Joelle Pineau says Meta has a strong culture of open protocols that predates its AI team.
“There's a strong culture towards open protocols at Meta that predates the AI team.”
Person
Joelle Pineau
2 episodes · 21 quoted claims
Joelle Pineau says Meta does not necessarily open-source its production models because that would expose it to undue attacks.
“We don't necessarily open source the models that are running into production. That would open us up, I think, to undue attacks”
Joelle Pineau says Meta released successive Llama language models and that each generation improved significantly, partly through community contributions.
“we release Llama one, Llama two, Llama three, and so on. And every generation gets significantly better.”
Joelle Pineau moved from academic AI research to industry in 2017, saying the kinds of research questions available were already diverging.
“I decided to make the jump in in 2017 because already we were seeing a divergence in terms of the type of research questions that you could take on.”
Joelle Pineau says the Fundamental AI Research (FAIR) team she led at Meta grew from about 100 people to roughly 600.
“the fundamental AI research team I was leading grew, you know, from maybe a hundred people to to 600 people or so.”
Cohere has two core pillars: building its own foundation models and an agentic AI platform that Joelle Pineau calls the core of its product strategy.
“we do build our own models, and the other one is this agentic AI platform that we build, which is really the core of the the product strategy”
Cohere's North is an agentic AI platform designed to take the company's models and deploy them to businesses and governments.
“we have North, an agentic AI platform. It's designed to ingest the models that our modeling team is doing and deploy that”
Joelle Pineau says Cohere's platform can be deployed on-premise so agents can use a company's internal data without that data being sent to Cohere.
“you can run locally with your own data, so you can have a conversation or your agents can see actually all the internal data from the company without that data coming all the way to Cohere”
Joelle Pineau cited a figure of about 150,000 OpenClaw agents having been created in the first days after its viral launch.
“I think the last number I saw was 150,000 of these agents getting created.”
Joelle Pineau defines prompt injection as using a prompt to insert information that leads an AI agent to behave maliciously.
“the ability through a prompt to inject information that would lead the agent to take on malicious behaviors.”
Joelle Pineau uses Waymo self-driving cars stuck in a parking lot as an example that interactions among many autonomous systems can be harder than classic dilemmas like the trolley problem.
“the problem that really turns out to be hard is like all the Waymo stuck in a parking lot, like spinning their circles”
Joelle Pineau says enterprise customers still expect platforms to provide high levels of privacy, unlike consumers who have largely accepted responsibility for their own data.
“On the enterprise side, there's still an expectation that the platform will will provide high levels of privacy.”
Joelle Pineau defines open source in research as sharing all the information used in the work so that the results can be reproduced.
“Open source is about essentially, like, sharing all the information that was used in the course of the research in a way that facilitates the the reproducibility of that research result.”
Open weight language models share the trained network's weights so others can download, run locally, adapt or fine-tune them.
“we've seen much more movement towards open weight models, meaning that the weights of the network are shared such that someone can download that and essentially, like, rerun the network locally.”
Joelle Pineau notes that open weight releases often omit the training code and training data, sometimes providing only inference code.
“you might not have the training code, you might have only the inference code, you might not have the training data.”
Cohere evaluates models using sanity-check benchmarks, benchmarks that drive progress, and a third group of experience tests that use A/B testing and human evaluators.
“And then we have a third bucket that's more the experience, you know, we have a whole harness for AB testing and for bringing in human evaluators”
As an undergraduate at the University of Waterloo, Joelle Pineau did a work term at the Flight Research Lab in Ottawa building a speech recognition system for helicopter cockpits.
“I did a a work term one fall at the Flight Research Lab in Ottawa, where we were essentially building like speech recognition system in the cockpit of helicopters”
Joelle Pineau says Cohere has teams in Canada, the UK, France, Germany and the US.
“We have teams based in Canada, teams in the UK, in France, in we have folks in Germany, we have folks in the US.”
Joelle Pineau says Cohere's research-led focus on multilingual models opened commercial markets in Korea and Japan.
“it's opened up new commercial markets because when we get to Korea and to Japan, you know, of course, there's like, oh, you have world-class models in the local language”
Joelle Pineau said she had recently been in Paris and Davos, where she heard a lot of interest in sovereign AI.
“I was in in Paris earlier this month, I was in in Davos earlier this month also”
Joelle Pineau has four children aged roughly 16 to 22 who use AI for homework and cover letters.
“I have four kids who are sort of sixteen to twenty two, covering sort of high school, late high school, till university.”