> The Olivia system is an HPE Cray Supercomputing EX system, with 448 GPUs and 64,512 CPU cores.
Training a sovereign LLM with this meager hardware as opposed to a LORA on some open source model seems like a huge mistake and a potential red flag.
There is no way these people have the resources to train a fully fledged LLM, so claiming that is their goal makes me think they don't intend for the LLM to be useful.
Which begs the question, whose money are they wasting - and why?
They successfully have made PoC finetunes before, so the next step is training fully fledged LLMs.
I don’t think they aim to anything worthwhile. The finetunes were incredibly broken. I’m guessing it’s more about having the method to do it. I’m not convinced it’s super useful but I’m not one to decide who gets to do what with the research funds.
One finetune I tried did make fun of humans expressing their feelings in the chat. Often.
One other finetune did hallucinate that it was a doctor and my baby had terrible diseases, every time I just wrote "hei" (with a generic neutral system prompt that likely triggered this behaviour though).
I think Olivia is big enough for what it’s used for. In my opinion it’s better to stay up to date and not waste too much money on hardware at the moment.
It may not be useful to anyone outside, but it's possible that one of the goals is institutional learning (that is, embedding the knowledge in how to build LLMs in an organization).
Even though it's nominally the national library behind this, they were probably chosen (as per the article) because they legally own and can use all NO material for this end. I'd guess researchers from related entities like unis will be involved in the process.
The largest problem is available training data actually.
They have already done experiments with dittrent sub 10b models with both fine-tuning and fully from scratch. And last I check the fully from scratch captured the language in a better way.
DeepSeek claims to have trained on something like 2k H800, this is ~0.5k GH200 … it’s not nothing. Sure they’re not going to _serve_ it at scale, but that’s not the point?
Also the line between “finetuning a base model” and “man this is a real good initialization” gets pretty blurry at scale.
And this is before anybody ever thought about optimizing the training process. (Currently it's just pytorch analyst-as-coder slop, with extremely overprovisioned quantizations, etc.)
I'm a Norwegian, and I use the national library almost every day for searching through texts. They have truly one of the best working user interfaces (and functionality) for searching through the massive amounts of text.
As a Norwegian this sounds like a mistake. Who will use this LLM? Where? For what? The underlying data could be made more easily searchable and digestible for agents in general if the goal is better knowledge of Norwegian culture.
Exactly, if there's one thing transformers are good at it's translation. One I've found particularly nice: any question ChatGPT can answer in English it can answer in French. I'm assuming Norwegian too. So there's no point.
The point is that norway willl have its own LLM. And will not have dependencies to another state or private company. The goal is not to be the best model. But to have a model that include more Norwegian data then other LLM and that it's not screwed against other sources.
>As Husnes put it; Norway is a small country solving a problem every non-English-speaking nation will face: how do you build AI that reflects your language, your culture and your history? AI needs custodians, not just builders.
I'm afraid the answer is, mostly you don't.
Such a thing requires strong political will that, at least in my environment, seems basically impossible to align.
The costs are prohibitive, but beyond that, the type of person who cares about local representation like that is either completely fine with letting foreign companies implement it (after all, you can use ChatGPT in Basque if you want to) or is against the idea of AI altogether.
I'm sure if Norway approached the American labs with goal of making a curated datasets for training, they would absolutely get in the training door, and those models would likely run circles around anything that could be domestically done.
That being said though, I can feel you cringing through the screen.
This can’t be right. 2 PB of flash is like $200k. It’s within reach of many individuals. Then again I guess you don’t need that much storage so maybe it is.
Boy pricing is pretty nuts these days. I have half a petabyte in Seagate enterprise drives myself and I didn’t pay anything close to that to acquire it. Such a pity about the flash storage. 2 years ago we built 200 TiB or something of flash using Samsung PM1633 or something and it was a fraction of the cost per gigabyte that $1m would imply.
> He asserted that any country with its own language that did not have a sovereign LLM trained in that language was at a disadvantage as a globally trained, English-speaking LLM would not know about that country’s history, news and culture that was described in the local language.
I don’t know this is true. But whatever sounds true enough and gets funding seems to be what flies these days.
Can confirm that. Norway may have a small population, but if you live there you'll think it's truly the center of the world (aside from the US. Norwegians love America)
5x 400gbit running to a 2U box whoa, the PCI lanes must have heat shielding.
More seriously there is a sensibility limit on extreme density where it's not needed. The idea that you're just going to magically get 2 TBit/s out of those ports seems unlikely even with tweaked software, and you're stuck with a power and comms hotspot that's liable to dictate the remainder of your network design.
At max utilisation that 2U would take 12 hours to drain, and only 12 hours assuming peak and likely unachievable throughput and the box otherwise being completely out of service. Not a great start
if you read the article 2pb is available as flash storage in the data pipeline, used to dedupe, clean, normalize, etc, for training from 60pb of raw data.
Training a sovereign LLM with this meager hardware as opposed to a LORA on some open source model seems like a huge mistake and a potential red flag.
There is no way these people have the resources to train a fully fledged LLM, so claiming that is their goal makes me think they don't intend for the LLM to be useful.
Which begs the question, whose money are they wasting - and why?
I don’t think they aim to anything worthwhile. The finetunes were incredibly broken. I’m guessing it’s more about having the method to do it. I’m not convinced it’s super useful but I’m not one to decide who gets to do what with the research funds.
One finetune I tried did make fun of humans expressing their feelings in the chat. Often.
One other finetune did hallucinate that it was a doctor and my baby had terrible diseases, every time I just wrote "hei" (with a generic neutral system prompt that likely triggered this behaviour though).
I think Olivia is big enough for what it’s used for. In my opinion it’s better to stay up to date and not waste too much money on hardware at the moment.
Even though it's nominally the national library behind this, they were probably chosen (as per the article) because they legally own and can use all NO material for this end. I'd guess researchers from related entities like unis will be involved in the process.
They have already done experiments with dittrent sub 10b models with both fine-tuning and fully from scratch. And last I check the fully from scratch captured the language in a better way.
What do you suggest, that they stop and wait until they have the right HW?
Also the line between “finetuning a base model” and “man this is a real good initialization” gets pretty blurry at scale.
Altogether a pretty presumptuous take.
Qwen was made on a cluster about that size.
And this is before anybody ever thought about optimizing the training process. (Currently it's just pytorch analyst-as-coder slop, with extremely overprovisioned quantizations, etc.)
Seems like making the frontier models know Norwegian and their culture is a better (or additional!) way to reach the end they are going for here.
I'm afraid the answer is, mostly you don't.
Such a thing requires strong political will that, at least in my environment, seems basically impossible to align.
The costs are prohibitive, but beyond that, the type of person who cares about local representation like that is either completely fine with letting foreign companies implement it (after all, you can use ChatGPT in Basque if you want to) or is against the idea of AI altogether.
That being said though, I can feel you cringing through the screen.
If you go with HDD arrays probably $50k
I don’t know this is true. But whatever sounds true enough and gets funding seems to be what flies these days.
Dell just launched a 2U that fits almost 10 petabytes in it. It's probably not 384 core capable but that is very doable right now, Epyc chips are 192 cores each! https://www.techradar.com/pro/dell-launches-record-shatterin...
More seriously there is a sensibility limit on extreme density where it's not needed. The idea that you're just going to magically get 2 TBit/s out of those ports seems unlikely even with tweaked software, and you're stuck with a power and comms hotspot that's liable to dictate the remainder of your network design.
At max utilisation that 2U would take 12 hours to drain, and only 12 hours assuming peak and likely unachievable throughput and the box otherwise being completely out of service. Not a great start