The Means of AI Production

Large Language Models are the tech du jour right now. However, they pose a danger to society beyond the obvious one.

The past few months have been wild. OpenAI launched ChatGPT, and it seems the whole world is predicting a pseudo-singularity. Large Language Models are the first step to an artificial general intelligence, or AGI. Or so these people contend.

There are a lot of assumptions baked into these prediction. Most of these predictions are seeped in the ‘status quo bias’. For Example; they are assuming the state of progress in Hardware would continue, or that there will always be more data for these LLM’s to evolve further into sentience and intelligence.

There is no reason to believe that we can maintain this Status Quo. For this essay though, we will let these assumptions stand. We will suppose that these models will advance into intelligent ‘agents’, if not AGI. These agents will do any of the tasks you ask them that are possible to achieve via software manipulation. Draw up a website, create a marketing campaign, etc. Whatever you desire.

The obvious danger is the loss of employment for white collar workers. Whole departments can be wholesale replaced by these agents. Plausible as it may sound, it is not really a danger. I will expand upon it, and present the real issue with intelligent agents.

Your Job is Safe

Not really, but also true. Lets say there is an Advertisement firm call Mavericks, Llc. Management has decided to employ an smart agent, called Hermes, to do most of the jobs. They have eliminated 90% of their Sales, Copywriting and Design departments. The remaining employees just provide prompts for Hermes. That are a lot of jobs lost.

Except, what is stopping the laid-off employees to use a similar agent to Hermes to prop-up a competing business to Mavericks? In fact, they have a better opportunity to compete since they would have a bloated management bureaucracy. These business can even be a single person and an AI agent.

However, since the status quo is being maintained, there is a good chance none of these people will have access to run such an agent. The hardware and infrastructure required to run such agents would remain in the hands of few large businesses. The barrier to entry would remain capital

A critique of AI economy

When the web first became widely available, it was seen as democratizing force. The gatekeepers of media and communication could no longer control the output of information. Anyone could launch a website or blog and directly compete with CNN and Der Speigel. The media landscape became noisy and chaotic, but also diverse.

What transpired since is proof of the naivete of this belief. The web became more centralized and brought new gatekeepers. Now the web is a collection of fiefdoms run by Bay Area despots and Chinese mega-corporations. They can decide whether you can exist, and how you are allowed to exist. Loose anti-trust regulations and the might of state power has elevated them to de facto digital governments. Some even have their own ersatz supreme courts.

AI optimists carry similar naive notions about the supposed AI future. And unlike the web utopia, the hope is unwarranted. I can’t just create, train and run my own LLM/GPT model like I could run my own website in 2003. Even with the requisite technical knowledge.

The resources required to do so are financially overwhelming. The only solution is to rent the infrastructure from Google, Amazon, or Microsoft, among other smaller competitors. However these, businesses will be actively competing with me.

How do we democratize the production and distribution of AI?

Agents of the world unite

During the crypto-fad, we were bombarded with claims the decentralized internet. It was bunk. However, we have had a real, peer-to-peer technology for decades. The only reason nobody uses it beyond media piracy is because there is no way to monetize it.

We need an evolution of Bittorrent that distributes not just data but also computation. This is not a new idea. Malicious actors have been using botnets to launch DDOS attack, send spam emails, and in an ironic twist, mine crypto. Many commercial mobile apps use the on-device neural net technology to train their business-specific models.

We need a way to implement it in a way that is open, secure, consensual on the device owners’ part, and most important of all, provides compensation for the said compute time. I do understand that the purely peer-to-peer global network has been pipe dream for decades now. The communication infrastructure and the electronics required for such a network makes it self-contradictory.

And I have no utopian delusions about this setup. Even with a million phones and gaming desktops at our disposal, we still would not be able to compete with the scale of Big Tech. This is not a manifesto for new egalitarian world order of AI, but a way to provide an alternative to current situation. An alternative to the oligopolies.


  • GPT3/4 summary of this essay: “This document discusses the potential dangers of large language models (LLMs) and their impact on the job market. While LLMs may replace white collar jobs, the real problem is the concentration of power and resources in the hands of a few large businesses. The author proposes a solution to democratize the production and distribution of AI through a decentralized, peer-to-peer network for computation.”
  • Counterpoint: This is a great interview that makes the case that the AI apocalypse will come even if the assumptions are incorrect.

I do not have analytics on this site. So leave a comment, even a nasty one, here or on your preferred social site, or shoot an email to feedback at the domain of this site. I also have weekly newsletter where I share wisdom I have acquired in the form of quotes and notes. You can check the archive here. You can sign up below to receive it every week.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.