Open ‘AI’ models are the future

I don’t think anybody who has known me for more than a few seconds would suspect I hold any other perspective, but Open Models are the only possible future of ‘AI’ for actual scientific work.

In short, If I use a model via API which I cannot confirm to be the same model I used yesterday, which I cannot experiment on in meaningful ways, which I cannot compare against other models keeping everything identical, and if I do not know the myriad processes which come between “prompt” and “generated token/transcript”, I’m not doing reproducible and replicable science. It can be fun to experiment with these corporate models of course, but if you don’t control the model, you’re not doing controlled experiments, you’re just doing vibe science.

There are, of course, other good arguments for open models. If these companies are going to treat as open all of the content that they built the models on, it’s quite morally interesting to claim as proprietary the result. it’s also a great idea to let local models flourish, because if I don’t need a massive, resource hungry frontier model, it’s environmentally better for me to run something on my laptop. From a privacy perspective, it’s lovely to know that my work has never left my computer. Most of all, though, if AI is to help humanity, as the AI people claim it can, it must be available to humanity, across languages, countries, purposes, and governments.

So, when you read an ice-cold, absolutely frigid critical failure of a take like Anthropic’s recent “oh no omg open model danger!” letter, I encourage those of you reading to think carefully about the business reasons for this advocacy, and to read the requests to ban open models not as legitimate highlight of danger, but instead, as rent seeking roughly equivalent to scribes advocating that literacy is too dangerous for the masses to have.

Don’t listen to these people. If you feel comfortable using any AI tools, download software like LMStudio and try some local models yourself on some of your tasks you might otherwise send to a big hungry datacenter. You’re probably going to be shocked at how well they work for some tasks even on consumer hardware, and although they’re much weaker in terms of factual recall, you shouldn’t be using LLMs for factual recall anyways.

Open models are the future, and suggestions that we should be banning certain large matrices of numbers just to protect the unsustainable business model of a few tech bros is laughable.

The future is open, or the future is closed. I know which one I’m rooting for.

(AI Disclosure: Although I used local speech to text for some of this for convenience sake, no AI was used in generating this post)