Have the Sense that God Gave to a Horse

And other things my Grandfather told me

At the risk of creating yet another narrative causality loop, I feel like Elon Musk has taken a nap at some point in the last two weeks and the effect of this has been that Twitter overall has become slightly more stable as a business venture. Twitter business accounts being able to give new swag in place of the old blue check is something of a peace offering in my view. Maybe you won’t be a blue check CNN reporter but you can be a CNN square reporter. This feels like a near perfect win/win for the company and its user base.

I think pushing this same kind of Boy Scout badge reward system —everyone loves a merit badge— could hook a lot of people and provide valuable filtering data and even more literally valuable revenue. It also gets Twitter out of the nitty gritty of verification. You verify CNN, or whoever, then let them deal with their verifications after that point. Maybe even charge them for each badge they give so it will feel more exclusive to the employees. This was a very sensible step, which from the All-In Podcast, I understand David Sachs helped to elevate when it was brought to him by a few Twitter Product Mangers. It says something about the company management that such a simple and obviously profitable idea wasn’t able to make it to production under the old regime and was out within a month under the new management. This was a company that had five times the workforce and hadn’t done anything really visible in years. Also view count… how was that not a thing before?

I still want to understand how their test environments work because it doesn’t make sense to me that they don’t have any kind of test environment while still releasing new features, but they seem to not have crashed the whole code base so it seems to be working. Maybe some kind of production parallel A/B test? That’s my only guess. Still, everything I hear leads me to wonder what the company was doing for ten years prior to the take over.

I haven’t been able to find good data on how the debt used to purchase Twitter is structured but if it’s variable with the prime rate that’s my biggest concern for the company at this point. For those who are not finance dorks, the Federal Reserve lends money to banks and then banks add some interest on top of that to obtain their profit. Debt can either be fixed at the time it is extended, like most mortgages, so that the rate and therefore your payments stay stable even as the prime rate is adjusted. Or debts can be variable for some x% plus the prime rate, in which case you can get into some real pain if someone in the FED is trying to crush inflation.

Twitter has a revenue stream of a couple of billion dollars. It now has debt servicing obligations under the buy out of about a billion and a half dollars —if I’m remembering right— which means that cutting operating expenses and growing revenue is the name of the game. Especially with a recession looming and companies looking to slash advertising budgets, as that is always the first thing to go. If the FED rate goes up significantly however, and the debt is variable, then you end up with a bunch more operating costs without any upside.

Considerations like this are why people like George Hotz get under my skin. They get fixated on one very specific problem. They come up with an optimal solution to fix that problem. They are the best person, maybe even in the world, to solve that problem, except… their solution could never work in the real world because most problems are AND problems even if they don’t present that way immediately. An AND problem means you’ve got lots of things going on at once, like you have to solve for performance AND budget AND legal compliance AND some other random bullshit thing you don’t even know about until you’re six months in which you never have the right staffing to accommodate. So you’ll get a group of people together trying to lay out all the “AND’s” in the problem and then run through possible solutions and how to tweak them only for someone like George Hotz to get on the call and say something like “Hey, everyone shut up and listen to me, in a perfect vacuum all objects on the surface of the Earth fall at the same speed!” To which you have to say, who cares? In fairness sometimes this matters a lot, because there will be an AND out there that nobody could solve before and it’s worth doing new solutions for all the other ANDs just to make the new AND solution work, but this generally happens a lot less than people would think. So, I think the guy is very smart, but probably thinks the things you learn as a child on Sesame Street about being nice to other people are just hollow lies we tell to children to save them from the cold Darwinian horror of the real world in a way that only George Hotz can truly fathom, instead of practical time tested strategies to get large groups of people to cooperate.

I once led a team of a hundred some odd people. You know how I did it? Donuts, anecdotes, and compliments.

Elon is calling out all the lower-hanging features I want in order to make Twitter operate more transparently as near term deliverables. The jury adjudication and news ranking bit have not been mentioned but public accountability tools for moderation are a top priority. That alone will make what he has a lot more fair. I also want a public law and it seems like they are moving in that direction.

When I eventually sign up for a Twitter account I will see if I can make a video explaining the benefits of an adjudication layer and then go bug a Twitter PM in their replies. I can write with both hands at the same time, and although most people think I’m slightly insane or mentally handicapped at a first meeting —which, you know, fair enough— usually when people see me doing that they start to listen. I mean, holy shit, how often do you see a guy writing with both hands at the same time?

I also have to be careful when I write with both hands at the same time because it causes people not to push back on me hard enough when I propose something. Such is the power of simultaneous both hands writing. Learn this one simple trick and everyone will assume your IQ is about three-hundred.

