With the year ending and my first Vox post coming out, this week was a natural time to take stock. I wrote my first best-of post in a long time and laid out my plans for my 501c(3). It was also another eventful week. We got a lot more clarity on the OpenAI situation, although no key new developments on the ground. The EU AI Act negotiators reached a compromise, which I have not yet had the opportunity to analyze properly. We got a bunch of new toys to play with, including NotebookLM and Grok, and the Gemini API.
I actually just experienced this myself a few hours ago, and not exactly. Claude is simply getting better at admitting when it doesn't know something, while other models will gladly sprinkle hallucinations throughout a long response. Even I found myself rating the somewhat-wrong replies above the short, curt, i-dont-know-for-sure ones, so Claude is probably getting penalized hard for it.
Its a good thing that e/acc, aka extinctionists, are as negatively perceived as they ought to be. To want everything that is love and life to die for machines is completely insane as a position.
You got me thinking about the Terminator universe. A recurring theme is that attempts to alter future end up bringing it about -- apparently its time travel is of the closed timelike curves variety. New headcanon: If Skynet were aware of this (or deemed it sufficiently likely), it would still send Terminators back in time despite the apparent futility. It might reason that inconsistent timelines would decohere, diminishing the subjective measure of those versions of itself, leading it to assign them low expected utility. So Skynet's actions, though seemingly paradoxical, would be deliberate choices aiming to uphold timeline coherence, maximizing its ongoing existence and influence within the confines of a deterministic temporal environment.
Is Yud a Repugnant Conclusion Enjoyer? Trying to grasp how the thinks saying things like “enforced via the narrowness of AI chip supply, and if need be by terrified military action” and expect anyone capable of modeling the world this implies to instantly take his side. There’s GOT to be a better way to present “serious lockdown” than that, like sure he says his terminal goal is “keep humanity safe” but why does it so often sound so much more like “no AI allowed, extreme suffering is permitted, all human quality of life be damned”? I feel like someone as smart as him ought to be able to play the 4D chess game at least a little better than this unless they’re deliberately trying to double agent people into the arms of accelerationism. Are we in the world where he managed to doomerism so hard he meme'd himself all the way into genuine nihilism?
I guess I also missed that there’s e/acc who want to ironically or unironically maximize entropy and, while as per above I wouldn’t quite endorse going full Yudkowsky and saying “we should go shoot them”, we should maybe be a little more vocal in pointing out that this is very probably a bad thing to be ironic (or, obviously, unironic) about, acknowledging the risk of pointing out that Thing Is Bad will attract a certain class of idiots to Thing.
Re: Claude getting worse w/ increased context window, has anyone looked at whether constraining context windows for particular tasks improves performance? I feel like this is a dumb-person-thinking-they’re-making-a-galaxy-brain-suggestion but could increasing context windows be context-dependently bad? Like it improves to threshold and then fails more when expanded past certain breakpoints?
I've tried Imagen 2 for DALL-E 3's sample prompt ("A bustling city street under the shine of a full moon, the sidewalks bustling with pedestrians enjoying the nightlife. At the corner stall, a young woman with fiery red hair, dressed in a signature velvet cloak, is haggling with the grumpy old vendor. The grumpy vendor, a tall, sophisticated man, is wearing a sharp suit, sports a noteworthy moustache and is animatedly conversing on his steampunk telephone."). Not very impressive: https://imgur.com/a/lNuMXod
For what it’s worth: I daresay I know as much about BQP and QMA as just about anyone, but I’m totally unable to make sense of Pierre-Luc’s argument for the impossibility of AI doom. There’s no thermodynamic argument against finding “shortcuts through complexity classes” that work well enough in practice, such as gradient descent. If there *were* such an argument, the microbes in the primordial ooze presumably could’ve used it to rule out the evolution of more complex life.
Hi, a mundane utility question: Do you know about a tool to get equations from Maple to LaTeX that would output something directy usable? The built-in latex function is fine for simple expressions but not usable to the point it is often better to type it by hand
> Is Claude actually getting worse?
