[mft_]

Open models are probably also comparatively astronomically expensive to train - just less so than the frontier models because they’re somewhat smaller, +/- the creators are more incentivised to focus on getting more from less compute because they’re have to, +/- they rely on distillation of the frontier models and this is more efficient.

But efficiencies aside; creation of open models still requires a lot of money and compute from a large organisation which is willing to accept zero return for that spend. This largesse is unlikely to continue forever; so the question is which will crack first, the frontier models’ business model or the fast followers’ generosity?

[demosthanos]

Yes, the problem with comparing open models to open source is that open source requires humans to volunteer their time. Open models requires humans to volunteer their money.

These two types of contributions have very different behavioral profiles, and it doesn't obviously follow that the historical success of getting people to collaborate socially on building software for fun and for the benefit of the community will translate in any meaningful way to the necessity of being able to raise enormous amounts of money to pay for enormous amounts of electricity.

[enos_feedler]

The biggest hurdle is whether humans volunteer their expertise. Not time or money. We need top talent to make the open models. Sponsorship is plentiful. Open source volunteers are less critical with LLM doing the grunt work. Its about talent contributing to the open

[calebkaiser]

This is not my experience working in the field the last 8 years. There is not a dearth of talented researchers, engineers, etc. who are willing to contribute to open models. Just look at the ecosystem generally, from academia to industry. So much research still happens in the open, and the open source community in ML is still massive and energetic.

Training a new frontier model just costs a lot of money and there are very few companies for whom it makes sense to do so, and even fewer orgs who are just going to gift those kinds of resources to an open source initiative.

[hobofan]

> open source requires humans to volunteer their time

Your idealistic of open source may require that, but in practice a huge part of open source is commercial and a large chunk of that is low on collaboration (across vendor boundaries).

[connicpu]

Technically open source requires some amount of monetary volunteering, it's just that the electricity to run a code editor and compile (most) open source code bases is within hobby budget for most people.

[chaosharmonic]

I don't think it needs to be framed purely as generosity. You just need a sufficiently self-interested actor that sees open ecosystems as a necessary part of reducing their own risk profile, relative to the alternative of complete reliance of a third-party business that can take an exorbitant cut and/or Sherlock them at any time.

Valve and SteamOS are a good example of what this idea looks like in practice. (Though they may also illustrate a third thing you need: a privately-run company, that has enough profit, and enough commitment from leadership to the company's vision, that they can make long-term bets without having to eventually bow to investors seeking short-term gains.)

[mft_]

> You just need a sufficiently self-interested actor that sees open ecosystems as a necessary part of reducing their own risk profile, relative to the alternative of complete reliance of a third-party business that can take an exorbitant cut and/or Sherlock them at any time.

This would be an argument for an organisation developing its own model; but not per se for releasing the weights openly.

The possible explanations (I'm aware of, which overlap somewhat) for spending large amounts of money on models then releasing them for free (i.e. the current Chinese approach) are soft power, marketing for a future paid model business (i.e. competing with the US models for customers and mindshare during the time you can't compete directly at the bleeding edge), and/or a geopolitical move to diminish the value of the US's frontier model companies.

[chaosharmonic]

I acutally see the reduced burden that comes with actually sharing resources.

To extend the previous analogy: Valve didn't make desktop Linux viable on their own. A lot of it is owed to another self-interested actor -- Google -- through the reduced need for dedicated desktop apps (largely pushed by Chrome), and various enhancements to wireless and power management that were necessary to make it a viable mobile platform (directly benefiting Android/ChromeOS, but then spreading out to laptops and mobile devices in general, including handhelds like the Steam Deck).

You can see it on a smaller scale in ecosystems like Android -- where handset makers regularly contribute features from their UI skins upstream, so that they no longer need to spend engineering resources maintaining distinct versions of theming engines/notification badges/multiwindow/various other stuff.

On a related note this is an argument for open source models, not just open weights. I think a lot of the diminishing returns relate to the opaque nature of most models.

[afavour]

I’m not exactly sure on the “how” but it only makes logical sense for (non-AI) companies to band together to fund the training of a shared model. Apple is a great example, AI is not their core business but they still require it.

The only thing that took us down a different path is the vast sums of VC funding pumped into the AI companies.

[marcosdumay]

It probably doesn't.

