My current thoughts on LLMs

This is a longer post, here’s the main ideas:

• I don’t trust OpenAI, Anthropic, Google, Facebook, Apple or any of the companies producing the frontier models
• We need to consider people’s and the environment’s needs over companies and AI
• LLMs as a technology are useful, and probably here to stay
• LLMs are like digital librarians: they can help people navigate through the huge swaths of information we all have access to and deal with on a daily basis

Trust

I don’t trust OpenAI, Anthropic, Google, Facebook, Apple* or any of the companies producing the frontier models.

*Apple: They really aren’t on the frontier, but they are one of the few companies prioritizing models that are private and run on your devices. But, they also want to own a proprietary platform and lock users in, generating more service revenue. So they are a mixed bag and it’s not clear cut. Still, I don’t think we should entirely trust their motives and treat them like the rest of the companies on this list.

The mistrust is in these entities as developers and operators of LLMs. The LLMs themselves, I think you need to evaluate trust in a different way. Currently, the LLMs themselves seem to be accurate enough that they are not intentionally misleading you. But, the corporate entities behind them, I don’t think you can count on them to be trustworthy.

There’s so much money pouring into this industry with the expectation of seeing a return on those investments. I think the bubble here will eventually burst and any companies left standing will need to start to show a profit. If they can’t burn through investor cash, where is that money going to come from? The users.

It’s not that I’m against paying for a product I find useful (in fact, I prefer that business model over most others). It’s that the gap between what people are currently paying and what has been spent to develop and run these models is enormous. I think there will be pressure to extract as much money as possible from the users and businesses built on top these LLMs. This could easily play out like it did with Microsoft in the 90s, with monopolistic business practices. At best, enshittification is the outcome.

So what do we do? Well, there’s not a great answer that I can see, but I think we should be skeptical of how AI gets woven into our lives and businesses. I think it’d be best if LLMs become an open commodity where it is easy to develop and run models on accessible hardware. I think avoid tying yourself and your business to the proprietary platforms offered by these companies. Don’t build ChatGPT apps, for example.

Regulation may be an option here, but that is a double edged sword for sure. I think our best bet is to proactively move toward open ecosystems and platforms and avoid the proprietary.

The environmental cost

All of the companies at the frontier are in a giant race to increase compute capacity, resulting in building new, larger, and more power hungry data centers.

Training and, to a lesser extent, hosting these models is incredibly resource intensive. And, of course, all that electric gets converted into heat, so keeping these data centers cool often consumes large amounts of water.

This is a big problem that needs to be curbed.

When these data centers are being constructed it’s not uncommon for the local power utility company to give the data center an incredibly low rate for the electricity they consume. Then the cost of the utility expanding their supply and upgrading the grid often gets pushed onto all the other consumers. I think utility regulators should make it illegal for utilities to give these kind of deals to AI data centers. Instead, the companies building and operating these data centers should bear the cost of the grid and supply upgrades that are needed. They should also require zero carbon, renewable power sources. Let the companies with billions to spend help push along the electric industry toward zero carbon goals.

Even worse is when AI data centers forgo the local utility and generate their own power using natural gas or some other carbon emitting source. This should also be stopped.

On the topic of water consumption, I think local municipalities also need to make sure they put safe limits and regulations around water use for this purpose. I think the best policy is to forbid data center construction near population centers and within city limits. Save the land and water for people.

The environmental and human costs of building out AI infrastructure must be considered while this industry is booming. We should not compromise on this.

LLMs as a Technology

While the economics and levels of investment around AI/LLMs are, I believe, creating a bubble, I think the technology itself, the LLM, is useful and here to stay. Even if today’s LLMs don’t become significantly more powerful or ever achieve AGI (however we define that), I think the current capabilities of these models are incredibly useful.

But, the technology, like all technologies, has its limits. I don’t think AI should replace humans in any workflow or creative endeavor. I think this is, at best, foolish and, at worse, dangerous.

Instead, I think the current models are best at augmenting humans in their endeavors.

There’s so much data on the internet and even within just a single company/organization, it’s way too much for any human or even a group of humans to keep up with. So, I think there’s a huge benefit to use LLMs like a digital librarians to help us navigate through it all. It’s like the computer in Star Trek: you talk to it in natural language, it goes off and searches through mountains of data and comes back with the information you need plus pointers on where to go for more detail.

I think LLMs can also be used to shield people from the mundane and mindless tasks we all do on a daily basis, freeing us up to focus on the creative and important endeavors. Let the LLM go through my email, filter out the spam, and surface the most important, and actionable messages. This and a million other use cases are all good fits for LLMs.

Conclusion

There’s a quote by Brendon Bigley on a recent episode of First, Last, Everything:

“…that’s not inherently a terrible idea. It’s just that the worst people alive got their hands on it first…”

He was talking about crypto and NFTs here, but I think something similar is relevant for LLMs. I think it’s regrettable how this technology is being rolled out, but the technology itself is not a terrible.

But, unlike crypto, I think LLMs are mainstream and not going away anytime soon, so we ought to find ways that we can leverage this technology and shape our relationship to it to maximize positive outcomes for people over corporations and capitalists.

I think success looks like people benefitting from and empowered by LLMs, rather than replaced by them. The environment is not sacrificed to power LLMs. We build open, communal systems and platforms rather than proprietary ones to enable a new generation of applications and systems.

If we focus on this, I think we can avoid dystopia. Well, at least an AI dystopia.