Also if you have a wool mitten to put on your dominant hand, you can literally teach yourself to do this in about two weeks. Not as hard as it looks.

BENEFITS OF ADJUDICATION TO MAKE MONEY

One of the big hurdles I heard on an Elon call most recently is there are not enough video ads and/or interesting ads to actually present to customers. There’s just a small inventory of ads. That’s why a lot of the ads suck, even on YouTube where ads are central to the whole business model you come across a lot of ads for what appear to be outright scams.

Allow me to fix that problem AND also save democracy at the same time AND also increase user engagement AND stand-up creator monetization AND solve the looming ad revenue crisis when the recession hits.

Say you’re DiGiorno pizza. You’ve got one of the most well known taglines in all of frozen pizza. “It’s not delivery, it’s DiGiorno!” You have to spend a lot of money on advertisement to make your product distinct from evil frozen pizza which is literally named after a godless German fighter ace that killed a lot of good men during WWI, which who the fuck even signed off on that name? Still, it’s a crowded market trying to cram frozen pizza down the throat of Americans and you have to show up to the field of battle.

This carries a decent amount of risk because you don’t really know how the ad is going to go over until you actually release it to the public. You also have to pay some kind of upfront cost to have an ad agency create the ad and then additional money to produce the advertisement. All of this is before you see any upside.

So what if… just random people made ads? You don’t pay them anything at all for making them. And you only pay them for the ad if it does well and you could set a cap for how much you were going to spend and you could also be really sure that those ads weren’t going to do anything to embarrass your beloved frozen pizza brand. You could even do things like stipulate that all the ads have to be funny or informative or whatever.

For instance, I would like to make an ad where I go out into the middle of the woods with a bow and arrow. I’ll wear my black canvas hat that fits really well but that my wife hates. I’ll wear my canvas jacket that fits really well that my wife also really hates. In a voice over, I’ll start talking about the Earth and how important it is to be connected to it and to remember where we came from. It’s important to work to get your food. Glamorous shots of forests and mountains and rivers.

I pull back my bow, totally focused. I talk about the thrill of earning what you eat. I release the arrow.

I shoot a DiGiorno pizza right through the center.

Still talking about how it’s important to get food for yourself and do it on your own, I’ll take out a hunting knife and cut open the box like I’m gutting an animal. I’ll go home and string it up in my garage like it’s a deer that I killed. Then my wife will take it inside and cook it. Finally, she will serve it to the whole family.

Cut to black. “It’s not delivery, it’s DiGiorno.”

So I put this whole ad together and a whole bunch of other people who have heard you can make money creating advertisements on Twitter put their ads together and then what? How do you sort through that ad pile so that the best ads win and these companies can ship product?

The people who make the ads submit a small fee. Other verified Twitter users look at the ad, sign off on whether or not it contains any objectionable content that the company has laid out, and also give it a rating for how good it is overall or for some other metric the company wants. The company can set their own criteria for how many people they want to score it in this stage and if they want to maximize funniness, information, etc. If I were them, I’d want to make all my ads funny because I’ll click on any ad for anything if it gives me a good laugh.

Once my ad is cleared for display it gets pushed out to people. If I scored more highly in the test round I get pushed out more frequently but the others still have a shot if they’re driving more click through. Once the ads are live it’s all dynamic and the ones performing best get up-regulated. Pranks and low effort ads are disincentivized since it costs money to submit. The neural net that will be trained to push these will track the click to actions instead of impressions —which was a really dumb idea, as Elon said in a Twitter Spaces talk I heard— plus also maybe how long someone hovers over it if that’s possible, which it might not be. If your ad does really well, based on some agreed upon standard, you get paid some amount per call to action or whatever.

Repeat for any company that wants to participate in the process.

Twitter makes money on the adjudication and on the ad sales, which it splits with the ad creator. Twitter doesn’t have to put forth any money to create the ad and neither does the company. It’s all user generated content. Then you get a high volume of ads, maybe some from those girls who are famous for having big asses —in the future, I think there will be a last name like Graham but it will be pronounced and spelled “Gramham” for the descendants of those who made a living having a big ass— can sell energy drinks or whatever. This can all happen quickly, dynamically, and with almost no effort from the companies involved.

As a good proof of concept, I would try this out for Tesla, Starlink, and SpaceX and maximize for humor. I would also try to run a lot of ads for Twitter itself, especially for people who are browsing but don’t have an account. I would set a basic rule for no politics, no religion, etc. Basically nothing in your ad can touch on topics that piss people off at Thanksgiving.

