catching them submitting a badly written assignment.’
There's more to the post than that. Further in the tweet thread, Thinkwert says he that GPT's bad writing looks genuinely different. When a student has an empty argument, they are usually halting and confused and contradictory, but ChatGPT is smooth and confident but disconnected from the core discussion. Some humans can do that, too, of course, so it's not absolute proof, but he says that it's enough evidence that he can often use it to coax a confession when the student is guilty.
There's also discussion about how to improve your system prompts to coax ChatGPT to write better. IMHO, if you can use ChatGPT intelligently to help you write a better paper than you could write on your own, that's fine. The important is to make sure that students learn about good writing rather than just to prove that they didn't "cheat".
So that generate-video-from-image sample is clearly struggling in a few key areas—it’s obvious anywhere the fabric has to move substantially, the proportions are a bit janky, and the faces are definitely obvious—but I wonder how much of that is the wireframe movement template not being detailed enough. It’s been an internet eon now but we already had a wave of memes turning still portraits into singers and, as you said, this kind of thing is one TikTok dance away from rocketing into the stratosphere. The animated/anime ones are especially bad and would need so much work to clean up as to make it not worth it (at present) but I strongly suspect the human portraits in the hands of an experienced rigger/animator could do some wild stuff without anyone noticing.
Basically, I suspect legal challenges to AI use in entertainment will hit porn before anything else.
> The [Belrose and Pope] post still puts existential risk from AI, despite all this, at ~1%. Which I will note that I do agree would be an acceptable risk, given our alternatives, if that was accurate.
Have you given much thought to your rough upper limit for acceptable risk? Belrose and Pope don't seem to mention a time-frame. I find it hard to think about risk without one, so I'm going to assume we're discussing this century.
Has anyone else in the upper left quadrant (from section 16) given their own rough upper limit?
Minor nitpick:
(1) This sentence appears TWICE:
"Totally agree with all this analysis, and yet, if media is and previously wasn’t fully controlled by people committed to preventing gains to humanity, that has some bearing on whether AGI can be expected any time soonish."
(2) You probably meant "strangle", not "strange" in this line (though it's funnier as written):
"Hawley’s upcoming bill about Section 230 is a no good, very bad bill that will not only strange generative AI in its tracks[...]"
See you at Solstice!
Am i the only one who would definitely pay to read Zvi on Moloch? (Surely not.)
"Surely if the user typed ‘Tony Danza hates puppies’ then that would not allow a third party to sue ChatGPT in the absence of Section 230, that’s obvious nonsense."
You can sue anyone for anything (any many frequently do). The special magic of S230 is that it creates a procedural process to get a certain class of cases dismissed early. Nothing in S230 changes the definition of libel, it just states explicitly who can be legitimately held accountable for libel.
Otherwise defending (even if you win) such lawsuits can be ruinously expensive.
The New York Solstice or the related meetup? Link?
If we are indeed in a simulation of the most interesting historical epoch, the run-up to ASI, could it be pivotal that the AI regulatory effort by the United States is led by such manifestly dim and incompetent people? This seems like it could be a key variable, but I can't tell which way it might cut in terms of p/doom.
As it relates to "AI control" of untrustworthy models, you might be interested in Redwood's new work on this here: https://twitter.com/bshlgrs/status/1734967328599945621