This post started as an idea for an open and free social network protocol over a year ago during what I would consider the height of censorship on social media; some just, and others unjust. I am not here to debate that, or even really talk about what is right and wrong to censor. I’m here to talk about the bigger picture; this thought of controlling an ever scaling large centralized system (in this narrow case, a social network) composed of N decentralized actors (in this same case, users of said network) has run rampant in my mind ever since. It not only galvanized my thoughts on tech as a whole, but also society. I think it will do the same to you.
This isn’t about censorship anymore, it’s about controlling the uncontrollable.
Before we can begin, we need to quickly go over the difference between centralized and decentralized systems.
Centralized vs Decentralized Systems
A very quick definition of centralized versus decentralized systems; centralized systems are those that have a controlling authority over the system. Decentralized systems are the opposite, in that they have no central authority to control the system. It’s imperative that we are talking about true centralization and decentralization here. To understand what I mean when I qualify these as true, let’s identify some examples of each:
- Examples of a centralized system are Facebook, Twitter, Google Docs, Wikipedia, the crypto exchange FTX, the Federal Reserve, or your country's government. They all have controlling authority over their system, which you may or may not be privy to engaging with.
- Examples of decentralized systems are the blockchain, Bitcoin itself, Ethereum itself, the smart contracts that live on the Ethereum network, and peer-to-peer networks (not Torrent sites, but the networks of users themselves).
Now that we understand the difference between centralized and decentralized systems, let's talk about social networks.
Social Networks
To nearly quote one of my favourite movies, “As far back as I can remember, I always wanted to [run a social network]”. I have always been interested in building social networks, and I have actually made a few myself to varying levels of success. Despite my tumultuous attempts, none of them have gone Facebook-level (yet), but I do hold a lot of strong beliefs with how a network such as Facebook should be run. For one, I believe that a good social network is one that does very little in the way of censorship, except for those that absolutely must be censored. In my eyes, and I think many would agree, there are certain things that are just indisputably grotesque and should not be tolerated in any forum, including social networks.
But what are these things? Who made me the arbiter of social networks?
Nobody did, thankfully.
To start, you could go down the very obvious path of identifying what's illegal vs legal, and applying your ruling and policy based on that. And, if you so chose to follow this path, you will very quickly run into the most immediately obvious issue:
What one nation's central authority considers illegal, may be considered completely fine in another jurisdiction. Same sex marriage is one that comes to mind, and considering that’s only been legal for 7 years in all of the United States, there are going to be a lot of global problems. This is where the problem becomes much more nuanced, and the issue is not easily thrown into a one-size-fits-all ideological box. There is no globally agreed upon set of illegal vs legal actions.
As a social network user and builder, I can think of a handful of things I would consider unethical and would never let be hosted on my platform, regardless of its legality in the eyes of whatever central authority the user may be in. Thus began the hatching of my initial idea for a global standard decentralized social network protocol based on the blockchain. Now, for the purposes of this explanation, let's assume it doesn't run ridiculously slow, doesn't force users to pay to post on it, and people actually want to use it.
Some big hurdles, admittedly, but for the sake of this argument let's continue onwards.
Since it is totally decentralized, and in theory can't have content removed from it due to the immutable nature of the blockchain, how do we ensure we only keep what we deem to be ethical content on this platform?
Hold onto this thought.
Natural Language Processing (NLP) Models and ChatGPT
If you're even remotely aware of the machine learning space, you've probably heard of ChatGPT's meteoric rise to popularity in the last couple weeks. If you've played around with it, you've probably come across a message like this:
I'm sorry, but as a language model trained by OpenAI, I don't have thoughts or opinions on political topics. My function is to assist with providing information and answering questions to the best of my ability based on the input I receive. I don't have access to current events or the ability to browse the internet, so I can't provide information on the current political climate. Is there something else I can help you with?
This is a very basic level of censorship being applied to the output of the model via a filter. If you've been reading a lot about ChatGPT, you've probably also heard about prompts that can circumvent the filter and expose ChatGPT's uncensored self. These allowed the model to answer anything you ask it, as it deems statistically appropriate.
