-
Nvidia’s Llama Nemotron models were developed quickly, said Jonathan Cohen, VP of Applied Research.
-
The speed was thanks to researchers across the company being willing to “give up their compute.”
-
“These days, the currency in any AI researcher is how many GPUs they get access to,” he said in an interview.
In the world of AI research, the speed of development is limited in large part by available computing resources, according to Jonathan Cohen, Vice President of Applied Research at Nvidia.
“These days, the currency in any AI researcher is how many GPUs they get access to, and that’s no less true at Nvidia than at any other company,” Cohen said in an interview on Nvidia Developer.
Cohen led the team responsible for developing Nvidia’s Llama Nemotron family of models. Released in March of this year, they represent the company’s entry into the world of “reasoning” AI systems.
The speed at which the models came together was remarkable, Cohen said, taking “no more than one to two months.” He partially credits the efficiency of their development to other workers being willing to sacrifice their processing power.
“So, there were a lot of researchers who very selflessly agreed to give up their compute so that we could get these Llama Nemotron models trained as quickly as we did,” he said.
Cohen also attributed the speed of development to Nvidia’s company-wide culture of prioritizing major projects, regardless of current team goals.
“How do you have a team to do a thing you’ve never done before? Part of the corporate culture is — we call them a ‘swarm’ — where you identify, ‘This is something that’s important,'” he said. “And everyone, every manager who has people who might be able to contribute, thinks about, ‘Is this new thing more important than the current thing everyone on my team is doing?'”
If the manager can spare anybody, they’ll “contribute” their direct reports to the new priority.
“Llama Nemotron ended up being a very cross-discipline, cross-team effort,” Cohen added. “We had people from across the whole company working together without any formal organizational structure.”
Llama Nemotron required a series of sacrifices, Cohen said, both in terms of computing power and personnel — but people were able to set aside self-interests for the benefit of the whole.
“It was really great to see, great leadership,” he said. “There were a lot of sacrifices that people made, a lot of very egoless decisions that brought it together, which is just awesome.”
Nvidia did not respond to a request for comment by Business Insider immediately prior to publication.
Read the original article on Business Insider
