LessWrong developer, rationalist since the Overcoming Bias days. Jargon connoisseur.
jimrandomh(Jim Babcock)
Karma: 17,748
This gives the AI a clear opportunity for big advantage. An AI can easily give complex and detailed answers to all six opponents at the same time, for the entire game, in a way a human cannot.
It’s worth noticing that, if a future AGI has taken a treacherous turn, it has that same advantage, but moreso: Not only would it be able to carry on simultaneous complex conversations with six key people at the same time, it would likely be able to scale that all the way to simultaneous conversations with every living human at the same time. Or at least every human that plausibly-matters, if compute is still a bottleneck.
It’s technically doable, but I think probably not worth implementing because the main benefit is mostly not there. With posts, the paragraph and the comment that replies to it may be very far away; but comments are usually shorter, so comments-replying-to-comments are not so far apart.
Huh, there is in fact a bug here, but it’s browser-specific (fails to close in Firefox, closes correctly in Chrome). Will investigate.
You can see the PR here: https://github.com/ForumMagnum/ForumMagnum/pull/6023. Calendar-time from first starting on the first prototype to today was ~6 weeks; accounting for unrelated concurrent projects, I’d say I spent 3-4 FTE equivalent weeks on it. The biggest time sinks were HTML-token-stream wrangling to find matches, mark where to highlight, and so on, and UX details of the hoverable comments themselves. There’s also a partially-implemented system in there for shifting things left (closing the ToC if necessary) to get a wider right margin and avoid scrolling, at certain screen widths, which might get deployed later.
It only shows one icon on purpose, so that if someone goes through a post responding to things point-by-point in a single long comment, and then a reader goes through the post expanding the comment each time, they don’t get flooded with duplicate views of the same comment. If you pin a comment, though, it will highlight where all the blockquotes were, including ones further down in the post, so you can track things somewhat that way.
Sure, this here is an unimportant comment that you can downvote..
There’s a special case for the post author, which exempts them from the karma minimum. If it’s not showing it’s because the visibility is set to “hide all”.
There isn’t currently a way to do it. It’s feasible to add, but requires adding a fair bit of new code to deal with vertical sizing and overlap issues. (Currently it relies on the fact that one comment icon aligned with the top of each paragraph is guaranteed to not create any overlap issues.) I’m not sure that expanded-by-default is good to have by default, though; one of the goals of the current design is to minimize intrusiveness; there’s a risk that, with side-comments visible-by-default on the first pass, people would have trouble making it to the end of a post without getting diverted into sidetracks..
Please give feedback on how this affects the reading experience! If the feedback is positive, this will leave beta and be enabled (for comments above a karma threshold) for everyone.
There’s an option on the new/edit-post page for authors to disable this on their post. (The option is currently only visible if you’ve also opted into beta features in your user settings.) This changes the default from show-10+-karma to hide-all. This may be useful if comments are likely to derail readers from finishing the post, or contain spoilers.
LW Beta Feature: Side-Comments
I’m sure there are many people whose inner experience is like this. But, negative data point: Mine isn’t. Not even a little. And yet, I still believe AGI is likely to wipe out humanity.
Mass-self-copying is one of the mechanisms by which an infrahuman AGI might scale up its compute and turn itself into a superintelligence. Fine-tuning its own weights to produce a more-powerful successor is another strategy it might use. So if you can deny it the ability to do these things, this potentially buys a little time.
This only helps if AGI spends significant time in the infrahuman domain, if one of these strategies would been the first-choice strategy for how to foom (as opposed to social engineering or nanobots or training a successor without using its own weights), and if humanity can do something useful with time that’s bought. So it’s probably a little bit helpful, but in the best case it’s still only a small modifier on a saving throw, not a solution.
Q: Why did the chicken cross the road?
A: We both know that you don’t know of any specific chicken having crossed any specific road. Your question does not state a lie, but presupposes it. This would not be called out as a lie under ordinary social convention, but a deep commitment to truth requires occasionally flagging things like this.
Presuppositions are things which aren’t stated directly, but which are implied by an utterance because if they weren’t true, the utterance would be nonsensical. Presuppositions that aren’t truth-optimized can be surprisingly pernicious.
I have not only told you a joke, I have also told you that chickens live in contexts where they can interact with roads, are motile, and are capable of having motivations. I have told you these things in a way that bypasses some of your epistemic defenses. Some of these things are true, and some are false. I didn’t tell you these things on purpose, but they were communicated nevertheless.
I think that explicitly unpacking presuppositions, and spotting the questionable ones, is a foundational skill of rationality that I’ve never seen expressed. I also suspect some other techniques might be warped versions of this one. For example, there’s a lot of overlap between the ontology of frames, and what you get if you unpack presuppositions, but presupposition-truth is much more concrete.
(Crossposted with Facebook: link)
Aight, I have a Meta Quest Pro to test now, so here are my first-day impression. I haven’t updated away from the bottom line of this post, though a few things were better and a few things were worse than I expected.
They really knocked it out of the park with the pass-through cameras. While wearing the headset under typical indoor lighting, I can see the regular world about as well as I see it without my glasses (I’m nearsighted −2.75).
The underside is open, rather than covered, which makes it much easier to peak out eg to check your phone compared to Quest 2. A side-effect of this is that you can see the same object in both normal and pass-through vision at the same time, and verify that it lines up perfectly.
Text clarity improved more than I expected (they improved the optics, but didn’t significantly increase the panel pixel density). The minimum acceptable font size is now a font size that most people (without my unusual preference for small fonts) would find to be normal.
