Archive for category Speculation

Using AI to moderate online discussions

This brief post is here solely as prior art to make it more difficult for someone to patent these ideas. Probably I’m too late (the idea is so obvious the patent office probably has a dozen applications already), but I’m trying.

GOALS

One of my pet projects for years has been finding a way to promote civil discussion online. As everyone knows, most online discussion takes place in virtual cesspits – Facebook, Twitter, the comments sections of most news articles, etc. Social media and the ideological bubbles it promotes have been blamed for political polarization and ennui of young people around the world. I won’t elaborate on this – others have done that better than I can.

The problem goes back at least to the days of Usenet – even then I was interested in crowdsourced voting systems where “good” posts would get upvoted and “bad” ones downvoted in various ways, together with collaborative filtering on a per-reader basis to show readers the posts they’ll value most. I suppose many versions of this must have been tried by now; certainly sites like Stack Exchange and Reddit have made real efforts. The problem persists, so these solutions are at best incomplete. And of course some sites have excellent quality comments (I’m thinking of https://astralcodexten.substack.com/ and https://www.overcomingbias.com/), but these either have extremely narrow audiences or the hosts spend vast effort on manual moderation.

My goal (you may not share it) is to enable online discussion that’s civil and rational. Discussion that consists of facts and reasoned arguments, not epithets and insults. Discussion that respects the Principle of Charity. Discussion where people try to seek truth and attempt to persuade rather than bludgeon those who disagree. Discussion where facts matter. I think such discussions are more fun for the participants (they are for me), more informative to readers, and lead to enlightenment and discovery.

SHORT VERSION

Here’s the short version: When a commenter (let’s say on a news article, editorial, or blog post) drafts a post, the post content is reviewed by an AI (a LLM such as a GPT, as are currently all the rage) for conformity with “community values”. These values are set by the host of the discussion – the publication, website, etc. The host describes the values to the AI, in plain English, in a prompt to the AI. My model is that the “community values” reflect the kind of conversations the host wants to see on their platform – polite, respectful, rational, fact-driven, etc. Or not, as the case may be. My model doesn’t involve “values” that shut down rational discussion or genuine disagreement (“poster must claim Earth is flat”, “poster must support Republican values”…), altho I suppose some people may want to try that.

The commenter drafts a post in the currently-usual way, and clicks the “post” button. At that point the AI reviews the text of the comment (possibly along with the conversation so far, for context) and decides whether the comment meets the community values for the site. If so, the comment is posted.

If not, the AI explains to the poster what was wrong with the comment – it was insulting, it was illogical, it was…whatever. And perhaps offers a restatement or alternative wording. The poster may then modify their comment and try again. Perhaps they can also argue with the AI to try to convince it to change its opinion.

IMPORTANT ELABORATIONS

The above is the shortest and simplest version of the concept.

One reasonable objection is that this is, effectively, a censorship mechanism. As described, it is, but limited a single host site. I don’t have a problem with that, since the Internet is full of discussions and people are free to leave sites they find too constraining.

Still, there are many ways to modify the system to remove or loosen the censorship aspect, and perhaps those will work better. Below are a couple I’ve thought of.

OVERRIDE SYSTEMS

If the AI says a post doesn’t meet local standard, the poster can override the AI and post the comment anyway.

Such overrides would be allowed only if the poster has sufficient “override points”, which are consumed each time a poster overrides the AI (perhaps a fixed number per post, or perhaps variable based on the how far out of spec the AI deems to the post); once they’re out of points they can’t override anymore.

Override points might be acquired:

  • so many per unit time (each user gets some fixed allocation weekly), or
  • by posting things approved of by the AI or by readers, or
  • by seniority on the site, or
  • by reputation (earned somehow), or
  • by gift of the host (presumably to trusted people), or
  • by buying them with money, or
  • some combination of these.

Re buying them with money, a poster could effectively bet the AI about the outcome of human moderator review. Comments posted this way go online and also to a human moderator, who independently decides if the AI was right. If so, the site keeps the money. If the moderator sides with poster, the points (or money) is returned.

The expenditure of override points is also valuable feedback to the site host who drafts the “community values” prompt – the host can see which posts required how many override points (and why, according to the AI), and decide whether to modify the prompt.

READER-SIDE MODERATION

Another idea (credit here to Richard E.) is that all comments are posted, just with different ratings, and readers see whatever they’ve asked to see based on the ratings (and perhaps other criteria).

The AI rates the comment on multiple independent scales – for example, politeness, logic, rationality, fact content, charity, etc., each scale defined in an AI prompt by the host. The host offers a default set of thresholds or preferences for what readers see but readers are free to change those as they see fit.

(Letting readers define their own scales is possible but computationally expensive – each comment would need to be rated by the AI for each reader, rather than just once when posted).

In this model there could also be a points system that allows posters to modify their ratings, if they want to promote something the AI (or readers) would prefer not to see.

My God, it’s full of… disk drives!

What’s going on with the X-37B?

X-37B (USAF photo)

read today that the USAF has launched one again, on, I think, the 3rd flight.

What is unique about this vehicle compared to most other space systems is:

  • It’s reusable and returns to Earth (lands on a runway, ala the Space Shuttle)
  • It has a long on-orbit dwell time (270 days was supposedly the spec, with the second flight lasting 469 days)
  • Some significant amount of on-orbit delta-v for orbital maneuvering
  • Payload capacity “similar to a pick-up truck”

What is this thing for? The government has been strangely silent about its purpose, leading to a lot of random speculation, none of which makes any sense to me. The only thing they’ve said is:

“The X-37B is a risk reduction vehicle for space experimentation and to explore concepts of operation for a long duration, reusable space vehicle.”

