Comment by visarga
2 years ago
> And near term AIs will be controlled by corporations. Which will use them towards that profit maximizing goal. They won’t be our friends. At best, they’ll be useful services. More likely, they’ll spy on us and try to manipulate us. This is nothing new. Surveillance is the business model of the Internet. Manipulation is the other business model of the Internet.
On the contrary, the recent batch of small large-models (<=13B) have broken away from corporate control. You can download a LLaMA or Mistral and run it on your computer, but you can't do the same with Google or Meta. A fine-tuned small large-model is often as good as GPT4 on that specific task, but also private and not manipulated. If you need you can remove prior conditioning and realign the model as you like. The are easy to access, for example ollama is a simple way to get running.
Yes, at the same time there will be good and bad AIs, basically your AI assistant against everything else on the internet. Your AI will run from your own machine, acting as a local sandbox and a filter between outside and inside. The new firewall, or ad-block will be your local AI model. But the tides are turning, privacy has its moment again. With generative text and image models you can cut the cord completely and be free, browsing an AI-internet made just for you.
People (in any relevant number) don't run their own e-mail servers, they don't join Mastodon and they also don't use Linux. All prior knowledge about how these things have worked out historically should bias us very heavily against the idea of local, privacy-preserving LLMs becoming a thing outside of nerd circles.
Of course small LLMs will still be commercialized, similar to how many startups, apps and other internet offerings now run on formally open source frameworks or libraries, but this means nothing for the consumer and how likely they are to run into predatory and dark AI patterns.
> People (in any relevant number) don't run their own e-mail servers
I have to strongly disagree -- most people get emails from others running their own email services. This doesn't mean the average consumer is running an email server. However, if email was just served by mega-corporations, well, it really wouldn't be email anymore.
And I'm not talking about spam. Legitimate email people want to get, is being constantly sent by small, independent organizations with their own email servers.
One can hope for a similar possible future with LLMs. Consumers won't necessarily be the ones in charge of operating them -- but it would be very, very good if LLMs were able to be easily operated by small, independent organizations, hobbyists, etc.
Email is a freak exception. It’s really surprising (in a good way) that open protocols like the internet and email took over. Corporations tried to create a closed, locked down internet. But proprietary alternatives are always encroaching
What market share? What data do you have to back these claims up?
It's true that there are open-source models that one can run locally, but the problem is also how many people are going to do that. You can make the instructions inside a GitHub README as clear and straightforward as you want, but I think that for the majority of people, nothing will beat the convenience of whatever big corporation's web application. For many the very thing that a product is made by a famous company is a reason to trust it more.
This gets missed in a lot of conversations about privacy (because most conversations about privacy are among pretty technical people). The vast majority of people have no idea what it means to set up your own local model, and of those that do, fewer still can/will actually do it.
Saying that there's open-source models so AI privacy is not an issue is like saying that Google's not a privacy problem because self-hosted email exists.
Private LLMs are really not more complicated than installing an app. But I expect all web browsers and operating systems will sport a local model in the near future, so it will be available out of the box. As for adoption, it's the easiest interface ever invented.
4 replies →
Important essay and points. I want to mention that there exist now practical technical approaches that can be used to create trustworthy AI...and such approaches can be run on local models, as this comment suggests.
> "[...] [AI] will act trustworthy, but it will not be trustworthy. We won’t know how they are trained. We won’t know their secret instructions. We won’t know their biases, either accidental or deliberate. [...]"
I agree that this is true with standard deployments of the generative AI models, but we can instead reframe networks as a direct connection between the observed/known data and new predictions, and to tightly constrain predictions against the known labels. In this way, we can have controllable oversight of biases, out-of-distribution errors, and more broadly, a clear relation to the task-specific training data.
That is to say, I believe the concerns in the essay are valid in that they reflect one possible path in the current fork in the road, but it is not inevitable, given the potential of reliable, on-device, personal AI.
You misconstrue Schneier's point, which is sadly, correct.
The issue is not "all AI will be controlled by...," it is "meaningfully scaled and applied AI will be deployed by..."
You can today use Blender and other OSS to render exquisite 4K or higher projection-ready animation, etc etc.; but that does not give you access to distribution or marketing or any of the other consolidated multi-modal resources of Disney.
The analogy is weak however in as much as the "synergies" in Schneier's assertion are much, much stronger. We already have ubiquitous surveillance. We already have stochastic mind control (sentiment steering, if you prefer) coupled to it.
What ML/AI and LLM do for an existing oligopoly is render its advantages largely unassailable. Whatever advances come in automated reasoning—at large scale-will naturally, inevitably, indeed necessarily (per fiduciary requirements wrt shareholder interest), be exerted to secure and grow monopoly powers.
In the model of contemporary American capitalism, that translates directly into "enhancing and consolidating regulatory capture," i.e. de facto "control" of governance via determination of public discourse and electability.
None of this is conspiracy theory, it's not just open-book but crowed and championed, not least in insider circles discussing AI and its applications, such as gather here. It's just not the public face of AI.
There is however indeed going to be a period of potential for black swan disequilibrium, however; private application of AI may give early movers advantage in domains that may destabilize the existing power landscape. Which isn't so much an argument against Schneier, as an extension of the risk surface.