AI Judge changed title after evaluation, original title: "Pleias CTO Pierre-Carl Langlais warns the EU's new EUROPA model will be obsolete due to its two-year timeline"
The project won 5.15 million GH200 supercomputing hours.
So another mid vintage LLM in two years no one will use. Great.
Europe has the talent, infrastructure and industrial capacity to build advanced AI systems.
We've chosen the EUROPA consortium to build an open-source frontier AI model in all 24 official EU languages↓
Many users reacted negatively to the EU selecting the EUROPA Consortium for an open-source frontier AI model, calling the effort underfunded, unrealistic in scale, and undermined by incompetence or soft corruption.
No Digg Deeper questions have been answered for this story yet.
"Come on, EU, when are you going to make real chaton fat"
"We have real chaton fat at home"
So another mid vintage LLM in two years no one will use. Great.
@Dorialexander > The selected project will receive up to 2.5% of the overall EuroHPC computing capacity for one year
And just took a quick skim, Domyn seems to do "take nemotron, run on "eu enterprise nice data" whatever that means?
So another mid vintage LLM in two years no one will use. Great.
You have to imagine Sisyphus happy (from latest blog)
So another mid vintage LLM in two years no one will use. Great.
Europe launching multiple competing AI projects with overlapping goals which compete with one another for access to limited public compute
Europe has the talent, infrastructure and industrial capacity to build advanced AI systems.
We've chosen the EUROPA consortium to build an open-source frontier AI model in all 24 official EU languages↓
@giffmana And optimizing somehow only for MMLU.
@Dorialexander > The selected project will receive up to 2.5% of the overall EuroHPC computing capacity for one year
And just took a quick skim, Domyn seems to do "take nemotron, run on "eu enterprise nice data" whatever that means?
@Dorialexander I also got curious to see what happened with the 2024 grants, and it seems to be a mixed bag:
@Dorialexander > The selected project will receive up to 2.5% of the overall EuroHPC computing capacity for one year
And just took a quick skim, Domyn seems to do "take nemotron, run on "eu enterprise nice data" whatever that means?
@Dorialexander We're so fucking cooked
@redtachyon Well we supported the one capable team for this, but not meant to be.
@Dorialexander @zehavoc At least it’s European
@Dorialexander I bet you whatever you want they’ll still have to write another euroHPC grant to get some compute
@zehavoc Likely. Whole setting is bizarre (they get a % of all available compute)
@Dorialexander wait!? why not @MistralAI ???
@giffmana The one we didn’t get 🥲
Funnily, Pleias first LLM compute plan was pretty close: a 30-50b range hybrid mamba mixture of expert trained on synth (back from January 2024, unfortunately never funded).
@mimrocker @Dorialexander @zehavoc This was obviously a joke
@antoine_chaffin @Dorialexander @zehavoc You don't know how many people believes that genuinely.
@Dorialexander the EU is nothing if not a consortium of consortia
@murchiston mixture of (non-)experts
@Dorialexander It means that they’ll have to distribute the training over all European clusters ?
Btw, never heard of that company before, have you ?
@Dorialexander This is the real problem. It's structural. I tell everyone who will listen.
It can't be one shot. If there is nowhere for iteration, nowhere for expertise to accrue and for processes to be refined, then you're never going to compete.
You have to imagine Sisyphus happy (from latest blog)
@antoine_chaffin @Dorialexander @zehavoc That's exactly the mindset that is so dangerous to our continent (in the context of AI).