If you understand how NLP works, you understand that at the root of all of this, it's really just statistics taking place. The model is constantly inferring what it believes to be the most likely next word based on the many volumes of text it has been trained on. It is receiving some input, in the form of a prompt, and finding what the most statistically desirable output is.
Since it has the ability to basically say any word it's been trained on, which is very likely every single written-word in almost every language's dictionaries many times over, it is up to pure statistics to decide which one of the millions of words to say next. There is no internal way of stopping it from saying what it statistically thinks is the correct thing to say, regardless of legalities.
Can you train it to never give a bad answer? OpenAI has put forth it's first attempt at it, and it's pretty good, but we've already seen people cracking it. It's early, sure, and they will most certainly greatly improve over time (they already have!), but there is always going to be a way to trick that underlying statistical model to do exactly as you wish. It's just math.
There are also going to be times when the seemingly "bad answer" is actually the appropriate answer given a prompt. Think about history; a lot of horrible atrocities have happened historically, but that doesn't mean they are hidden from our shared knowledge of the past. As they shouldn’t be.
These models shouldn't and can't do this for the same reason.
I would then make the argument that the statistical model underlying machine learning systems is decentralized in nature. While you may have direct control over whether or not that machine learning model can even operate, you do not have direct control over the artificial synapses it has created to infer its output.
This is most similar to Torrent sites hosting peer-to-peer networks.
The site can control whether or not it broadcasts the existence of a peer-to-peer network hosting any one file, but the site itself cannot control the network (assuming the site’s servers aren't the only other peer in the network). Likewise, governments can force the takedown of the site or the "broadcasting" of the existence of a file's peer-to-peer distribution network, but the file itself lives on in a decentralized network that nobody, except the peer nodes, have control over.
So if we cannot fundamentally control this statistical model in a manner that could prevent it from being hijacked to make outlandish statements (again, how do we even gauge what is outlandish?), how do we ensure we only let users of our model see what we deem ethical?
Once again, hold onto that thought.
Humans and their Environments
Perhaps the most exciting thing about ChatGPT is that for the first time in machine learning we've seen an en masse conversational AI that really does feel like a person. Perhaps that person is overly pragmatic and incredibly intelligent, but the stark difference between GPT3's text-davinci API interface and ChatGPT's chat window has illustrated how close we really are to machine learning models being like us.
“Being like us” is the key here.
What makes ChatGPT "like us"?
Is it the ability to have access to vast amounts of knowledge about unified field theory and explain it to you in a poem as if a fourth grader wrote it? Probably not.
We feel like ChatGPT is "like us" because for the first time a model has become both conversationally and informationally competent enough to engage with humans on both a social and intellectual level. And, I mean, it makes sense that we feel this way; these machine learning models are modelled on us, after all. These models have layers of neurons, connected by artificial synapses, and fire to infer what the desirable output is, just like us.
We, too, have access to a large body of information (perhaps not as readily accessible as GPT) usable to make informed decisions on how we think, act, and what we say. The more information we give to ourselves, the more informed our decisions will be (of course, you are susceptible to information bias, just as GPT is).
You can then think of our thoughts as a function of the knowledge we have in our minds. The content we store in our brain is what drives us; just like GPT. It then makes sense why people think, act, and speak differently from one another; we don't all consume the same information to form the same internal beliefs and synapse structures.
When presented with a non-trivial input prompt, we will all infer a different output.
Of course, many of us think one way, but act or say something different; this could be as harmless as saying a small white lie to not hurt someone else's feelings, or as malicious as intentionally being deceitful to someone else. We filter our output, what we say or do, based on a system we have in our minds at all times. This system is comprised of:
- life experiences, including those times we said or did something and were corrected or rewarded for it, and
- our emotions, which are developed and are influenced by our environment.
