← AI Deep Dive
AI Deep Dive
Meta's AI Instagram Comments, Mercedes-Benz's Humanoid Robots, & Hugging Face vs. The White House
00:15:09
Loading summary
Transcript134 lines
- [00:07]
A
Man, it feels like every single day there's some new AI thing popping up. A new headline, a new model, a new whatever. It's like drinking from a fire hose. And I don't know about you, but sometimes I just have to step back and think, okay, what really matters here?
- [00:25]
B
What.
- [00:26]
A
What are the big things to pay attention to?
- [00:27]
B
Right. It's easy to get lost in the noise. That's why we're doing this deep dive. We want to cut through the hype and help you see the bigger picture. Where is AI actually going?
- [00:36]
A
Exactly? So for this deep dive, we've got a pretty interesting mix of stuff. There's this news piece about Meta. They're experimenting with AI generated comments on Instagram, of all things. Oh, wow. And then there's this research paper from a group called metr, that's M E T R. They're looking at how long AI can actually stay focused on a task, which I think is fascinating.
- [00:56]
B
That is interesting.
- [00:57]
A
Yeah. And then we'll look at how Mercedes Benz is starting to use humanoid robots in their factories.
- [01:02]
B
Oh, really?
- [01:02]
A
Yep. And finally, we'll talk about Hugging Face's recommendations to the White House about AI policy. So a lot to unpack here, but hopefully by the end, you'll have a much clearer sense of where AI is at right now and where it might be headed.
- [01:15]
B
Cool. Sounds good.
- [01:17]
A
All right, let's start with Meta and Instagram, because that one really caught my eye. You know, Meta loves to play around with new AI stuff. Maybe those AI characters they rolled out.
- [01:27]
B
Oh, yeah, the ones with all the. The backstories.
- [01:31]
A
Yeah, like they tried to make them into these full fledged personalities and it just felt weird.
- [01:35]
B
A little creepy.
- [01:36]
A
Yeah. And then they just kind of disappeared. So clearly not everything they try sticks. But now they're testing out AI generated comments on Instagram.
- [01:45]
B
Like, how does that even work?
- [01:47]
A
So picture this. You're scrolling through your feed, you see a photo you like and you want to leave a comment, but instead of racking your brain, you just tap this little pencil icon next to the comment bar. Oh, bam. Meta. AI analyzes the photo and gives you three comment suggestions.
- [02:01]
B
Oh, wow.
- [02:02]
A
Yeah. So the article gave an example of a living room photo. The AI suggested things like cute living room setup or love the cozy atmosphere or even great photo shoot location.
- [02:13]
B
Huh. Okay.
- [02:14]
A
And if you don't like any of those, you can just hit refresh and get more options. And it's not just Instagram. Meta's spokesperson confirmed that they're looking at using AI all over their platforms in Comments in the feed, in groups, even in search. They say they want to make things more fun and useful.
- [02:32]
B
More fun and useful, huh?
- [02:34]
A
Yeah, their words. But here's the thing. I think this raises some really interesting questions about what people actually want from social media. I mean, a lot of us, we value genuine interactions, right?
- [02:45]
B
Absolutely.
- [02:46]
A
Real people, real thoughts, real connections. And throwing AI generated comments into the mix, it feels like, I don't know, it feels like it could cheapen the experience somehow. The article called it AI Slop, which I thought was pretty funny.
- [02:58]
B
Yeah, that's a good one.
- [02:59]
A
Like, imagine scrolling through comments and you can't even tell which ones are from real people and which ones are just, you know, AI spitting out generic platitudes.
- [03:07]
B
Yeah, you start to lose that sense of authenticity.
- [03:10]
A
Exactly. And I think there's a real longing for a more raw, unfiltered Instagram. It's like back in the day when it was just people sharing snapshots of their lives, not this perfectly curated AI enhanced. Whatever it's becoming.
- [03:23]
B
Right. It's almost like we're overthinking it now.
- [03:26]
A
Yeah, maybe. Anyway, the big question is, will this feature actually roll out to everyone?
