building playful affordances for alien lifeworlds. talks about cognition, ecology, and maths


catball
@catball

If you feel like a little schadenfreude, here's an article about Google admitting that they won't be able to keep pace with open-source LMs that are doing more with less:

Which also mentioned LoRA regarding reducing the parameters needed for LMs:

Some quotes from the article I appreciated (emphasis theirs):

We have no secret sauce. Our best hope is to learn from and collaborate with what others are doing outside Google. We should prioritize enabling 3P integrations.

People will not pay for a restricted model when free, unrestricted alternatives are comparable in quality. We should consider where our value add really is.

Giant models are slowing us down. In the long run, the best models are the ones which can be iterated upon quickly. We should make small variants more than an afterthought, now that we know what is possible in the <20B parameter regime.

[...]

In many ways, this shouldn’t be a surprise to anyone.


ireneista
@ireneista

now we can get on with trying to avoid an entirely different kind of dystopia, the one where instead of megacorps deciding how this stuff gets used, it's not possible to prevent unethical uses of these models because there's no single point of control

this is a good thing, it's closer to our ideal world. you cannot have decentralized power structures while having the ability to prevent misuse of technology, because any mechanism which can prevent misuse can also prevent challenges to power - and will.

however, it creates a huge moral burden on the community that works with these models, to foster public conversations about where the ethical lines need to be and how we can hold each other accountable for them.

we look forward to that work. it is important. it is a core part of building the society we wish to live in.


You must log in to comment.

in reply to @catball's post:

thank goodness tbh, I was hoping this was true but there was always a chance it wasn't. I think if Google doesn't think their sauce is special, then literally no one's is. There's 0% chance in my mind that there's some ridiculous advantage that Google wouldn't be aware of.

To be fair, Damore wasn’t hired to be a huge prick. This is someone good enough at LLMs to be paid to do it by one of the best talent-scout organizations in the business. I see your salt and I add a teaspoon, maybe.

Agree @modulusshift !

I think the most compelling argument I've heard among folks in the field in favor of google / openAI is that they have already built integrations between their language models and other services.

I feel like the edge Google has is mainly in that they have a lot cash to push a lot of people after this, plus datacenters / hardware / infrastructure to support their pursuit

As research advances allowing LMs to operate on less params and with less overhead, and as open corpora become more robust, Google's hardware and cash advantage will become less prominent in the market, and the niche for LLMs becomes a more accessible commodity

My dream though is that everyone gets sick of LLM hype and realizes that putting a conversational agent in front of a service doesn't make the service any better (maybe even worse)

I guarantee that this is why the CEOs are in Congressional hearings talking about their fears of AI turning bad. They couldn't care less about "evil AI," because that's literally what a corporation is, a set of rules that diffuses responsibility for when people feel exploited. They care about making AI too expensive for competitors to run, now through legal compliance, since hardware is no longer an obstacle...

in reply to @ireneista's post:

I think the most striking observation in that analysis is that the big companies' huge data sets aren't an advantage: small, well-curated training sets are what produces high-quality LLMs.

Everything in these companies is focused around scale. And here is an arena in which scale (at least, the way they think about it) is a disadvantage.

The way scale could be an advantage is that these companies could staff up with teams experts who curate better high-quality data sets than the outside world. Except that's the exact opposite of how they look at the world.

Another implication of that observation is that there may well be an incentive for people building LLMs to be exceptionally mindful of bias in the training sets they're curating, and even a competitive advantage to teams who do this better.