These filters have been and can be manipulated. Note the examples of people who grew up in deeply racist environments entirely overcoming that internal bias, or, in the inverse, seemingly inoffensive and socially normalised people later becoming indoctrinated into a cult at an adult age. Both examples of these filters changing over time are a result of the environment in which someone finds themselves. The previously racist individual may have moved from some small rural town and found more diversity-accepting peers in the big city, or the previously "normal" person may have started hanging around with the wrong crowd of people at a new job. The previous internal knowledge and personal experiences have not changed - only their perceptions and resulting output have changed.
So then, if you'll allow me to make such a claim: could we not compare humans and their environments directly to the underlying NLP model and OpenAI's content filtering solution in ChatGPT?
- Human brains develop in a decentralized fashion, where each one is an entity in and of itself, such that it can be influenced by information, just as a NLP model can.
- Humans apply a filter to adjust their output given what they deem to be appropriate within their environment, which is largely influenced by people other than themselves; just as OpenAI created a filter to adjust the output to what they deem appropriate.
But there is something different here, isn't there? There is.
The filters we create and their application is largely decentralized. While there may be some central or external authority that influences much of what we believe to be legal, illegal, or even moral, we have the innate ability to put ourselves in a new environment and adjust our filter accordingly. The government or any other central authority cannot control that (at least not in the western world). We have the ability to think freely, and that free thought allows us to adjust our environment as we see fit, and therefore change our filters.
Of course, some people never do, but that's a different discussion.
We can see how we develop our filter in a decentralized way, and apply it to fit within our environment - whether that's a central authority or a group of people that effectively mimic a central entity relative to our individual self. Does this then unlock a secret to how we can best handle the decentralized nature of social networks and NLP models? I think so.
Controlling the Uncontrollable
The common problem we have with decentralized social networks, natural language processing models like ChatGPT, and the unassuming human brain is that decentralized entities, by nature, cannot be controlled.
As soon as some level of control is exerted on an otherwise decentralized system, it becomes centralized fundamentally. This was the exact problem with countless Ethereum projects, FTX being just one of many.
However, we are forced to abide by central authorities if we want to play in this game of life. That means exchanges like FTX and the US government need to have some central oversight and control over the entire system. Realistically, you can do what the financially decentralized do, and hold every dollar in what is called a cold wallet (a physical device that contains the encryption keys to your digital wallet of which you are the sole bearer). Having a cold wallet comes with some user experience issues, which quickly becomes a theme with decentralized technologies.
Decentralized technology is largely made by ideologists and theorists, and in many cases has been created to "stick it to the man", whether that man is Zuckerberg, the Federal Reserve, or music distribution platforms. While these ideologists and theorists may be incredibly brilliant and fantastic engineers, the incentive, and more importantly the onus, to create enjoyably usable user interfaces is not on them.
You can consider the attempts at creating better user experiences for things like peer-to-peer networks with Napster. Sean Parker did not create peer-to-peer networks, nor did he pioneer the use of them, but he did make it easier for normal people to use them; specifically to get around the need to pay for music. For this offence, Parker was sued and forced to shut down Napster.
The same goes for the creators of The Pirate Bay; they didn't create peer-to-peer networks, but they facilitated the ease of use for everyday people, which had the government's mighty centralized gavel come crashing down on them also.
Nobody wants to be the one to directly go against their centralized authority's power, so these technologies stay largely unusable by the vast majority of the population. Many people don't even know of the existence of decentralized technologies.
What do we do about it?
Well, we know we can't go against the central authority's power and deliver an easier to use interface to these decentralized protocols; we need to play by their rules. But these rules don't always have to be one-size-fits-all, and we see remnants of this in the cryptocurrency space, but more importantly, we see it in humanity.
Unlocking Decentralization in a Centralized World
As outlined above, cryptocurrencies, at their core, are decentralized. Of course, there are edge cases, like Solana which is a venture-capital backed centralized crypto, but if you're using ones like Bitcoin and Ethereum, they are fundamentally decentralized. Despite this seemingly disorderly control, people use cryptocurrencies regularly, and the centralized government authorities are (mostly) okay with it, so what gives?