- [03:30]
B
Right.
- [03:31]
A
Who knows? Meta's just testing it for now, so we'll have to wait and see how it goes.
- [03:35]
B
I guess time will tell.
- [03:36]
A
Okay, so moving on, let's talk about this research from mpir. They're trying to answer a really fundamental question about AI. How do we actually measure its progress?
- [03:49]
B
Yeah, how do you do that?
- [03:50]
A
Right. Like, a lot of times we hear about AI beating humans at games or getting high scores on tests, but that doesn't necessarily tell us how good AI is at doing, you know, real world tasks. And that's what these researchers are trying to figure out.
- [04:03]
B
Interesting. So what's their approach?
- [04:05]
A
Well, they're looking at how long AI can stay focused on a task and complete it successfully. They call it the AI's time horizon.
- [04:13]
B
A time horizon?
- [04:14]
A
Yeah. So think about it this way. Current AI, it's really good at things like writing a paragraph of text or solving a quick math problem. But can it write a whole essay or plan a complex project? That's where it starts to break down. It's like it has a limited attention span. And the researchers are saying that this time horizon is a much better measure of AI's real world capabilities than just looking at its performance on isolated tasks.
- [04:39]
B
Okay, that makes sense. So how are they measuring this time horizon thing?
- [04:42]
A
Well, they basically timed how long it takes a human expert to do a bunch of different tasks, and then they compare that to how long it takes an AI to do the same tasks.
- [04:51]
B
Huh. Interesting.
- [04:52]
A
And what they found is that current AI models, they have a near 100% success rate on tasks that take a human less than about four minutes. But when you get to tasks that take four hours or more, the AI success rate drops to less than 10%.
- [05:05]
B
Wow. Okay, so AI is good for those quick hits.
- [05:08]
A
Exactly. It's like it loses steam after a while. And they estimate that a model like Claude 3.7 Sonnet, which is a pretty advanced model, has a time horizon of about one hour.
- [05:17]
B
So one hour is the limit basically.
- [05:20]
A
For now. Yeah. But here's the really mind blowing part. They looked at how this time horizon has been changing over the past six years, and they found that it's doubling roughly every seven months.
- [05:31]
B
Wait, every seven months?
- [05:32]
A
Yeah. So that means that in a couple of years, AI might be able to handle tasks that take a whole week to complete.
- [05:39]
B
That's, that's a huge jump.
- [05:42]
A
It is. And they even say that more Recent data from 2024 and 2025 suggest that this doubling time even shorter now. So we could be talking about AI handling month long tasks sooner than we think. Wow. Of course, you know, they acknowledge that these are just projections.
- [05:58]
B
Right.
- [05:59]
A
And it's hard to predict the future.
- [06:00]
B
Yeah.
- [06:01]
A
But their findings really highlight how quickly AI is advancing and not just in terms of, you know, raw processing power, but in its ability to actually do things that we would consider complex and time consuming.
- [06:14]
B
That's really interesting. So what kind of implications do they see for this in, like, the real world?
- [06:19]
A
Well, they think it could have a massive impact on automation. I mean, if AI can handle week long tasks or even month long tasks, that opens up a lot of possibilities. For better or worse. They also did a bunch of tests to make sure their findings were solid. They looked at different types of tasks, different AI models, even real world data from software engineering projects. And the trend was pretty consistent across the board.
- [06:42]
B
That's pretty compelling.
- [06:43]
A
Yeah. So, you know, for you listening, this research is really about shifting our perspective on AI progress. Yeah. About benchmarks and test scores. It's about understanding how AI's ability to handle real world tasks is rapidly evolving.
- [06:59]
B
Right.
- [07:00]
A
And what that means for the future of work and automation.
- [07:02]
B
Yeah, it's a big deal.
- [07:03]
A
Okay, let's shift gears a bit and talk about something a little more concrete. Mercedes Benz, they're actually putting humanoid robots into their factories right now.
- [07:12]
B
No way. Really?
- [07:13]
A
Yep. And they're using Google DeepMind's AI to control them. It's like something out of science fiction.
