← AI Deep Dive
Loading summary
Transcript105 lines
- [00:07]
A
Ever feel like just keeping up with AI and tech news is like a full time job? I mean, it's, it seems like new models are constantly dropping, pricing structures are shifting, and then you've got studies that make you question like everything.
- [00:20]
B
Yeah, it can be a real whirlwind trying to stay on top of it all.
- [00:22]
A
Absolutely. So that's what this deep dive is all about, helping you cut through all that noise and figure out, like, what really matters.
- [00:28]
B
Right. We're going to be your guides today, helping you sort through some of the latest and greatest in AI and digital platforms.
- [00:35]
A
Yeah. So basically we've been digging into some recent news and we found some really fascinating stuff covering, like, new AI models, changes in how these AI coding assistants are priced, even a study that looks into the data that's actually used to train these AIs, and then some big updates to how advertising works on Spotify.
- [00:54]
B
Lots to cover.
- [00:54]
A
Okay, so to kick things off, let's talk about Meta, because they just dropped a whole new lineup of AI models called Llama 4. And they've got some interesting names like Scout, Maverick and Behemoth, which is still in the works, but they didn't even wait for like a normal workday. They released all this on a Saturday.
- [01:09]
B
That's right. Straight into the weekend.
- [01:11]
A
Yeah. And what really caught my eye was that they're saying these models have like this broad visual understanding because they were trained on so much data, like text, images, video, all unlabeled.
- [01:25]
B
Yeah. And you know, that unlabeled part is really interesting because usually AI models, they learn from this really carefully organized and labeled data. Right, but training on unlabeled data, that's a whole other approach.
- [01:36]
A
Right, so it's more like letting the model just figure things out on its own.
- [01:40]
B
Exactly. It's about finding patterns and connections without being told exactly what to look for. And then saying it has broad visual understanding that suggests it's getting much better at understanding images alongside text and video.
- [01:52]
A
You know, and the word on the street is that Meta was in a real rush to get these models out there.
- [01:57]
B
Yeah.
- [01:57]
A
Apparently those open source models coming out of China, especially from a company called Deepseek, they kind of lit a fire under Meta.
- [02:03]
B
That makes sense. The competition is really heating up.
- [02:06]
A
Yeah, and I heard they even had teams like, dedicated to figuring out how Deep Seek managed to make their operations so cost effective.
- [02:12]
B
That's the key. Right now it's not just about making a better model, but making one that's actually practical to use.
- [02:19]
A
Right. So if you're a developer or just, you know, curious to tinker around. Scout and Maverick. Those are already available on llama.com and through platforms like Hugging Face.
- [02:29]
B
Yep. You can get your hands on them right now.
- [02:31]
A
And then they've also got Meta AI, which is their assistant that's built into, like, WhatsApp, Messenger, Instagram, and it's now powered by llama4 in a lot of countries.
- [02:41]
B
That's a pretty big move right there.
- [02:43]
A
Yeah, but it looks like those really fancy features that work with, like, all different types of data, you know, that multimodal stuff that's only in the US and only in English for now.
- [02:52]
B
Yeah. They're probably still working out the kinks, you know, making sure it works smoothly in different languages and all that makes sense.
- [02:58]
A
But here's where it gets really interesting. The llama 4 license.
- [03:02]
B
Okay, what about it?
- [03:03]
A
Well, it looks like they've put some restrictions in place. If you or your company is based in the eu, you can't actually use or distribute these models.
- [03:13]
B
Wow, that's a big deal.
- [03:14]
A
Yeah. And just like with the previous Llama models, if your company has over 700 million monthly active users, you need a special license directly from Meta.
- [03:25]
B
So they're really keeping a tight rein on who can use these models, especially the really powerful ones.
- [03:30]
A
Exactly. And on a more technical note, this is actually Meta's first time using what's called a mixture of experts, or moe, architecture for their models.
- [03:38]
B
And moe, that's a really interesting approach. It's all about efficiency.
- [03:42]
A
Yeah. So instead of having one massive model that does everything, they break it down into these smaller expert models.
- [03:48]
B
Right. It's like a team of specialists, each one really good at a specific thing.
- [03:52]
A
So when the model gets a task, it figures out which expert is best suited and sends that part of the problem to them.
- [03:58]
B
And that way they can have this huge overall model with tons of knowledge, but they don't have to use all of it for every single task.
- [04:05]
A
Right. So it's like having a brain surgeon, a rocket scientist and a chef all on call, but you only have to pay for the one you need at that moment.