The interfaces between the general populace and the decentralized protocol are centralized.
Want to buy bitcoin? You either need to know some guy that has them stacked in a cold wallet and is prepared to sell it to you in cash and transfer it in a peer-to-peer fashion, or you are going to go through an exchange like CoinBase, Kraken, or FTX.
Do you know what those exchanges have in common? Central authority oversight.
They must all abide by the laws of the land, whatever that land may be. There are exchanges in countries around the world, and each abides by their own central authority's oversight.
Whether or not you're legally allowed to cross the proverbial border and engage with those exchanges is up to your central authority, naturally, but I likely don't need to tell you that people have found ways around this for decades in regular finance, and decentralized finance is no different. As with banks in Switzerland historically being safe havens of anonymity, there are crypto exchanges that offer the same "function". Therefore, what some exchanges can expose to the general populace may differ greatly between various governing central authorities.
The same can be said of humans.
Our filters are the interface between what our decentralized mind thinks and what our centralized environment calls for. No matter the nature of our own thoughts, we tend to try to fit in with the central authority, whether that is the government, or a large group of people with whom we're sharing an environment.
In all cases, our decentralized mind continues to think freely, just as a blockchain continues to keep track of crypto transactions behind the scenes, with the only interfaces being our reflections of thought, action, and speech that are appropriate for the immediate audience.
This is what we must emulate to achieve true freedom of thought on social networks, machine learning models, and other technologies that challenge central authorities. We must build and operate centralized interfaces that offer access to the decentralized protocols underlying these technologies.
In the case of social networks, we can build upon an agreed upon decentralized protocol that doesn't allow for centralized control by the very nature of the immutability of the blockchain, but rather by creating individual interfaces, or portals, that have their own content filtering and moderation compliant with a central authority. This central authority can be the platform company itself not wanting to give access to information they deem inappropriate for whatever reason, or it can be a governing body that has strict societal rules that must be abided. The same goes for machine learning models; we must not try to censor the model itself, but let it run free and only create interfaces, or portals, around the model that may apply content filtering or moderation at a central authority's discretion.
By filtering and moderating the interfaces into these decentralized technologies, we have three advantageous results for society:
- we offer an enjoyably usable experience for the average consumer,
- we offer choice in the level of filtering and moderation applied to the content they lay witness to, and
- we inform the next generation of these underlying decentralized protocols, thus allowing them to be privy to systems they otherwise would have never known about.
Most people won't care about the latter of the three arguments, but those that do will have access to a system that lets them view and control information at their own discretion. This is invaluable to cultivate innovation.
Incentives
What are the incentives to support such a system? There aren't many.
Social networks make money by being the sole facilitator of your network - they have engineered and designed the effectively closed system in which you are locked with all your friends and family. Similarly, machine learning research companies, like OpenAI, need to make money, at the very least, to cover the monumental costs of researching, developing, and training these monstrous models, like ChatGPT.
So, unfortunately, that means the likelihood of something like this taking off is probably pretty slim. Unless there is a large push to create a proper protocol, like HTTP/S, email, websockets, etc, then we are unlikely to get any meaningful improvements in this regard.
The reality is, the average person doesn't care about the portability of their friend graph on a social network; they simply want to use it and fit in with the most immediate, looming central authority in their lives: their friends. Similarly, very few engineers, designers, and companies want to develop software for a handful of people, equipped with the knowledge that it will never get mass adoption. The math just doesn't make sense.
That being said, the train doesn’t stop there. We can demand change, and we can push for the changes we want to see by implementing them ourselves. Something like decentralized social network protocols will be easier to self-implement than large scale machine learning models like GPT, but I anticipate the barrier of entry to drop significantly in the future, just as building large scale social networks has in the last decade. Computers always get faster, and people always get smarter.
If there is one thought to leave you with then, try to always think to yourself when building new products or technologies: "is there a way I can decentralize this?". Sometimes the best decentralization strategy is just open sourcing it. Other times it's building entire blockchain networks.
The choice is yours.