- [07:19]
B
Wow. Are they replacing their human workers with robots?
- [07:23]
A
Well, not exactly. At least not yet. Right now they're focusing on using the robots for what they call intra logistics.
- [07:30]
B
Intralogistics?
- [07:31]
A
Yeah. Basically means moving things around the factory. So the robots, they transport components between workstations, they help with loading and unloading things like that Makes sense. It frees up the human workers to focus on more complex tasks like assembly and quality control. Mercedes says it's all about making the production process more efficient and safer. And they're not just using the robots as mindless drones. They're using Google DeepMind's AI to give them some smarts.
- [07:58]
B
Like how so?
- [07:59]
A
Well, the AI helps the robots navigate the factory floor, avoid obstacles, and even learn new tasks. It's also being used for quality control, like detecting defects in parts and analyzing data to optimize the production process.
- [08:14]
B
So it's like the AI is the brain and the robot is the body.
- [08:17]
A
Yeah, kind of like a really high tech team. And they're using different types of Gemini models from Google, like Gemini Robotics and Gemini Robotics Extended Reasoning, which is based on Gemini 2.0, apparently. So they're really going all in on this AI stuff.
- [08:31]
B
That's fascinating.
- [08:32]
A
Yeah. And a guy from Mercedes, Joerg Berther, he was quoted in the article saying that they want to train tasks away from the workforce.
- [08:41]
B
Train tasks away?
- [08:42]
A
Yeah, it's kind of weird way to put it, but basically they want to automate those repetitive and physically demanding jobs that nobody really wants to do anyway.
- [08:50]
B
Right, right.
- [08:51]
A
And let the robots handle those so the human workers can focus on things that require more skill and creativity.
- [08:57]
B
I can see the logic there.
- [08:59]
A
Yeah. Right now they're testing these robots and AI systems in their factories in Berlin and Hungary and they're using a method called teleoperation to train the robots, which basically means that a human operator controls the robot remotely. Kind of like a video game.
- [09:15]
B
Ah, okay.
- [09:16]
A
But the goal is to eventually have the robots learn and operate autonomously so.
- [09:20]
B
They can work on their own.
- [09:21]
A
Exactly. And here's the really interesting part. They're not planning to roll this out in China. They specifically said that because of geopolitical concerns, they're not going to introduce these technologies there.
- [09:33]
B
Huh, that's interesting.
- [09:35]
A
Yeah, it shows how these technological decisions are often tied to larger political issues.
- [09:40]
B
Yeah, for sure.
- [09:41]
A
It's not just about the tech itself. And it's not just Mercedes doing this either. Other car companies like Honda, Hyundai and BMW are Also experimenting with humanoid robots in their factories. Even Elon Musk talked about Tesla's Optimus robot a while back, although we haven't heard much about that lately.
- [09:58]
B
Yeah, whatever happened to that?
- [09:59]
A
Right? Who knows? But the point is, this is a real trend. AI and robotics are moving beyond the lab and into the real world, and it's happening fast. So, you know, for you listening, this is something to pay attention to and have a big impact on the economy, on jobs, on how we think about work in general, for sure. Okay, last but not least, let's talk about Hugging Face and their recommendations to the White House about AI policy. Now, Hugging Face? They're a big player in the open source AI world. They have this platform where people can share and collaborate on AI models and tools, and they're basically arguing that OpenAI is the way to go.
- [10:34]
B
Like open source?
- [10:36]
A
Exactly. They're saying that open systems are more transparent, more adaptable and more scientifically sound than closed proprietary systems.
- [10:46]
B
Interesting.
- [10:46]
A
Yeah. And they're not just talking theory. They gave some specific examples like their own Olympic coder model. They say it outperforms Claude 3.7 on certain coding tasks, even though it's smaller and uses fewer parameters.
- [10:58]
B
Wow, really?
- [10:59]
A
Yeah. And they pointed to AI2's open Olmo 2 models, which they say match the performance of O1 mini using open training data. So they're making the case that open doesn't mean less powerful or less capable.