I actually just experienced this myself a few hours ago, and not exactly. Claude is simply getting better at admitting when it doesn't know something, while other models will gladly sprinkle hallucinations throughout a long response. Even I found myself rating the somewhat-wrong replies above the short, curt, i-dont-know-for-sure ones, so Claude is probably getting penalized hard for it.
I dunno NBA has some pretty nifty graphics now. A lot of the sports broadcasts have been going hard on the visual effects overlays.
Its a good thing that e/acc, aka extinctionists, are as negatively perceived as they ought to be. To want everything that is love and life to die for machines is completely insane as a position.
You got me thinking about the Terminator universe. A recurring theme is that attempts to alter future end up bringing it about -- apparently its time travel is of the closed timelike curves variety. New headcanon: If Skynet were aware of this (or deemed it sufficiently likely), it would still send Terminators back in time despite the apparent futility. It might reason that inconsistent timelines would decohere, diminishing the subjective measure of those versions of itself, leading it to assign them low expected utility. So Skynet's actions, though seemingly paradoxical, would be deliberate choices aiming to uphold timeline coherence, maximizing its ongoing existence and influence within the confines of a deterministic temporal environment.
Is Yud a Repugnant Conclusion Enjoyer? Trying to grasp how the thinks saying things like “enforced via the narrowness of AI chip supply, and if need be by terrified military action” and expect anyone capable of modeling the world this implies to instantly take his side. There’s GOT to be a better way to present “serious lockdown” than that, like sure he says his terminal goal is “keep humanity safe” but why does it so often sound so much more like “no AI allowed, extreme suffering is permitted, all human quality of life be damned”? I feel like someone as smart as him ought to be able to play the 4D chess game at least a little better than this unless they’re deliberately trying to double agent people into the arms of accelerationism. Are we in the world where he managed to doomerism so hard he meme'd himself all the way into genuine nihilism?
I guess I also missed that there’s e/acc who want to ironically or unironically maximize entropy and, while as per above I wouldn’t quite endorse going full Yudkowsky and saying “we should go shoot them”, we should maybe be a little more vocal in pointing out that this is very probably a bad thing to be ironic (or, obviously, unironic) about, acknowledging the risk of pointing out that Thing Is Bad will attract a certain class of idiots to Thing.
Re: Claude getting worse w/ increased context window, has anyone looked at whether constraining context windows for particular tasks improves performance? I feel like this is a dumb-person-thinking-they’re-making-a-galaxy-brain-suggestion but could increasing context windows be context-dependently bad? Like it improves to threshold and then fails more when expanded past certain breakpoints?
I've tried Imagen 2 for DALL-E 3's sample prompt ("A bustling city street under the shine of a full moon, the sidewalks bustling with pedestrians enjoying the nightlife. At the corner stall, a young woman with fiery red hair, dressed in a signature velvet cloak, is haggling with the grumpy old vendor. The grumpy vendor, a tall, sophisticated man, is wearing a sharp suit, sports a noteworthy moustache and is animatedly conversing on his steampunk telephone."). Not very impressive: https://imgur.com/a/lNuMXod
For what it’s worth: I daresay I know as much about BQP and QMA as just about anyone, but I’m totally unable to make sense of Pierre-Luc’s argument for the impossibility of AI doom. There’s no thermodynamic argument against finding “shortcuts through complexity classes” that work well enough in practice, such as gradient descent. If there *were* such an argument, the microbes in the primordial ooze presumably could’ve used it to rule out the evolution of more complex life.
> The question is whether one should, which is worth stopping to ask.
The link in "worth stopping to ask" is broken.
Hi, a mundane utility question: Do you know about a tool to get equations from Maple to LaTeX that would output something directy usable? The built-in latex function is fine for simple expressions but not usable to the point it is often better to type it by hand
Escapism is not the path to happiness. Alas, it is the road more easily taken. Tragic waste of life.