I haven’t found someone else with a Quest Pro yet to try out eye contact and face tracking in a call, but I was able to try it out on an avatar in a mirror. Eye direction, blinking, winking work, and most mouth movements work. Eyebrows don’t move, sticking out your tongue doesn’t come through, and I don’t think it can distinguish real smiles from fake smiles. Translated onto a slightly cartoonish avatar, this is enough to get me over the uncanny valley.
The weight distribution is good, with the thinner front and the battery moved to the back, but I don’t like the head strap (compared to the Quest 2 Elite Strap I was using). It puts too much of the headset’s weight onto too small an area of the forehead, and as a shear rather than a compression. I think it can be remedied with an aftermarket modification to add a fabric strap across the top, but the strap is much less customizable/replaceable than it was on Quest 2. This seems like a major weak point and some people will probably find it to be a dealbreaker for extended use.
I paired it with a bluetooth mouse and keyboard, opened a web browser and tried it as a laptop replacement, and it Just Worked. (Choice of a virtual environment with hands, keyboard, and browser window visibile, or pass-through camera with a browser window floating in virtualized physical space. Caveat that transferring a keyboard into a virtual environment with computer vision only works for a whitelist of supported keyboard models, and I was lucky in that the first keyboard I tried happened to be a supported one). It’s not at a level where I’d choose it over my Macbook Pro under normal circumstances, but it’s within striking distance, and I *would* use it that way if I was stuck in an airplane seat without room to properly angle a laptop screen, or looking at something highly confidential in a public place.
The controller tracking is solid. It’s the same experience as Quest 2, except the dead zones are gone, as are the bulky tracking rings.
You have officially blown my mind. I seriously cannot believe that AI can subtly mess with video color in real time based on known effect on eye movement, that is absolutely nuts and the applications are limitless in the short- and long-term. No wonder Apple stock keeps going up, there’s probably all sorts of things like that which I’m not aware of.
This isn’t an AI thing at all, it’s an optics thing.
One of the core problems of VR optics is that the panels emit three different wavelengths of light (R, G and B), and these bend differently when they pass through the lenses. If you naively display an image without correcting this, you wind up with red, green, and blue partial images that have separated from each other. In order to fix this problem, you predict the lens effect, apply the opposite effect to the drawn image, and have the distortions cancel. The problem is that the headset’s position on your face is imprecise, and if you shift the headset a millimeter in any direction, the R, G and B images (as perceived by the eye) move in different directions. If you’re trying to display black-on-white or white-on-black text, moving the color channels a pixel apart has a major effect on readability.
Before I read this, I thought I was cool for knowing that oscillating adaptive refresh rate could yield known measurable effects e.g. while shopping online. That’s nothing compared to what chromatic aberration can do. Thank you very much for sharing, my career has benefited profoundly by me learning this.
This paragraph is profoundly confused in a way that I can’t fathom.
foveated rendering (uncertain value, but in the best case might effectively quadruple your GPU speed)
I’m definitely not an expert in this area, but I can’t imagine this being possible unless the headset was hardwired to a data center or something. Have we really gotten to the point where that much ML can fit on a gaming PC?
Again, nothing to do with ML. Foveated rendering is a fancy way of saying “don’t spend GPU cycles drawing parts of the screen that the user isn’t looking at”. It only works if you have an eye-tracking camera that tells you which part of the screen the user is looking at.
I strongly doubt eye-tracked mouse input is going to be good enough for anyone to want (if they don’t have a disability that makes them unable to use their hands). The three reasons for eye tracking are for augmenting the inverse optics (chromatic aberration correction in particular works better if you have millimeter-precise information about where on the user’s face the headset is sitting), for foveated rendering (uncertain value, but in the best case might effectively quadruple your GPU speed), and for letting other people in a conversation see your eye movements (very important for getting past the uncanny valley).
Lots of people look at Oculus and tell stories about how this is supposed to support Facebook’s advertising business in some creepy way or another. I think you can refute this just by looking at revenue numbers: the Oculus platform gets a 30% cut of game revenue (similar to Steam), and this is quite a lot of money per user.
Transformative VR Is Likely Coming Soon
Nuclear winter/nuclear famine being fake is something I’ve suspected for awhile (without specifics, eg I never looked into the size of food commodity stockpiles.) I avoided talking about it partially because it seems infohazardous; “a nuclear war reliably inflicts famine on the victor” seems a lot like a noble lie that someone might spread to reorient militaries away from starting one.
I’m confused about how to relate to this, because I still think any nuclear exchange would be catastrophic for everyone (including the victor, and even assuming that use-nukes-to-shoot-down-other-nukes tactics work as intended). But, even if it’s good for everyone to be deceived about this from a nuclear-risk perspective, I think the collateral epistemic damage is bad enough that we shouldn’t have noble lies in this area; people need to know the truth about how resilient or how brittle the food supply really is, in order to be able to evaluate other risks like climate change or supervolcanoes.
Nitpick: US per-capita calorie consumption is 3782kcal/person/d (source). Some of that is waste and pet food which can be reduced in a disaster scenario, but most of it isn’t. Use of 2000kcal/d as a placeholder number for typical dietary intake is a pet peeve of mine; for most adults, 2000kcal/d is a mild weight-loss diet. A lot of people get tripped up by this when they first start trying to quantify their intake, and falsely assume that the 2000kcal/d number is typical and normative when in fact none of the careful sources are using it that way. (Actual needs vary based on height, activity level, age, weight, and other factors; the way to get a real estimate for the calorie requirements of a specific adult is to use the Harris-Benedict equation to get a BMR and multiply it by an activity factor.)
Nope, https://www.lesswrong.com/posts/CG9AEXwSjdrXPBEZ9/welcome-to-less-wrong has 2003 and https://www.lesswrong.com/posts/yCWPkLi8wJvewPbEp/the-noncentral-fallacy-the-worst-argument-in-the-world has 1758.