Hacker News new | past | comments | ask | show | jobs | submit login

Google can look up into their index and can remove whatever they want to, within minutes. But how that can be possible for an LLM? That is, "decontaminate" the model from certain parts of the corups? I can only think of excluding the data set from the training and then retrain?

As a side note, I think LLM frenzy would be dead in few years, 10 years time frame at max. The rent seeking on these LLMs as of today would no more be a viable or as profitable business model as more inference circuitry gets out in the wild into laptops and phones, more models get released, tweaked by the community and such.

People thinking to downvote and dismiss this should see the history of commercial Unix and how that turned out to be today and how almost no workload (other than CAD, Graphics) runs on Windows or Unix including this very forum, I highly doubt is hosted on Windows or a commercial variant of Unix.




> almost no workload (other than CAD, Graphics) runs on Windows or Unix including this very forum

About a fifth to a quarter of public-facing Web servers are Windows Server. Most famously, Stack Overflow[1].

[1]: https://meta.stackexchange.com/a/10370/1424704


> About a fifth to a quarter of public-facing Web servers are Windows Server

Got a link for that? Best I can find is 5% of all websites: https://www.netcraft.com/blog/may-2023-web-server-survey/


20% of workloads running on Windows should result in corresponding number of jobs as well but that's not what I see.

Most companies are writing software with software developed on Linux first and for Linux first (or Unix) and later ported to Windows as an after thought. I'm thinking Python, Ruby, NodeJS, Rust, Go, Java, PHP but not seeing as much of C#/ASP.NET which should at least be 20% of the market?

Only two explanations - either I am in a social bubble so don't have exposure or writing software for Windows is so much easy that it takes five times less engineering muscle.


There are plenty of .NET jobs, and .NET (Core, particularly) is really easy to write.

That said, I'd guess the difference is that the startup and big tech world (i.e., "software companies") like our fancy stacks, but non-software companies prefer stability and familiarity. It makes way more sense for most companies to have a 3-man "bespoke software" department (sys/db admin, sr engineer, jr engineer) on a stack supported by a big company (Microsoft) where most of the work is maintenance and the position lasts an entire career. It's a big enough team to support most small to middling businesses, but not so big that the push to rewrite everything in [language/framework of the week] gains traction.

The practical conclusion is that these companies have few spots to fill, and they probably don't advertise where you're looking.


>>either I am in a social bubble so don't have exposure or writing software for Windows

you clearly are.... There are TONS of windows only software out there, and most INTERNAL systems that run companies, these internal LOB apps, often custom made for the companies, many many many of them (probally more than 50%) are windows server apps.

For example GE makes a Huge Industrial ecosystem of applications that runs a ton of factories, utilities, and other companies... Guess what all of that is windows based.

Many of the biggest ERP's run on MS SQL Server which until very recently was Windows Only, and most MS SQL Servers are still on windows server

To claim only 20% of all workloads are Windows shows an extreme bubble most likely in the realm of WEB BASED DEVELOPMENT, as highlighted by list of web technologies, php, node, etc..


.NET is huge in banking, iGaming, traditional industries. Python/PHP are kinda outliers found here and there. JS is eating both Java and .NET's lunch and ofc frontend.


But, wasn’t the reason proprietary unixes died out at major work horses because of a nearly feature comparable free alternative (Linux)?

Extending the analogy, LLMs won’t die out, just proprietary ones. (Which is where I think this tech will actually go anyway.)


LLMs won't die out but proprietary LLMs behind APIs might not have valuations of hundreds of billions of dollars.

Crowd source, crowed trained (distributed training) fast enough, good enough generative models that are updated (and downloadable) every few months would start to erode the subscriber base gradually.

I might be very very wrong here but it seems like so from where I see it.


> But how that can be possible for an LLM?

Well, it seems to me that's part of the problem here.

And it's their problem, one they created for themselves by just assuming they could safely take absolutely every bit of data they could get their hands on to train their models.



The argument may be that having very large models that everyone uses is a bad idea, and that companies and even individuals should instead be empowered to create their own smaller models, trained on data they trust. This will only become more feasible as the technology progresses.


Windows and MacOS (and their closed source derivatives) are probably at least as large as Linux, even including all the servers Linux is deployed on. Proprietary UNIX did not "die out"; Apple sells about a quarter million of them every year.

The majority of the world's computing systems runs on closed source software. Believing the opposite is bubble-thinking. Its not just Windows/MacOS. Most Android distros are not actually open source. Power control systems. Traffic control systems. Networking hardware. Even the underlying operating systems which power the VMs you run on AWS that are technically open source. The billions of little computers that form together to make the modern world work; they're mostly closed source.


Google have their "Machine Unlearning" challenge to address this specific issue - removing the influence of given training data without retraining from scratch. Seems like a hard problem. https://blog.research.google/2023/06/announcing-first-machin...


> But how that can be possible for an LLM?

They should have thought of that before they went ahead and trained on whatever they could get.

Image models are going to have similar problems, even if they win on copyright there's still CSAM in there: https://www.theregister.com/2023/12/20/csam_laion_dataset/


People thinking to dismiss this should, period. Consider that Open AI and similar companies are the only ones in the AI space with the market cap to build out profitable hardware projects which open source can't. Or maybe every investor is just dumb and likes throwing millions of dollars away so they can participate in a hype train.


Unix kinda still does the same thing now as before.

Future big ai models might be totally different in quality, and latency.


maybe they should build a better LLM? maybe they could ask the AI to make a better system. after all, tech and ai is so powerful that they could do virtually anything, except having accountability as it turns out.




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: