Hacker News Reader

Comment by calebrob6 | Hacker News Reader

It looks like it needs to go through toxicity testing -- https://twitter.com/WizardLM_AI/status/1780101465950105775

I can't imagine how toxicity in a LLM can be harmful. Can't we write stories with toxic characters, for example?

It feels like the greatest minds of our era are creating an amazing piece of technology. And we are hindering it in the name of corporate cover ass and bullshit jobs.

if you read the large theory papers on human-computer LLM interfaces, the people who build these are genuinely worried about "harm" . Impolitely, it appears from the outside that the kind of researchers that they have attracted over a decade to do this fantastically tedious and abstract work, are covered in emotional illness symptoms personally, and developed a culture of incessantly declaring "harm" in every shadow of every corner; at the same time, corporate black-hearts have money on the mind, and are genuinely worried about "harm" in the form of consumer retaliation in the marketplace, massive legal liability for civil rights blunders, and losing the sweet spot to a competitor; then government at the executive level, at the Nation-State, are obsessing about obtaining and implementing AI for competitive advantage against just about every other group of people you can name -- as long as no one can prove that they implemented "harm" while getting unprecedented competitive advantages at scale over populations of unwitting civilians, their geopolitical rivals, and probably other political types of a different stripe.

So no, it is not "bs jobs" at all .. but worse

That's another point, but I agree. The government's incentives are to become bigger and more powerful. LLM control could be a fast lane to become a totalitarian state. Governments around the world might not do that by choice, but the incentive is true and is present.

If you think that you want to live in a world where your life is heavily influenced by machines that were trained on the idea that you don’t deserve to exist, then yes, “toxicity” isn’t a problem.

But you don’t think that. Even if you think you do.

But instead, they are trained on what some corporation or government thinks is good for me. How's that any better? Do you trust them to be neutral and act in your best interest? Who defined toxicity?

A "neutral" policy here is probably a still unsolved philosophy problem.

I already live in this world. That is why I don't want my models censored; they're already spewing that toxicity.

(It's just your brand of toxicity, so naturally, you don't think it's a problem.)

You can't even define toxicity in an objective and verifiable way, because it's inherently subjective.

Trying to make rules for a machine to behave in a decidedly nontoxic way is a fool's errand, then.

You're also assuming that AI is going to be used to heavily influence people's lives, but there's a good chance that all it's good for is ripping off copyrighted material and generating clipart that's good enough for powerpoint presentations.

AI is probably going to change the world in the way that NFTs did. And self driving cars. And the Alexa.

> It feels like the greatest minds of our era are

This is the new “we could put a man on the moon, yet…”. No, “the greatest minds of our era” are not working in adtech or building LLMs. It’s easy to forget, but there is a whole world outside of computers, and being good at it does not equate to being “a great mind”. It is absurd to believe the greatest minds are all working around the same problem spaces.

Fair, but that's not central to my argument. My point still holds.

So this is a good thing for people like me who don't care if some piece of software is "toxic" with me or not.

It was trained on content from the internet. I'd be massively surprised if it somehow wasn't toxic. Humanity (or a small portion of it) is full of assholes. As much as that sucks, shouldn't the embeddings reflect the reality of the training content? If you want fluffy bunnies and flowers and happy people holding hands, shouldn't you just train it on that content?

Toxicity testing? That a self-imposed requirement by Microsoft?

I don’t think they forgot about Tay.

https://spectrum.ieee.org/in-2016-microsofts-racist-chatbot-...

I asked GPT-4 to rewrite the refrain from Eminem's "Forgot about Dre" but change it to "Forgot about Tay" and make it all about chatbots... this is the best one it came up with:

Nowadays, every bot wanna chat

Like they got something to say, but these LLMs

Are too toxic to use, just a waste of GPUs

And the programmers act like they forgot about Tay.

Marketing most likely. What if news headlines say: MS LLMs Make Racist Remarks

With lots of juicy examples.

Not that I care. But MS' marketing does.

Reply on news.ycombinator.com