- [04:13]
B
Exactly. It's much faster and uses way less computing power.
- [04:16]
A
Okay, so let's talk specifics. Maverick. This one has, like, a whopping 400 billion parameters, but only about 17 billion are active at any one time. And those are spread across 128 different experts.
- [04:29]
B
Yeah. And Meta saying, this one is great for general tasks like chatting and creative writing, you know, your everyday AI stuff.
- [04:36]
A
Right. And they're even claiming it beats out some pretty well known models like GPT4.O and Gemini 2.0 on certain tests.
- [04:43]
B
Those benchmarks are always interesting. Right. It gives you an idea of how they stack up, but it doesn't tell the whole story of how a model performs in the real world.
- [04:49]
A
Right, then we've got Scout. This one's a bit smaller. 109 billion total parameters, 17 billion active, cross 16 experts. And this one's strengths are in summarizing documents and really getting deep into code.
- [05:02]
B
That's a critical area these days.
- [05:04]
A
And get this, it has a 10 million token context window.
- [05:07]
B
Wow. Okay, so for those who aren't familiar with tokens, that's like being able to read and understand several entire books at once.
- [05:12]
A
Yeah, and apparently it's efficient enough to run on a single high end Nvidia.
- [05:15]
B
Graphics card, so that makes it much more accessible for researchers and developers who don't have access to these massive data centers.
- [05:22]
A
Now, Behemoth, this one's still in training and the numbers are just mind blowing. Nearly 2 trillion parameters. 288 billion active across 16 experts.
- [05:32]
B
That's getting into some serious, serious territory.
- [05:34]
A
And Meta's internal tests show that it's already beating out some of the top models like GPT 4.5 and Claude 3.7 Sonnet, especially when it comes to, like, STEM stuff.
- [05:45]
B
That's pretty impressive. But you know, with that many parameters, it's probably going to need some serious hardware to run.
- [05:50]
A
Yeah, definitely not something you're going to be running on your laptop. But here's an interesting thing. Meta specifically says that none of these Llama 4 models are quite like some of the other reasoning models out there.
- [06:00]
B
Right. Those reasoning models are designed to be more like fact checkers, giving really reliable answers, even if it takes a little longer.
- [06:07]
A
And then one last thing about Llama 4, Meta is claiming they've tweaked these models so they're less likely to refuse to answer certain types of questions.
- [06:15]
B
Hmm.
- [06:15]
A
And they're supposed to be better at giving more balanced perspectives on controversial topics.
- [06:19]
B
Well, you know, it's a tricky thing trying to make sure these AI models are both helpful and unbiased.
- [06:24]
A
For sure. Okay, so let's shift gears now and talk about GitHub Copilot, which is that AI coding assistant that a lot of developers love.
- [06:32]
B
Yep, I've heard good things about it.
- [06:34]
A
So they're introducing this new pricing structure with something called premium requests. And it's basically like this. If you want to use those more advanced AI models for certain tasks, you Know, like doing really complex coding or editing multiple files at once. You have a limited number of these premium requests each month.
- [06:53]
B
Okay, so they're basically creating tiers of access based on which AI model you're using.
- [06:57]
A
Exactly. So if you're using the standard model, which is OpenAI's GPT4.0, you've got unlimited requests. But if you want something fancier, like Anthropic's 3.7 sonnet, then you've got to watch those premium requests.
- [07:09]
B
And those limits depend on your subscription level.
- [07:11]
A
Right, so Copilot Pro subscribers get 300 premium requests a month, business users get 300 as well, and enterprise subscribers get 1,000.
- [07:20]
B
Okay, makes sense. And if you run out, you can always buy more. Yep.
- [07:23]
A
Or you can upgrade to their new CoPilot Pro plus plan, which costs $39 a month, but gives you 1500 premium requests and access to what they're calling the best models, including GPT 4.5.
- [07:36]
B
So they're definitely betting that there are enough people willing to pay for the premium experience.
- [07:41]
A
And it seems like Copilot is a pretty big deal for GitHub's bottom line. I read that Microsoft CEO Satya Nadella said that Copilot was responsible for over 40% of GitHub's revenue growth this year and that it's already bigger than GitHub was when Microsoft acquired it.
- [07:57]
B
Wow, that's a huge impact. It really shows how these AI tools are transforming the developer world.
- [08:02]
A
Okay, so now for something completely different. There's a new study out that's adding fuel to the whole debate about whether AI models are being trained on copyrighted material.
