The next wave of innovation seems like it’ll be much less centralized into a handful of big companies. Even more powerful closed door LLMs will appear, but already so much utility out of these that can run on consumer hardware.
Can’t build anything significant with leaked models.
Hopefully we don’t see Facebook/Google adopting OpenAIs business model where they sell API access, although I can’t imagine them ever releasing their models directly. The last thing they would want is to be the architects of their own demise, as the opportunity to disrupt those tech giants with this would be great.
I do agree that large orgs are likely to have the best models, but a lot of the low hanging fruit/initial utility is already covered by models that run on consumer hardware
EDIT: You seem to have deleted your response but I am posting my reply anyway:
Microsoft has mentioned today that it cost them many hundreds of millions of USD to accommodate OpenAI models. I studied NLP at Stanford including LLMs. Only Google had a smaller transformer that could somewhat compete (Reformer) but the rest went with the brute force GPT-like models and 1000+ GPUs, $10M+/training run etc. I haven't seen a single post anywhere explaining how to train such large models (a smaller model on 4-8 GPUs doesn't count). "Facts are clear"
It's a widely known meme at this point, but to reiterate: For a popular large model, the largest part of the cost will be spent on inference, not on training. If we assume inference on end user device, this cost disappears.
And even if you have the million to rent a cluster, there is a very deep question of the optimal architecture, dataset and hyperparameters to train the best model possible under given constraints.
Quotes from the paper: Our model is trained using a codebase that builds on Megatron (Shoeybi et al., 2020) and DeepSpeed (Rasley et al., 2020) to facilitate efficient and straightforward training of large language models with tens of billions of parameters. We use the official PyTorch v1.10.0 release binary package compiled with CUDA 11.1. This package is bundled with NCCL 2.10.3 for distributed communications.
We trained GPT-NeoX-20B on twelve Supermicro AS-4124GO-NART servers, each with eight NVIDIA A100-SXM4-40GB GPUs and configured with two AMD EPYC 7532 CPUs. All GPUs can directly access the InfiniBand switched fabric through one of four ConnectX-6 HCAs for GPUDirect RDMA. Two NVIDIA MQM8700-HS2R switches—connected by 16 links—compose the spine of this InfiniBand network, with one link per node CPU socket connected to each switch.
And if you are interested in 176B-scale training, read the BLOOM-176B and OPT-175B papers and research logs.
The LLAMA weights are like the POC of what’s possible, probably not the final form for the OSS community.
All you can see running in the wild are quantized LLaMA variants (4 or 8-bit) whereas the original model is 32-bit.
The field is moving so quick.
And we know how open source eventually took over pretty much all code. I expect this to be the same, but instead of four decades, it will be four months
It's huge, 160B parameters. It's open, it's free, it's been out there for a while.
Is something wrong with it? Using outdated training data? Using outdated training techniques? Or is it just not well known enough -- ie, not in the news at the right time.
And OpenAI's known models (at least the ones for which the architecture is publicly known) https://platform.openai.com/docs/model-index-for-researchers aren't pinnacles of parameter efficiency either: Deepmind's Chinchilla and LLaMA beat them on benchmarks while being much smaller.
Just an opinion, but if we truly aim for "green computing" and carbon footprint reduction, we should pay a lot of attention to the parameter efficiency of the models that get run in our datacenters - and we should invest money into public availability of the most efficient foundation models.
In other words every carbon footprint reduction optimisation is used to yield more efficient output with the same (all available) carbon footprint.
On the flip side, it overperforms it on every other language. It seems clear to me that it is more useful for popular applications (where a large userbase will use their own tongue).
It also has a 7B version. Although it is much weaker than LLaMA, it could be pretrained to match, especially through distillation from LLaMA. By now, it is clear that even Chinchilla is underestimating how much we gain from training more: training is more efficient thanks to new techniques, and even in the paper, some lines are stopped early before reaching Pareto (likely because it ran out of tokens, but LLaMA makes it seem that >1 epoch training is fine).
Llama performs better on benchmarks while a fraction of the size.
[1]: https://news.ycombinator.com/item?id=32079558
With all the appropriation going on from both commercial and academic organisations I think we need a new name for these kinds of models. Currently I am leaning towards “restricted”, as they are neither open, nor proprietary.
If it's behind a SaaS, I think you can.
Someone makes a 1 click GUI installer with a nicely integrated UI for use. No command line use at all.
We’ll get model and weight distributors for updated knowledge.