LLaMA 2: Meta's next AI model release | Mark Zuckerberg and Lex Fridman

Please support this podcast by checking out our sponsors:

GUEST BIO:
Mark Zuckerberg is CEO of Meta.

PODCAST INFO:

SOCIAL:

[LexClips]

Guest bio: Mark Zuckerberg is CEO of Meta.

[np]

Can't wait to see LLaMA 2! I hope it does end up being open sourced in the pursuit of innovation, as we've seen some of the research already done has been competitive with GPT-3.5.

[wodot]

We need AI that can answer any question, with no "ethic concerns" or bs like that. Knowledge is knowledge, even if it's knowledge for doing bad things, knowing in on itself ain't bad.

[Regoflika]

Llama 2 released today in partnership with Microsoft!

[woolfel]

Yannic and several others started Open assistant that's 100% open source. It is already happening.

[simonabunker]

All I heard was "better scammers on WhatsApp". Open sourcing Llama properly does sound interesting though - it's clear that a lot of innovation was done with the leaked version, and Google also sees open source as the future.

[DraftedByTheMan]

Hard to listen to Zucks word salads. So many words without telling you anything.

[Axl]

I agree with Mark. Transformer architecture either with RLHF or without it is not the way to create AGI. It will not scale.

[zandrrlife]

This is exciting. If they open source more powerful models, I can assume they will be gravely handicapped, due to safety concerns. All that matters is the research though. I hear so much about removing biases from these models, but I assert these biases can be utilized in compelling ways. Inducing specific psychological profiles, for behavior modeling and prediction. Even to boost the models theory of mind. I do find value in a LM model inherit bias. We just need better steerability, gained through interpretability insights. Then those biases can properly be gated. An interesting research route I'm waiting for groups to explore is -- adversarial attacks through deliberately inducing specific types of biases, allowing for model jailbreaking. How can those risks be mitigated? I feel this is an important latent space. Just my 2 cents. Zuck love you man, hair? 😂

[stefanomatic]

Wow. It’s been a while since I listened to an interviewee spout so little with so much. ChatGPT could trim his words down to a few sentences, and still keep all the content and context of what he says, in an more meaningful way.

[iggswanna]

facebook for business is the most UNHOLY GOD DAMN HYPER MANIACLY COMPLICATED interface i have EVER SEEN. cuz you have your profile linked to your page whos linked to a business acount whos linked to a corporate account whos linked to your personal account again whos linked to another business acount who is linked to the same corporate account... AND I CANT FRIGGIN CANCEL THE CORPORATE ACCOUNT

[lisalove]

When the "tech" and " bro" are in balance 🌗

[moonstrobe]

Working hard on neutering the next model, so it's as braindead and obstinate as possible, when you ask it to write anything with remotely spicy content.

[JamesWitte]

Wikipedia is defs no longer crown sourced lol

[elon--musk]

Mark Zuckerberg you are a good guy, thank you

[hermes]

It is going to be as popular as the metaverse

[allatcap]

Someone tell em to fix n update facebook support its a joke. Im starting not to trust this guy

[robbyhood]

Wikipedia is the greatest website ever created? Maybe if you said the greatest weapon I would agree with you, the greatest CIA propaganda weapon ever created.

[joeyloizzo]

Is he human? That is not how people talk. Not a genuine dude, very reserved, like a PR rep promoting a product

[sootedninjas]

his full of it and becoming more and more irrelevant