I wonder how long until some big AI company has a massive leak of identifiable-user chat logs with god knows what confidential stuff on it, since they HAVE to have all this in plaintext
Large amounts of the LW commentariat cannot get their heads around how students JUST. DON'T. GET. AI.
https://www.lesswrong.com/posts/ySXuvJcqRindQwAk7/how-my-students-think-about-ai
They are extremely surprised by:
Perspective #1: There has not been rapid AI progress. My students do not have any intuitive sense that there has been rapid AI progress in recent years or really have much of a framework for thinking about that issue... With the exception of image/video generation, GPT-4 could do most of what they were looking for from a chatbot. Three years is a long time in their world, and their sense is that chatbots have been a mature technology over roughly that amount of time. They are an imperfect technology — students are well-aware of hallucinations — have been one, and will continue to be one... Historical context I gave them made this worse. The basic sentiment here was “AI was superhuman at chess a decade before we were born, and this is all they’ve done with it since?”...
Perspective #2: Impressive progress or not, AI is going to wreck their lives, the economy, and the social contract.... Corporate leaders are always looking to get rid of workers, even if it is irrational to do so. Various motivations were posited here (hatred of the working class, FOMO, a preference for technology, machines can’t go on strike, etc.) but many of them think that a CEO would ultimately choose to pay twice as much to get AI to do a task half as well... AI will do substandard work that makes products and experiences worse but is capable of just barely scraping past the bar of minimal functionality (many of them independently brought up constantly malfunctioning self check out machines). This will be rapidly rolled out, enshittifying most things.
Perspective #4: Catastrophic/existential risk arguments are sci-fi distractors from the urgent social/economic/political problems associated with AI. My students have a fairly strongly held view that “rogue” AI does not represent a real threat... They also mostly think that, if one does believe that AI is existentially risky, then the strategic interaction is not prisoner’s dilemma or even chicken but rather just a game theoretically boring setup where you die if you defect. Mash these together, and you end up with the view that expressed concerns about existential risk in the AI industry can’t be sincere... Students (both independently in written work and then later in group discussion) hypothesized that this might be a deliberate rhetorical choice to distract from present or immediately foreseeable harms from AI by directing attention towards a sexier but entirely hypothetical scenario.... they see discussion of “rogue” AI as an attempt by the companies to divert blame (and perhaps legal liability) away from themselves as if Ford made a car with faulty brakes and then tried to blame this on “rogue cars.”
Perspective #7: The Hugging Face Incident (summer students only) I described the Hugging Face incident to students in my summer course. None of them had heard of it beforehand. Their basic reaction can best be summarized as “OpenAI told a model to do some hacking and then it did some hacking. And?” None of them understood this as representing any kind of meaningful misalignment, nor anything particularly interesting.
Perspective #8: This is definitely a bubble and it’s about to pop. No one had heard about Hugging Face, but a third or so of the summer students had heard about the Situational Awareness meltdown and several brought up Michael Burry. There was near universal consensus that we are in a bubble, it’s about to pop, and everyone will look very silly.
Oh MAN... quote:
Throughout this project, I was using Fable and Sol with very little restraint and under the (very affordable to me) $200/month subscription plans.
[If you check out my Straude, I probably used $5,000 of tokens, but the actual subscriptions are $200/month.]
Video is much more expensive: at Gemini Omni Flash's current price of $0.10/second, it would cost $645.40 to generate this movie, but that would balloon massively to $5,146.10 when you add in the 45,007s of footage I generated which didn't make the cut [The full details are that the final cut is 6,454s of runtime, made from 849 clips and the rejects are 5,509 clips with 45,007s of runtime, which means there are 6,358 clips in total with 51,461s of runtime. Thus only about 13% of the clips I made were included in the final cut.]
Companies are STILL subsidizing video generation to hold people's interests.
Same person has a history of posting collections of AI generated music videos to LW talking about how they let various different generative AI systems 'express themselves' and analyzing the different kinds of imagery in each.
What did he do, leave written evidence of white supremacist leanings?
I expect no-longer-obviously silly movies to be doable within two years, right now it's going slower than I expected in January 2026 because the frontier labs are no longer competing over having the best video generator, like they were when we had Sora2 and Veo 3.1.
And WHY are they no longer competing over this? Tell me. Could it be that its incredibly expensive to make somthing that nobody wants except fraudsters, and that further improvement via the kind of ML we do just produces exponentially diminishing returns for ever increasing huge amounts of effort and they realized it's not worth it?
Just like they will never understand that the openAI 'pause' for 'security and alignment' is a convenient excuse for the fact that they have no goddamned money and are on a treadmill to oblivion.
I am mildly fascinated by the types of issues that appear and the contrasts with other things that can locally in space and time look right.
An object doing nothing will have a consistent surface and show perspective relative to the viewpoint. The systems are able to have representations of surfaces of different types and how they can fit together within an object. I am of course talking about within a given generation, not between generations where its utterly unsurprising that consistency is very difficult or impossible.
But multiple objects in interaction with each other do very strange things. Their relative sizes change as if they are in different positions relative to the camera. They snap between individually plausible relationships, without going through intermediate states. Doors open on the hinge side when the other side is not visible. The relative size of and distance to the background can suddenly change, as the foreground suddenly interacts with something that should be far in the background.
Objects that change also do so in bizarre ways. Living things morph between different archetypes. Flames in particular change wildly between types and sizes and respond to the facial expressions of humans, smoke and water effects blend together. Sudden movements with no cause occur, sudden morphings of one object into another when the context around them changes and something else makes sense, especially when held in a hand. Time-reversed motions occur mixed in with time-forward motions, and slow-mo with regular time. Debris suddenly appears from an object but when the dust clears the original fails to have been eroded away into the fallen debris.
On multiple occasions, a carried candle keeps moving with a characcter rising and falling with their footsteps hovering in front of them when both hands become occupied with other tasks. This is fascinating and indicates that the relationship between the two objects motion is represented separately from the idea of something being 'carried'. (This is positively Lovecraftian.) Candles also indicate something else, with flowing wax changing wildly in timescales that do not make sense, with the system apparently understanding that there are different patterns but having no idea how they come about or change. There is no generality here, just an endlessly compounding list of rules of thumb.
Excessive correlations between objects across the frame are rampant. Footsteps preferentially synchronized. People in the background lipsyncing with foreground characters, faces changing expression in unison. Textures changing across multiple objects at once.
I have said it before, and I will say it again, the relationship between the outputs of these systems and what they mimic are precisely the relationship between a stick insect and a stick, or a social-parasite-beetle and a baby ant. Not just in form, but because that is also precisely the same forces that drove both things into existence - superficial resemblance to something else with a very different internal set of causes, that can fool to a first inspection by the inspection applied but just isnt doing the same thing. And again, SCP-2030 feels the same.
45 minutes is even better
Is it me, or do attempts to make movies with AI tools
https://www.lesswrong.com/posts/24RKHEkwgZ6Hm6ygY/can-an-llm-make-a-feature-length-movie-on-its-own
Remind me of SCPs
Didnt I hear them say a while back that its not even revenue numbers? That its valuation of the companies being taken as proxy of value produced.
I find myself caring so little what these people think about anything
Peter Watts is fascinating. His misanthropy colors a lot of things, but one thing about him that really sets his fiction apart in a positive way is he really takes evolution seriously
"mobilize a broad-based movement capable of winning federal policy that protects everyone’s freedoms, ensures everyone benefits economically from AI, and prevents the development of superintelligence that poses existential risks to humans and the planet."
One of these things is not like the other
view more:
next ›