I’m struggling to understand from this paper whether the approach is better in the general sense (all cases, with wider models seeing greater benefits) or purely for wider models (with narrower models seeing detriment)?
If it’s the former this could effectively halve finetuning cost overnight which would go a significant way towards enabling a wider array of use cases for LoRA.
Don’t we already have open source achieving temporal consistency up to 15 seconds? I might be misreading the article but the main achievement in Sora is consistency over 30min+ generations
There are a lot of places e.g. Replicate where you can finetune and deploy language models. You’ll need GPUs, but you can simply select a hardware class and pay per second of usage, similar to AWS Lambda or similar serverless infrastructure.
Serverless AI is quickly becoming popular precisely because of the scenario you’re describing — it’s currently still pretty hard to deploy your own GPU stack, not to mention crazy expensive to run eg an A100 24/7, plus orchestration for scale up/down. It’s why so many model authors don’t host their own demos anymore and simply toss it on HuggingFace or Replicate directly.
Serverless providers will basically do the infra for you, as well as make necessary GPU reservations, then charge you a slight premium on the reserved price — so you’d pay less than on-demand on GCP/AWS, while they benefit from economies of scale.
I do imagine at some point soon GPUs will become cheap and more commercially available so you could away with hosting your own VPS in the distant (or near?) future.
> Software Procurement by Federal standards is relatively straightforward
> FedRamp and FIPS compliance
It’s odd to see these in the same sentence. FedRAMP is so insanely complex/difficult to achieve in a straightforward way. Even by your own estimate for a series E startup (with lots of capital and the ability to spend >18 months< on compliance) there’s a 3M$ variation in cost.
That rules out every startup or SME in software and that’s why you have Palantir, half baked tech that rarely delivers/is somehow more universally hated in USG than ServiceNow. Yet able to seize the space and hike prices endlessly due to compliance being so difficult to achieve — they realize/accept this as their edge as well and it’s why they so aggressively pursued IL6.
The good news is that this is going away and USG is strongly reconsidering its approach here. CMMC, imo, is a huge step in the right direction.
Tiptap is immensely popular, supports both Vue and React, is super lightweight and has a huge community of extensions. I've been using this since 2019 and can't recommend it enough. A number of comments here are asking for functionality that is built in, easily available as an extension, or easily add-able!
Esri is an anticompetitive company leeching taxpayer dollars that only exists because ArcGIS trapped the government before many govcon laws preventing vendor lock in were put in place.
We would have 10x the mapping capabilities and much higher velocity on defense tech without this parasitic company.
Btw, also recently fined for blatant sexism against hundreds of female employees.
Yeah, it would be way better if they just released it right away, so that political campaigns can use AI generated videos of their opponents doing horrible/stupid things right before an election and before any of the general public has any idea that fake videos could be this realistic.
you joke, but the hobbling of these 'safe' models is exactly what spurs development of the unsafe ones that are ran locally, anonymously, and for who knows what purpose.
someone really interested in control would want OpenAI or whatever centralized organization to be able to sift through the results for dangerous individuals -- part of this is making sure to stymie development of alternatives to that concept.
[1] https://huggingface.co/models?sort=trending
[2] https://hype.replicate.dev/
[3] https://github.com/trending
reply