There's a reason we let companies specialize in some kind of service and buy it from them.

LLMs aren't looking like they'll be highly differentiated like software, so their market will probably be competitive. What negates the main reason Open Source software exists.

[munk-a]

LLM training doesn't carry the same NIH risks that normal internal software bloat does. They are relatively simple to setup training for and analysis of accuracy/recall can be automated.

This leaves the price differential between a private third party and an internal initiative as barely more than the cost to train the model[1] - perhaps that's where we'll end up, a centrally trained model will represent an economy of scale that can leverage that difference into a margin it can profit off of but your business being purely profit driven by that training expenditure seems like a ridiculously thin margin.

So where does that leave the AI companies? If their LLMs are off the shelf-once built products they have a strong advantage for casual low usage but enterprise customers will have a huge cost incentive to roll their own - if the LLMs require continuous retraining and the frontier keeps moving then enterprise customers will find a packaged service more attractive and likely continue to subscribe for more accuracy but casual low usage will likely shift towards "good enough" models. It seems inevitable that they'll lose half the market and it seems difficult to discern their long term profitability[2].

1. Costs can, I think, reasonably be reduced to hardware depreciation and energy - if trends continue with cloud resource availability (it's possible this won't be the case if large compute providers start pulling resources offline to build a moat but I think they'd likely prefer the reliable compute income over model income which has several other competitive weaknesses). Hardware depreciation would normally be pretty negligible and equal across different training entities, right now we have a chip shortage but given the demand that can't last too long so I'd consider hardware to be fungible - and energy is entirely fungible - they're both hard to moat.

2. Outside of AGI, who knows if AGI will be or what even counts for it at this point - but I think if AGI isn't a doomsday scenario we fall back to one of the two above scenarios - either the frontier is ever moving and they can retain enterprise customers or the frontier seizes up and everyone can just use an off the shelf offering. In either scenario they don't have a lot of moat to deal with for their products unless they can restrict compute which is why Alphabet, AWS and MSFT are the only players I could see realistically coming out of this as an AI vendor winner and I'm not even certain if it'd be a good idea for them if it'd hamstring their cloud profitability.

[inigyou]

If not for VC-funded LLMs there wouldn't be any LLMs.

[LunaSea]

Most of the innovations needed for LLMs came from people at Google.

[munk-a]

A fair amount of ML/AI innovations came out of the market in general. Neural networks are a useful tool to solve a variety of problems... LLMs specifically were a more recent interesting market to develop but I've yet to see anything that could give a market player a real competitive advantage. It feels like we just invented a new hammer and now that we know how to build it it isn't that hard to build one yourself. The all purpose hammers are, of course, unreasonable to build - but those don't seem to be that useful. I don't really need Claude to be able to generate sonnets when I'm programming so I think specialization is the place we'll see genuine markets form.

[echelon]

> came from people at Google

Who had to leave to build anything.

[inigyou]

So VC-funded.

[mrandish]

The company was VC-funded as a search engine but by the time they made significant investments in AI (DeepMind etc) they'd been a publicly held company earning multiple billions a year from advertising for a decade.

[inigyou]

Google is a VC. Their side projects are VC projects.

[afavour]

[citation needed]

Historically speaking a lot of inventions have come about without things like VC investment. Either way, there’s probably little point in debating it, just because VC funded companies control the market now doesn’t mean they should indefinitely.

[sph]

How does it work if people flock to open models but they're too expensive to train? What is the financial incentive to do so?

I seem to understand open models are mostly coming from China, and the benefit of training and releasing them for 'free' is a powerful geopolitical weapon against the Western/US economy that at this point depends on OpenAI & co. to succeed.

Will the West make open models illegal?

[echelon]

> Will the West make open models illegal?

We better not.

> What is the financial incentive to do so?

If we'd been sharing all along (as we should have been), we probably would have gotten even further along in the development of the tech.

Think of everything we could do if every researcher on the planet had first class access to the frontier. No academic fallback models. No crude API access. No limits, but direct access to the weights and the ability to lobotomize, splice, and dice.

We could pour intelligence from one container to the next without paying a tax or wearing a blindfold. All without spilling a drop.

*Open* *Must* *Win*

[inigyou]

> No X. No Y. No Z, but Q

[Danox]

That would be strike one like putting tariffs on every country you disagree with, such a strategy would fail.

[surgical_fire]

If by West you mean the USA, maybe.

Other countries in the westen hemisphere, probably not.

[LorenDB]

"You wouldn't download an LLM"

[sph]

You wouldn't crash the stock market by preferring Chinese models.

[mcv]

I absolutely would. Let the bubble burst.

[dominotw]

so using glm will crash my stocks and make me poorer? damn. thats quite a high price to pay for use then

[Danox]

The bubble would burst for OpenAI, Anthropic, Oracle, and Microsoft…

[mcv]

It'd better happen sooner than later. This bubble is distorting everything and turning many executives insane. See https://ludic.mataroa.blog/blog/ai-mania-is-eviscerating-glo...

[mrandish]

> releasing them for 'free' is a powerful geopolitical weapon...

I agree that, currently, the Chinese govt is not only allowing but tacitly encouraging open weight model releases. However, I don't see it as an attack. I think it's more of a strategic delaying move to slow the revenue to frontier models while China works to catch up. This strategy will likely change over time.

> Will the West make open models illegal?

In the U.S. this seems highly unlikely due to the current administration's generally laissez-faire approach to tech as well as the U.S. constitution severely limiting the government's latitude to constrain economic activity.

As we saw with the temporary Mythos restriction, there are legal mechanisms to limit tech on certain grounds, but over time such limits are subject to close judicial and constitutional review. The Mythos embargo was also likely driven in part by the administration's anger at Anthropic for choosing to block the DoD from using their products for mass domestic surveillance and warfighting. I doubt we'll see any meaningful restrictions on OAI or other large companies. It'll be nearly 3 years before a different admin is in office and could enact serious limits and by then it will be too late for fundamental bans.

There are vested interests in most governments, such as intelligence agencies, law enforcement and the military, who would prefer to restrict some AI from broad use. As we saw with strong encryption, they'll only be able to delay and constrain, not stop, such a broadly useful dual-use tech. The geopolitical, economic, competitive and civil liberty interests are similar between strong encryption and AI, setting up a similar game theory dynamic. While it can be argued AI poses some potential danger, the specter of any such threat is abstract and not immediate.

On the other hand, the tech is obviously too economically essential and competitively vital to risk 'falling behind'. While there will certainly be attempts to ban, limit or constrain AI, the well-funded, highly organized commercial interests and civil libertarians will deploy lobbying, legal challenges and public opinion to ultimately prevail.

[bilbo0s]

In the U.S. this seems highly unlikely

Aren't these the same guys who won't even let us have Chinese cars?

I'm not as confident as you that they will keep allowing us access to technology as strategic as AI models out of China and elsewhere that undercut US models in the market.

To everyone reading, download open models from anywhere as soon as they are released. You really have no guarantee at all that access to those models won't be cut off in the future with the stroke of a President's pen. Those downloads are your insurance policy. You'll always be able to access whatever you've already downloaded.

[figassis]

I would definitely pay a monthly subscription to help fund a non profit compete with Anthropic and OpenAI. I already pay subscriptions for myself and 2+ other people. It's a non brainer to be able to pay for the training of better models that I can then run myself for many more. I hope someone starts this, I think this model would work. I'd start it today if I had the team and initial capital to bootstrap the infr. I know VCs won't fund it, but we definitely will, enthusiastically and continually.

[nijave]

My (unverified) AI research claimed generally Chinese models are cheaper to train because Chinese data scientists are cheaper to hire and they're also under more pressure to optimize training cost due to limited hardware availability

Seemed believable but not sure where that's true

Chinese AI companies are generally smaller tho and the models they're releasing are also smaller (I think estimates put OpenAI and Anthropic SOTA into trillions of parameters)

[ubercore]

I can't imagine even with the crazy salaries at frontier labs, staff costs make much of a difference.

[holoduke]

In the US and Europe its extremely expensive indeed. In China it is much lower and expected to be much lower in the future when China can bake their own high end chips. You know China is focussed on optimizing production and manufacturing. Unlike the west that primarily focus on regulations that make things super expensive.

[Danox]

Engineers at the top of the pyramid in China, Lawyers on top of the pyramid in America.

[slashdave]

Except: distillation.

It is not so strange that models from certain organizations are always a few months behind.