I’m sure that’s true to some extent, but I’m sure that’s not all that is going on; otherwise why all the hush-hush?

Some of the theories I’ve seen claim that it’s:

  1. For spying on the Chinese manned space program
  2. For spying on random spacecraft in orbit
  3. Some sort of on-orbit anti-satellite weapon
  4. For repairing satellites in orbit
  5. A ground-attack (or ICBM interception) weapon (rods-from-god or similar)
  6. An orbital bomber
  7. For stealing satellites, like in You Only Live Twice
  8. Some professor at the Naval War College said “the Air Force has always wanted a crewed space plane, and this was the closest they could get”.

None of those make sense. One at a time:

1 and 2 – Spying on other spacecraft

The DoD is already pretty good at pointing ground telescopes at spacecraft in orbit. Remember they offered to check out Space Shuttle Columbia’s missing tiles back in 2003? Matching orbits and a close approach to an opponent’s spacecraft would be tremendously provocative (it’s not as if you can do it secretly), and how much more can you really learn by looking at a spacecraft up-close vs. from a distance anyway? Not enough to justify this doubtless hugely expensive program, I’m sure. If despite that, you really want to do this, you don’t need a vehicle that can land. As for spying on the Chinese manned space program, why would that be of any interest at all? (Besides, it’s been in the wrong orbit for that.)

So I don’t buy those explanations.

3 – It’s an ASAT weapon

So why does an ASAT weapon need to land? No matter how much the thing costs, it’s got to be cheaper to just build and launch a new one every so often than land it, refurbish it, and re-fly it.

4 – Repairing satellites in orbit

That’s crazy. Maybe if it were manned; if that’s what you want to do it’d be way cheaper to fly repair techs on a SpaceX Dragon. And what possible purpose would there be in having the repair-bot sit on orbit for a year, or have the capability to land?

5 – Rods from God

Again, no need for such a thing to land. If you really want to build that, sure build it, but why complicate things by having it land? Makes no sense.

6 – Orbital bomber

First, that would violate the Outer Space Treaty. I can see people wanting to get out of that treaty, but I think the US would do so explicitly rather than in this not-very-sneaky way. But, again, why does it need to land? Maybe you can make a case that the USAF doesn’t want to leave nukes on orbit forever – they want a way to get them back eventually. But there are many simpler ways to accomplish that – it can de-orbit the warhead with a parachute (ala Corona, as well as the whole Apollo program), they could plan on a future OTV or manned spacecraft to collect warheads in 15 years, etc. Plus, who needs another way to deliver nukes anyway? The Soviets are not coming back, and the Chinese only want to sell us stuff.

7 – Pac-Man (waka waka)

Snatching someone else’s satellite out of orbit would be an act of war. And difficult, because you don’t know where it’s CG is or how much damage it’ll take by being bounced around on re-entry. And why do you need your snatch machine to sit on orbit for a year? You can always just launch it when you need it.

8 – Manned spacecraft wannabe

Except they already built the full-size version (that flying pink elephant, the Space Shuttle), and this one doesn’t carry anybody. And a Dragon will be way cheaper, and NASA is already funding that. Not to mention Orion (oops… too late, I mentioned it. It’ll probably never fly anyway.) And, of course, they’ll have trouble getting even one brave soul to sit in a space the size of a pickup truck bed for a year (let alone finding room for oxygen, water, and food for that duration).

So what’s it for, then?

OK, here’s my theory. It’s a spy satellite full of disk drives.

The payload bay has some sensors, probably high-resolution cameras; maybe other things too. The rest of the payload is disk drives.

The X-37B collects data (pictures, maybe sigint, maybe other things), and stores it on the disk drives. Every so often it changes orbits in order to be able to look at some particular thing at particular times (or just to keep the other guy guessing). Once the disk drives are full (a year or so), or sooner if the data is needed on the ground sooner, it lands.

It lands so that the data on the disk drives can be read off.

Why not just radio the data down? Because there is way, way, too much of it.

Suppose the payload bay is 4 x 8 x 3 meters (roughly a full-size pickup truck bed). That’s 96 cubic meters (96 million cc). A 3.5″ hard drive is 101.6 x 25.4 x 146 millimeters, that’s 377 cc. So there’s enough room for about 250,000 drives. Figure a tenth of that after allowing room for the sensors, power supplies, and cooling (cooling is a big deal in space). 25,000 drives at 3 TBytes each is 73 petabytes. (BTW, that’s about $2.5M worth of disk drives; peanuts for the DoD.)

73 petabytes over a year is 25 gigabits/second. That’s 24×7, including when not over a convenient ground station.

How does that compare with the bandwidth available for a satellite downlink? I don’t know exactly, but the whole X-band is only 500 MHz, as is the Ku band. The Ka band is 2.5 GHz. That’s the whole band. You do the numbers.

[Edit, March 2013: There’s a mistake in the numbers above – a pickup bed is about 4 x 8 x 3 feet, not meters. But on further thought the X-37B is almost certainly using SSDs instead of rotating media, and the density of that is a lot higher. So I think the two mistakes roughly cancel out, without changing the conclusion.]

What can you do with that?

The Earth’s surface area is 510 million square kilometers (about a third of that is land). Let’s assume 24 bits/pixel (you can divide that up into bands as you like). Unless I made a mistake in the math, 73 petabytes is enough for pixels 40 centimeters on a side, of the whole planet.

Or, somewhat bigger pixels, but multiple copies in the disk drives, so hardware on-board the spacecraft can compare old images against new ones, and identify differences. You get the idea.

Makes Google Earth look kinda…lame.

[Disclaimer: This is just a guess; I have no inside info. Just imagination and a calculator.]