The more I read and write about AI, the more I become convinced that Big Tech incumbents absolutely have to strangle decentralized AI in its cradle before it destroys everything.
Let’s look at this interview Ben Thompson collaborated with OpenAI’s Sam Altman and Microsoft CTO Kevin Scott. Pay special attention to the following parts.
From Microsoft’s perspective, is this a funnel for new products, or do you see winning search as an end goal in itself?
KS: So I think you made a very important point. That said, even if the advertising economics of this system don’t have the same economics as “regular search,” it would be great for Microsoft if it could capture market share. I think we have a lot of capability here. Part of the reason is because we’ve done so much performance optimization work and we’re very confident in terms of cost and understanding what our business model is. He worked at Google as a pre-IPO employee, and what I do know is that the search business today is very different from the search business he was in 20 years ago. Understand what your business model is when it comes to ad units. Microsoft is more than capable of doing all of that profitably.
SA: There’s so much value here. I can’t imagine that we can’t do that. Think about how to ring the cash register bell. [Emphasis added]
I recently said the same thing to an interviewer who asked me about search and Google. My point was that Google has appeared in search for the first time before. That business model didn’t exist until Google came up with it (via acquisition). I argued that no matter what kind of user experience BingGPT and Bard evolve into, don’t expect another profitable model to emerge.
But I also made another point to the interviewer that was not captured at all above, but is of great importance to anyone thinking about technology policy at the moment. For the business model to work, decentralized AI must first be disrupted.
Centralized and decentralized
Implicit in Scott and Altman’s vision is a set of assumptions about how AI-powered chat will eventually “ring the cash register” as a new query interface for most information.
- Users access a centralized server and enter text into a hosted box.
- Advertisers access the same server and appear in front of all users.
- With Microsoft acting as an intermediary, advertisers and users are somehow connected to each other.
- Alternatively, users pay Microsoft directly for their queries via a subscription or micropayment scheme.
In other words, Microsoft’s ability to squeeze profits out of the experience of interacting with LLM is predicated on billions of users continuing to flock to a small number of centralized services to have their queries answered. So this is a vision that assumes a world of centralized AI.
But what would happen if we found ourselves in a world like this? decentralized AI Instead? What if you could download an app that answered your current questions from all of Wikipedia and Reddit, and in some cases access both of these sites to obtain new data?
What if some of my data sources are my favorite news websites and forums, and they all sign up to provide data to the app, and I receive a cut of the revenue generated by the app? ?
Or what if those apps leverage open source language models and are kept up-to-date by accessing current data sources via APIs? We are seeing such apps being published on their own with the ability to answer any question from a huge archive of past questions.
Decentralized AI is a real threat
To provide some technical background as to why the app-based decentralized AI vision described above is very possible, the models needed to make this happen could each be on the order of a few gigabytes in size. Please consider that there is. For example, the stable diffusion model files that power image generation range from 2.5 to 4.5 GB, depending on the version, and were trained on 240 TB of image data. This is an amazing level of compression.
So, for example, the average size of a model that needs to answer 75% of random questions about the world might be around 3GB or so. That’s about the size of a large mobile game download.
Why go to a website hosted by Microsoft or Google and type your query into a text box when you can download a model that reliably answers your questions about your training data? Your favorite recipe site If you need a recipe for , you might go to that site and ask for advice instead. their model.If you want the current NYT or WaPo consensus on Ukraine, why not go to those sites and chat? their Bot? Why should Microsoft or Google get involved in this?
The answer, of course, is that they don’t need to be involved. Decentralized AI can and will eliminate them completely. Assuming it is allowed.
But this is a big premise, as the future of decentralized AI is by no means guaranteed.
But before we explain who and why are trying to kill decentralized AI, there are a few caveats.
- Using models to answer questions requires significant computing power. However, as this is an active research area, these inference costs can and will be reduced through innovation. Also, have you looked at your cell phone lately? There is no shortage of computing power, and cell phone manufacturers are always looking for ways to harness it. After a few product cycles optimizing the hardware for query execution, it’s not hard to imagine getting very fast local performance for many different types of models.
- Yes, models still make up facts. This hallucination is a big problem, but one that everyone is dealing with. The model will be able to more faithfully represent the facts in the data source.
We must fight for a decentralized future
I wrote about this in detail on Substack. Forces gather against decentralized AISo I won’t repeat it here. But to summarize, the aforementioned model files that represent the “brains” of AI such as Stable Diffusion and ChatGPT can very easily be treated like digital contraband and wiped from the internet.
Everyone from Googlers to Google-hating ex-Googlers to indie artists to profiteering lawyers are saying these model files are child pornography, 3D printed gun files, pirated movies, spam, and malware. We work hard to build a rationale for why we are subject to censorship. .
Below are some of the rationales currently being considered for banning decentralized AI.
- All model files are full of copyright violations because they were trained using copyrighted data.
- Generative text models can harm marginalized people because they can induce “hate speech” to them.
- Generative text models catastrophically increase the threat of “disinformation.”
- Generated image models are used for fake pornography of real people, many of them children, without their consent.
We don’t even need new laws to ban these model files. All that is needed is agreement among a few major companies that these files and the apps and sites based on them are a threat. I imagine that the following platforms can, and probably will, work together to create what amounts to a de facto ban on decentralized AI.
- google play
- Apple’s App Store
- Amazon Web Services
- cloudflare
This means a world where everyone can host their own models, backed by their own data sources, but facts are never guaranteed.I’ll go history lessonsI think it’s probably not likely.
It seems increasingly likely that centralizing forces will succeed in treating unauthorized model files like contraband. Five years from now, we’ll still be running all our queries on servers hosted by one of the big tech platforms.
I hope I’m wrong, but I know that if we’re going to have decentralized AI, we’re going to have to fight for it.