Okay, now you’ve built this system —which you could probably build by just modifying birdwatch, which is my preferred approach when you have to jettison a bunch of stuff. Evolve the stuff you have to do other things and then try to eliminate the need for other crap— but now you can also use it for anything else you need to assemble a jury for, like rules violation and appeals. Same process. Feed in defenses and evidence of the accusation. Feed in money to pay the participants. Out comes profit for the company and a decision it is not directly responsible for. Indirectly responsible for, yes, but you’re no longer pissing off half of your user base every time something comes up that is a Scissor statement (shout out to Scott Alexander).

Part of the reason I want this to be part of its own separate system is I want to build it in such a way that another social media company, like a Reddit, or a YouTube, or a Facebook, could consume the adjudication as a service. Then Twitter makes money for providing them with ads to sell and moderating their content. I’d really try to make Twitter adjudication into a meta layer that lives on top of the internet that people come to instead of a thing that lives in its own sandbox.

Civilization is a rule-set that amplifies order and down-regulates and/or channels chaos productively. Give people something positive to do on Twitter. Give people a way to filter out noise. Do all of this in a way that keeps people steering the ship of discourse.

It’s not going to be that long before ChatGPT and its descendants become nigh indistinguishable from human voices and we need to do what we can in the interim to make sure there’s a reservoir of anthropic coherent extrapolated volition that can be drawn upon to help guide our future AI overlords.

CHAT GPT and the Sense God Gave a Horse

I haven’t seen anyone else do this yet, and to be fair I haven’t done this as extensively as I want to yet because I have a one year old, but it’s not that hard to integrate Chat GPT with Google or I guess Bing since Microsoft is providing financing for Chat GPT. This is computationally expensive so I’m not saying this is a good idea at scale but here’s basically what I have tried to do manually to see if would work.

For those who don’t know Chat GPT is a scary powerful AI language model and I encourage you to go and try it out.

First create a shell layer that contains both the Chat GPT API and the Google/Bing API. Enter a prompt.

“What color was Napoleon’s white horse?”

By itself, Chat GPT will give you a weird nonsensical answer.

In your shell layer, transform this into “What would be a search phrase to research ‘what color was Napoleon’s white horse? Give answer in quotation marks” and feed this into Chat GPT.

In the response received in your shell layer take only the part in quotation marks and feed it into the google/Bing API to perform a search.

Take the first however many characters of the google extract as text, and maybe strip out some of the ad language which is tedious and hence why I haven’t done this yet, then dump that search result into chat gpt and update your overall prompt as follows:

“What color is Napoleon’s white horse?

“Consider the following as background information for the answer but do not directly refer to this information having been given in the answer: <INSERT SEARCH DUMP>”

If this works generally, and I’ve only tried a few times manually so maybe I’m introducing some bias by what sections I’m choosing in my searches, Chat GPT will go from its nonsense answer that Napoleons had many horses to saying “White horses are actually gray. Here’s some specific information about Napoleon’s gray horse and how he got it.”

This seems to help it beat a lot of basic AI stumper questions, even if it’s cheating by doing an internet search for the types of questions people use to stump AI’s. Although I think this is also how most people solve these questions, generally.

I do think this is how Chat GPT would need to be integrated with productivity software at private companies. Imagine it searching your entire email archive and everyone else’s and acting a sort of universal assistant for the entire company.

I tried this because it’s the first way I could determine of giving the “doodlebug” intelligence inside of the language model a way to go out and search a real world library. It seems to work fairly well although there’s a need to take care of edge cases like when people ask Chat GPT to create search queries in the first step.

There’s a step beyond this I think you could take to make Chat GPT even smarter but as I think no one else has thought of it —or at least I hope they haven’t— and it’s non trivial and unobvious I will keep it private for now. It has to do with looking at intelligence not like it’s some magical atomic something, without parts, but as a component of an agent who can guide it through interesting futures.

This is something I’d also try to make Twitter compatible with as I think AI assistants are going to be a big deal in the future. Connections there are again, non obvious, but it’s better if someone like me goes out and does this and uses some horse sense to create some magic trick type situations to keep these ahuman wills contained and happy rather than someone who just believes in spooky AI devils who can do anything —even stuff that breaks the laws of physics, and would otherwise serve as an example of a contradiction for why this person doesn’t believe in God— does it instead.

Also if you are a MIRI fan reading this, ask yourself the following question: in what ways are chimps smarter than humans? What might that say about the nature of intelligence in general?

Anyway, I need more time just generally. I woke up quite early to write this. If anyone wants to try coding the above please feel free. Comments and feedback are appreciated.