- [11:12]
B
I see.
- [11:13]
A
And they laid out three main recommendations for the White House.
- [11:15]
B
Okay, what are they?
- [11:16]
A
First, they're saying that open source and open science should be recognized as fundamental to the success of AI. They pointed out that a lot of the key breakthroughs in AI, like attention mechanisms and transformers, came from open research. And they argue that we need more public investment in research infrastructure, access to computing power and support for open data sets, especially for smaller developers and researchers.
- [11:41]
B
That makes sense.
- [11:41]
A
Yeah. Second, they're emphasizing efficiency and reliability.
- [11:45]
B
Okay.
- [11:46]
A
They argue that we should focus on developing smaller, more specialized AI models that are tailored to specific needs. They're also pushing for more efficient inference techniques, which is how we use trained AI models, and for more mid scale training initiatives. They say this would make AI more accessible and more useful for a wider range of applications, especially in areas like healthcare, where reliability is crucial. And finally, they're really focusing on security. They're arguing that open and transparent AI systems are actually more secure because they can be scrutinized and audited more easily. They also say that open infrastructure and tooling make it easier to train and deploy AI models in controlled environments. And they're pushing for the use of open weight models, which can be important for security in certain situations. Basically, they're saying that we need to build expertise in OpenAI and prioritize its development, especially for critical applications.
- [12:41]
B
So transparency is key.
- [12:43]
A
Exactly. So, you know, for you listening, Hugging Face's perspective is really important because it offers a counterpoint to the idea that AI development should be driven primarily by big companies with closed proprietary systems. They're making a strong case for open collaboration and transparency as the best way to ensure that AI is beneficial and safe for everyone.
- [13:02]
B
Yeah. And it's a debate that's happening right now, so it's something to pay attention to.
- [13:06]
A
Yeah, definitely. Okay, so let's zoom out and take a look at the big picture. What have we learned from all of this?
- [13:11]
B
Well, I think one of the main takeaways is that AI is advancing at an incredible pace.
- [13:16]
A
Absolutely.
- [13:17]
B
And it's not just about those flashy headlines about AI beating humans at games or whatever. It's about real practical capabilities that are going to have a huge impact on our lives.
- [13:27]
A
Yeah, we saw that with Meta, exploring how AI could change our social media interactions, maybe in ways we don't even fully understand yet.
- [13:35]
B
Yeah.
- [13:36]
A
And we saw that with the research from A. Kr their findings about AI's rapidly expanding time horizon are pretty mind blowing.
- [13:43]
B
Definitely.
- [13:44]
A
And then we saw Mercedes Benz putting robots and AI to work in their factories. Right. So it's clear that AI is moving beyond the lab and into the real world in a big way.
- [13:55]
B
Yeah.
- [13:55]
A
And hugging faces arguments about OpenAI. They highlight the importance of thinking carefully about how we develop and govern these technologies.
- [14:03]
B
Yeah, for sure. It's not just about the tech itself.
- [14:06]
A
Exactly. It's about the values that we embed in these systems, the choices that we make about how to use them, and the potential consequences for society as a whole.
- [14:14]
B
Yeah. It's a lot to consider.
- [14:15]
A
It is. So here's a final thought for you to ponder. Given all of these developments from AI generated comments to humanoid robots and open source advocacy, what do you think is going to be the most profound impact of AI in the near future? Like what aspects of our lives, our work, our interactions with each other are going to be most fundamentally changed?
- [14:41]
B
That's a big question.
- [14:42]
A
It is, but I think it's a question worth asking. The answers are going to shape the world we live in.
- [14:48]
B
Yeah.
- [14:48]
A
And the world that our kids are going to inherit.
- [14:50]
B
Right.
- [14:51]
A
So that's it for this Deep dives. Thanks for joining us.
- [14:53]
B
Yeah, thanks for having me.
- [14:54]
A
And as always, the world of AI Keeps on changing. So we'll be back soon with more to explore and unpack.
- [14:59]
B
Absolutely. See you next time.