#AI #ScientificML #CloudLLMs #ChatGPT #Gemini #Claude #Ente #Firefox #Mozilla

Types

First, we should be clear what we're talking about here, because “AI” can mean a ton of different things. AI could mean any and all of:

1. Scientific machine learning, used within limited contexts to elucidate complex relationships within data and tease out more complicated models for which we don't currently have analytical forms.
2. Cloud-based LLMs, of the chatbot style (e.g. ChatGPT, Gemini, etc).
3. Local ML models run on-device (e.g. Firefox's on-device translation feature, Ente's machine learning, etc).

Putting them all into the same bucket is sort of silly, right? So let's address each in turn.

Scientific ML

Scientific machine learning models tend to be limited in scope and the good ones build in validation protocols. I might have my own concerns here around whether having a model for which we don't have an analytical form is useful. There are certainly variants here which attempt to extract equations from data (which brings its own challenges), but the concerns here are around the scientific usefulness. The data all has good provenance (usually generated by the researcher themselves), the scope is limited (e.g. we want to better predict the behavior of this specific material), and energy usage etc is wrapped up into the broader HPC system (this could change of course, if we need to scale up the HPC systems to handle more ML work, but that's a concern in the future and assumes these types of methods are broadly applicable in every context and that we need to train these models on the HPC...quite often, if a model is well-designed, it can be trained locally anyway, bypassing this issue entirely).

Cloud-based LLMs

Cloud-based LLMs like ChatGPT, Claude, Gemini, and so on raise the most alarm bells for me, and these are the ones I am completely against. I find issues with data provenance (did folks consent to have their data used for training these models?) and the fact that these companies seem to want to gatekeep knowledge. The ideal case for these people is that we can no longer function without LLMs and we pay them money indefinitely to continue to be able to function. There are the privacy concerns (these companies are storing your chats and can produce them on request) and the environmental concerns (there are enough articles out there about this, from the renewed building of gas power plants to folks' water becoming impotable to the fact that these data centers are being built in drought-prone areas). There are the ethical concerns around the fact that the pitch to companies is that they can replace all of their workers with these LLMs. There is the fact that they have no notion of 'truth' (so they don't 'hallucinate', they just spit out bullshit —– hallucination implies that they have a working model of the world and are imagining things which are not there, which simply is not true of these LLMs). There are bandaids around some of these concerns. Some companies provide interfaces to these LLMs without storing chats (DuckDuckGo) or storing them client-side encrypted (Proton). But nothing addresses all of these concerns.

Then there are the practical concerns. I simply don't see a use for them. I enjoy coding. I enjoy thinking about how to design interfaces and classes and modules so that extending them is easy. I enjoy debugging (within reason, of course). I enjoy deriving equations to figure out if something is correct. And I enjoy thinking. Why would I give that up to an LLM? Why would I prefer chatting with a sycophantic plausible-sounding bullshit generator rather than an actual human being with their own ideas and their own viewpoints? So much of this just feels like hype for hype's sake, where something sounds 'cool' so people jump on the bandwagon.

The basic fact is that if I have to double-check everything an LLM states anyway, I'm better off doing the research myself. I'm better off using Google Scholar (as long as that's still around /sigh) to search for research papers than asking an LLM to find stuff for me. Is there a good chance it will find stuff? Sure. Is there also a decent chance it will hallucinate papers that don't exist or imply a paper says something it doesn't? Also yes. The number of times I've looked at a DuckDuckGo AI summary and found it claiming something which simply isn't true is...non-negligible. It's not super high, but it's also disturbingly common. And that's the issue, right? It's not so good that I can just rely on it, but it's also not so bad that I can just assume everything is bullshit. It's at this level where most of what it's saying is probably accurate, increasing the cognitive overhead beyond what it would take for me to just search for stuff on my own.

Local models

Local models, I think, are the future of “AI” insofar as we want natural language interactions with machines. They are cheaper to run. They can be designed with good data provenance (sidestepping potential lawsuits and ethical issues). They don't require gigantic data centers. But they're harder to build for all of these reasons.

Something like Ente's on-device ML for face recognition and semantic search is great because the facial recognition data is encrypted before being uploaded to the server and all ML processing happens on-device. Firefox, similarly, has an on-device translation option which doesn't touch the cloud. I know Ente uses models with good provenance, but I can't figure out where the Firefox Translations dataset is. At the very least, though, any local ML model must have open datasets and open weights and biases for it to be ethical.

Conclusion

My strong feelings on 'AI' mostly derive from my strong feelings about cloud-based LLMs. There is a bait-and-switch tactic where folks will imply someone is against other forms of ML (including scientific ML or local models) just because they are against what OpenAI, Anthropic, and the rest are doing. I'm not against AI or ML in general, but this space has been co-opted by these giant companies doing highly unethical things. We can simultaneously be pro-ML (used judiciously) while also being anti-'AI' (as defined by these giant companies).