ETH Zurich and EPFL are launching the “Swiss AI Initiative”, whose purpose is to position Switzerland as a leading global hub for the development and implementation of transparent and reliable artificial intelligence (AI). The new Alps supercomputer based at the Swiss National Supercomputing Centre (CSCS) provides the supporting world-class infrastructure.
Those are great news. I’m more than happy with the Swiss AI Initiative. I do not think that they’re developing artificial intelligence, or even a language model (just like a taxidermised cat is not a cat), but they’re still developing an open source alternative to a nascent technology that showed some applications.
Now, let me chew on a specially inane HN comment.
> trustworthy AI // >> “… ensure legal, ethical and scientific criteria are met.”
Please… We are talking about stochastic models. This means that we are in the domain of Math, not the domain of Philosophy, and not Law either. // Evaluating a stochastic model, even a multivariate model, involves only two dimensions. Even if it is running on 10k GPUs. Even if it has been trained on billions of data points. // The two dimensions are: // 1) Reliability // 2) Validity
…and that is all. It is Statistics 101. Not only that, it is the most fundamental part of Statistics 101.
I’ll explain why this comment is fucking stupid with a simple proof by counterexample.
Let’s say that your site suggests usernames for your users, based on 6-characters words. And you want it to generate those words automatically.
You’re OK with non-existent words as long as valid for English; e.g. skwerl and wugwug are OK, fdskjd and ngaaaa are not.
So you solve this through a stochastic model; it picks the first letter at random, and then “predicts” the following letters based on the preceding ones. You did a great job, here’s a representative sample of the output: derbie, malcon, hitler, [insert here the N word], vacant, lacoid.
The model is perfectly reliable and valid. And yet there’s still an issue with the output - you don’t want a username like hitler or the N word in your bloody site. This issue is in a third dimension - “ethicality”. Proved by counterexample.
Granted, LLMs are far more complex than that hypothetical username generator, but the same reasoning applies. You actually care about the output of the LLM because it’s going to be used by human beings. You don’t want it to vomit shite that prescribes self-harm, or crimes against humankind, or things that are technically correct but misleading without the context to interpret them accurately.
Those are great news. I’m more than happy with the Swiss AI Initiative. I do not think that they’re developing artificial intelligence, or even a language model (just like a taxidermised cat is not a cat), but they’re still developing an open source alternative to a nascent technology that showed some applications.
Now, let me chew on a specially inane HN comment.
I’ll explain why this comment is fucking stupid with a simple proof by counterexample.
Let’s say that your site suggests usernames for your users, based on 6-characters words. And you want it to generate those words automatically.
You’re OK with non-existent words as long as valid for English; e.g.
skwerlandwugwugare OK,fdskjdandngaaaaare not.So you solve this through a stochastic model; it picks the first letter at random, and then “predicts” the following letters based on the preceding ones. You did a great job, here’s a representative sample of the output:
derbie,malcon,hitler, [insert here the N word],vacant,lacoid.The model is perfectly reliable and valid. And yet there’s still an issue with the output - you don’t want a username like
hitleror the N word in your bloody site. This issue is in a third dimension - “ethicality”. Proved by counterexample.Granted, LLMs are far more complex than that hypothetical username generator, but the same reasoning applies. You actually care about the output of the LLM because it’s going to be used by human beings. You don’t want it to vomit shite that prescribes self-harm, or crimes against humankind, or things that are technically correct but misleading without the context to interpret them accurately.