Hacker Newsnew | past | comments | ask | show | jobs | submit | dada216's commentslogin

Should you run a warehouse without forklifts?

My only issue about LLMs doing scientific advancements is that the actual models that did this and the training process are industrial secrets and not open weight releases.

I don't want the cyberpunk dystopia, I want the star trek utopia, that's where my attention is focusing on.


Star Trek is fiction. The role of super-advanced and sentient computers in Star Trek is deliberately kept limited in scope, so as to advance Roddenberry's belief in the unshakable human spirit and noble human yearning for discovery or whatever. Whenever machines are presented as capable of supplanting or exceeding humanity, as in the TOS episode "The Ultimate Computer", or anything involving Lore or the Borg in TNG, it's universally depicted as existentially threatening.

The access to those models is not really open to anyone though.

Through wars and revolutions. Imagine a few generations detached from us.

I'm a millennial, Im starting to think I need to write down for those future generations how it felt to be raised in a mostly analog world and die in a world were LLMs achieved what they just did and what they might still.


If the models that were used to do this were open, if we were already living, or at least have a decent plan to build, a society where all the scientific advancements that can be built with such models would actually improve people's life I would be extremely joyful.

I am terrified we are already living in the cyberpunk dystopia


Even with the model open, you can't run it. When we read some of these break throughs, they use 10,000 agents running on 20,000 GPUs for days. It still comes down to wealth, those with more capital will have an edge. You might not have the compute to run such a model or if you do have the compute, it will be slow and require tons of energy.

The valuable models are kept in house.

The governance model of the Anthropic IPO gives absolute power forever to the founders.

They claim to strive towards the good of humanity.

The silent part is that they will be the ones deciding what good means for the rest of us.

This will end in tears.


ITT it's that huge max output token of 1mln tokens.

It's pretty far beyond other models (Opus 5.5 is at 128k).


I do :) alongside cloud of course, hybrid is the way to go in my view.

I run a bunch of openshift clusters for very private container workloads and VMs.

You can't beat a local fiber channel SAN for speed and reliability, especially compared to cloud software defined storage, and even on the networking side of things if you worked out your local connectivity well using real networking vs cloud providers software defined networks is so much better.

CSIs are hit and miss, some vendors are really really good, some are meh (hitachi, even though they are improving a lot with the latest relases), but the moment you get an IBM nvme san and start to tune all layers you get performance that are out of this world, some of our dbms workloads are very io sensitive and they improved 10x with nvme fabric.


In this specific case there is quite a simple solution, google (and oai/anthropic as well) should share revenue with the website they scrape to provide the ai summaries, they literally did this with adsense and youtube!

Sure, there's some technical hurdles, like identifying who's actually using the info they read, but impressions are a thing in ad networks, then there's broader questions about evaluating what training data was useful and used to ultimately produce what token, I'm still just taking about the kind of ai summaries in search related to the article, seems doable with the same system we use for ad networks.

When it comes to coding and more generalized knowledge there's already been interpretations about copyright laws and fair use that seem to allow what oai and ant are doing and how they are capitalizing on it, doesn't mean we should rethink them.

I'm sure there are technical solutions that can be translated to this sort of issues as well that could be derived from ad networks, I mean, imagine a world where your code and your blog post get a score on their usefulness to train LLMs and you get a share of the revenue, wouldn't this be better for all players? they get good training data, we get paid for it.


> should share revenue with the website they scrape to provide the ai summaries, they literally did this with adsense and youtube!

https://x402.org can facilitate that, we just need forcing the Big AI to pay out.


They would quickly have zero traffic if they demanded payment.

Cloudflare experimented with this model as well.

> In this specific case there is quite a simple solution, google (and oai/anthropic as well) should share revenue with the website they scrape to provide the ai summaries

The actual problem here is so weird that this doesn't actually work.

Consider something like Stack Overflow or Reddit. The people answering questions weren't getting paid to begin with, so you're solving the wrong problem.

And weirdly, the thing the AI summaries are doing is the thing sites like that have long been trying to do themselves. If someone has a common question that has already been answered, the AI summary answers it for them and then they don't show up to post the 5000th duplicate of the same question. It even lowers the operating costs of the site since then 5000 people don't have to load the same page when they don't need to post anything new.

Whereas if they have a question that hasn't been answered, there is nothing for the AI summary to summarize, so then the user isn't satisfied with it and proceeds to actually visit the site and post the novel question that hasn't been answered yet. In some kind of theoretical sense that should work. The site only gets used for new questions.

But it actually doesn't work, because it breaks an important part of how sites like that function, which is that the users vote on the answers. When you have the common question, you don't need to post anything new, but you upvote the correct answer, which gives the person answering the question a sense of accomplishment and lets them know that people value the thing they spent the time to write. And if you try someone's answer and it doesn't work then it gets downvoted and the low score lets other users know to be skeptical of it.

If the AI summary shows the answer without the feedback mechanism then it breaks the site. The people who post good answers then can't see anyone interacting with them and feel like they're wasting their time, and if that happens the site is dead.

The question is, how do you fix that when "just show me the answer immediately" is the thing the user of the search engine actually wants?


Good points.

Maybe there's technical solutions?

For sites like the one in the article the feedback loop and payment it's there with revenue sharing.

For web 2.0 (remember this term? Haven't used it in a while) you're right, the community building part is hampered by the LLM middle layer, but as you said if you're going past "just show me the answer immediately" and actually click on the site you may start to participate and this could actually improve the quality of contributors?

I kinda hate it but the "marketing funnel" is a valid model to apply here, and I'm sure there is an equilibrium to be found which benefits everyone, regardless of size and the kind of platform involved, history is not on my side here, but I need to be optimistic to keep being sane.


This was a hit on him, the suspicion I have is that it's his position on Open Weight models that's problematic for other players.

I don't care about stock values, I do care about open weight releases to be the economics formula the LLM industry adopts


It can't be a hit on him when he put himself in this position in the first place.

Do you think he's the only CEO that has affairs? The hit is surfacing them. He's also not the only one who wanted to send GPUs to China and Nvidia didn't skirted the export ban on their own.

It could be both.

Sure, but he needs to have put himself in the position to begin with. I find so much conversation about this strips him of agency and responsibility, like he's a victim somehow.

That's what "both" means. Someone wanted to take him out. His personal failings gave them the means to do it.

Bedsides coding agents, which only work extremely well if you are a developers or you train a person with the right inclination I have plenty running that work quite well, they do data extraction and data entry from emails and documents. I run them with Chinese and other open models on a bunch of H100, they cover tens of thousands of emails per month and as many documents.


My 2 cents: it would be beneficial to involve companies like RH and the teams behind many distros to work through those, it's not just about validating and patching but also about delivering them to the systems involved, a lot of people need to be involved, hence collaboration is required


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

Search: