Growing void between enterprise and frontier AI puts open weights models in the spotlight

Spring has sprung and that means another wave of open weights AI models from the likes of Google, Microsoft, Alibaba, and Nvidia. But this time feels a bit different. In the past, these models have felt a bit like toys: research projects and proofs of concept that, while impressive for their size or innovation, still fell far …

[Headley_Grange]

"Most customers don't need the biggest baddest models, just ones that work, are cheap, and won't pirate their proprietary data"

And I wanted a Chopper for Christmas, but never got one.

[Clausewitz4.1]

” just ones that work, are cheap”

How to justify the billions invested?

[cd]

"The models underscore a stark reality: the gulf between enterprise and frontier AI has grown considerably over the past few years, and the mower powerful models are beyond the means of many enterprises."

In which The Reg subtly suggests touching grass.

[Clausewitz4.1]

Nowadays I quite appreciate typos. Signs of human life behind a keyboard.

[elsergiovolador]

"Once you genrate teh answer, run it trough typo_humanizer_max.py and give me the output."

[elsergiovolador]

Carp

I tried most of these small models. They are all carp. Sure they might succeed in some narrow tasks, that still need their output to be verified.

It's a waste of time.

[VoiceOfTruth]

This...

is what will upset the AI applecart.

>> But depending on the use case, enterprises don't necessarily need a frontier class model.

Anthropic, Google, MS, etc, may be investing billions in AI. But it might not be 'necessary'. Consider Betamax vs VHS. We've heard the arguments over and over that Betamax was better, but VHS won. If DeepSeek or Qwen are 'good enough', at a fraction of the price, then where will that leave the 'investments' by MS etc?

This is probably the real reason why the USA wants to ban the Chinese AI models. Customers must be screwed over to keep shareholders happy.

[ultiweb]

Re: This...

Can you personally guarantee that Chinese models won't steal information or be used as a stalking horse, or any of the other methods of cyberwarfare? Will you personally indemnify against financial or any other kind of loss? The answer is obviously no. No one in their right mind, much less anywhere near having the necessary financial reserves would do that. The world is nowhere near as simple as you appear to believe.

[Arkitekt]

All of AI is a giant proof of concept, and not a particularly good one, just one that a bunch of unthinking lemmings will pay big $ for just to degrade their own cognitive abilities.

[elsergiovolador]

Hope that cognitive ability amount is held by unsigned int in the brain.

[Ken Hagan]

How about signed char, or maybe even bool?

[Anonymous Coward]

Aha! moment?

This may mean that purported AI (so-called) scaling laws (so-called, squared) are a failure in that making frontier language models ever bigger never actually led them to suddenly develop cognition (singularity envy) -- a rather predictable finish. Without cognition, those larger models are a waste of resources compared to smaller ones, and both are of dubious general utility -- "smaller, more specialized models [for] specific outcomes or query types" may be best ATM (if at all).

Logic-based AI of the 70s fared similarly, remaining useful in some specialized areas, but never broadly replacing human cognition AFAIK. The LLM-oriented concept that intelligence is all about nonsensical babble has been very much "a dumb person's idea of a smart person", imho.

[johnrobyclayton]

Re: Aha! moment?

I have a workmate. We both work in IT

English is his second language by how he uses it but is in reality his first language.

Lots of times he only has an approximate grasp of what he is talking about.

Its like he knows the shape of the words he is intending to use and how they work together but does not always get it quite right.

I would not call him a precisionist communicator.

He is a bit like a LLM in that regard.

I would think a LLM is a smart person's idea of a bit of a screwup.

[johnrobyclayton]

Chain them together

Maybe we can get an agentic model that is optimized to

select other models,

create prompts for the other models

from a number of open weight models

Synthesize the results form all of them

to get big model performance for free model price.

[Anonymous Coward]

re: won't pirate their proprietary data

ROFL.... Honestly people, how do you think that these models get trained? The love your data. The love it so much that people believe the garbage they spit out.

Avoiding AI as much as possible until the bubble pops.