- [08:10]
B
Yeah, this is a big issue with all the lawsuits from authors and programmers.
- [08:14]
A
Right. So these researchers, they came up with a clever way to investigate this and see if AI models are actually memorizing copyrighted content.
- [08:20]
B
It's like trying to peek inside the black box of AI.
- [08:23]
A
Yeah. So they focused on what they call high surprisal words, which are basically words that are unexpected in a given context.
- [08:29]
B
So like a word that you wouldn't normally see in that particular sentence or phrase?
- [08:33]
A
Exactly. They took snippets of text from fiction books and New York Times articles, and they hid these high surprisal words, and then they asked the AI models to guess the missing word.
- [08:44]
B
Interesting. So if the model could consistently guess those unusual words, it would suggest that it had actually seen that specific passage before during its training.
- [08:53]
A
Yep. And they tested this on models like OpenAI's GPT4 and GPT 3.5. And what they found was that GPT4 did show signs of memorizing parts of popular fiction books, especially those from a data set called Book mia, which has a bunch of copyrighted e books.
- [09:09]
B
Okay, so there's some evidence there.
- [09:11]
A
And they also found evidence of memorization from New York Times articles, although not.
- [09:14]
B
As much that makes sense since the New York Times has a massive amount of content.
- [09:18]
A
One of the study's authors, Balasha Ravachandra, she said that we need to be able to really examine these AI models and audit them scientifically. And she called for more transparency from companies about what data they're using to train their models.
- [09:30]
B
Yeah, it's a valid point. If we're going to trust these AIs, we need to understand where their knowledge is coming from.
- [09:35]
A
Okay, last but not least, let's talk about Spotify. They've just announced some major updates to their advertising platform and it seems like they're trying to make things easier for advertisers to buy, create and track their campaigns.
- [09:48]
B
So they're really investing in their ad business.
- [09:50]
A
Yeah. The biggest news is the launch of something called the Spotify Ad Exchange or ssx. And this allows advertisers to access Spotify's users through real time auctions. And they're already partnering with big names like the Trade Desk and Google display and video 360.
- [10:06]
B
So they're moving towards that programmatic advertising model where advertisers can bid on ad space in real time.
- [10:12]
A
Exactly. And are also making improvements to their self service Spotify Ads Manager, giving advertisers more control over targeting and measurements. And then there's something called Spotify Gen AI Ads which lets advertisers in the US and Canada use AI to generate ad scripts and even voiceovers.
- [10:30]
B
Wow. So you can just tell the AI what you want to advertise and it'll create the whole ad for you?
- [10:35]
A
Pretty much. They're also extending their measurement tools and partnering with third party measurement providers.
- [10:40]
B
So they're really trying to give advertisers that confidence that their Spotify ads are actually working.
- [10:45]
A
Exactly. And Spotify's global head of advertising, Lee Brown, he talked about how Spotify's free users are engaged for like two hours a day on average. And that Gen Z specifically sees Spotify as a positive escape from all the negativity online.
- [11:00]
B
Yeah, I've heard that too. It's their little sanctuary from the doom scrolling.
- [11:04]
A
Yeah, in their 2024 Culture Next report, they found that 71% of Gen Z said Spotify is like an antidote to doom scrolling.
- [11:11]
B
So it's more than just a music platform for them. It's like a positive space.
- [11:14]
A
And then quickly they've got these in house agencies, Creative Lab and aux. Creative Lab works directly with brands on campaigns and AUX is their music consulting agency.
- [11:25]
B
So they're offering a full suite of services to advertisers.
- [11:28]
A
Right. They're not just selling ad space, they're trying to be partners in creating those connections with listeners.
- [11:33]
B
And that Gen Z audience is a prime target for a lot of brands.
- [11:37]
A
Okay, so wrapping up this deep dive, we've covered a lot. Meta's new Llama 4 models with all those interesting licensing issues, GitHub, Copilot's new pricing structure, the study about AI training data and copyright, and Spotify's big push to attract advertisers.
- [11:53]
B
It's a lot to take in, but it really shows how dynamic this whole AI and tech landscape is right now.
- [11:59]
A
Absolutely. So for all you listeners out there, what part of all this do you think will have the biggest impact in the near future? Or maybe what are you most curious about as this all continues to unfold?
- [12:10]
B
Yeah, it's a question worth pondering. And we encourage you to keep exploring these topics. Maybe go back and read those original articles and just keep that curiosity going.
- [12:18]
A
Understanding all this. It's an ongoing process, but hopefully this deep dive gave you some food for thought.