Olmo 2 Challenges Big AI, AI Bots Take Over Airbnb, & Insta Founder Warns About Chatbot Hy

Summary

AI Deep Dive Podcast Summary
Episode: Olmo 2 Challenges Big AI, AI Bots Take Over Airbnb, & Insta Founder Warns About Chatbot Hype
Release Date: May 4, 2025
Host: Daily Deep Dives

---

Welcome to today’s detailed summary of the AI Deep Dive Podcast hosted by Daily Deep Dives. In this episode, the hosts explore four significant AI developments: the emergence of smaller yet powerful AI models like Olmo 21B, Airbnb's integration of AI-powered customer service bots, critical perspectives on AI engagement strategies voiced by Instagram co-founder Kevin Systrom, and an uplifting story of AI bringing historical photographs to life. Below is a comprehensive breakdown of each topic discussed, complete with notable quotes and timestamps.

---

1. Smaller AI Models: Olmo 21B Challenges the Giants

The episode kicks off with a discussion on Olmo 21B, a newly released AI model from the Allen Institute for AI (AI2). Despite having only 1 billion parameters, Olmo 21B is proving to be remarkably potent, outperforming larger models from tech giants like Google, Meta, and Alibaba in specific tasks.

Key Points:

• Parameters Explained: Parameters are the internal settings that guide an AI's functionality. While 1 billion parameters seem substantial, in the context of colossal language models, Olmo 21B is relatively modest.

  Host A [01:11]: "You've got the core tech, the models themselves, but then also these very human applications, how it affects us day to day or even connects us to the past."

• Open Sourcing: Olmo 21B has been open-sourced, making its code and extensive training datasets (Almo Mix 1124 and Dalmanomics 1124) available on platforms like Hugging Face for free. This democratization allows researchers, developers, and enthusiasts to access, replicate, and build upon the model.

  Host A [02:02]: "The code, the huge data sets they use to train it… it's all up on Hugging Face, free to use."

• Performance and Training Data: Trained on an immense 4 trillion tokens, Olmo 21B excels in areas such as arithmetic reasoning and factual accuracy, as demonstrated in benchmarks like GSM8K.

  Host B [04:07]: "AI2's benchmarks show it beating some well-known models on things like arithmetic reasoning."

• Caveats and Safety Concerns: Despite its impressive capabilities, AI2 warns against using Olmo 21B for commercial purposes due to potential "problematic outputs," including harmful or inaccurate information.

  Host A [04:32]: "Don’t use this for commercial stuff yet because it can still produce what they call problematic outputs."

Insights: The introduction of Olmo 21B highlights a trend towards more efficient and accessible AI models. By balancing performance with accessibility, AI2 is pushing the boundaries of what smaller models can achieve, fostering innovation while acknowledging the importance of responsible AI deployment.

---

2. AI Integration in Customer Service: Airbnb's Strategic Move

Next, the conversation shifts to Airbnb's deployment of AI-powered customer service bots, marking a strategic move to enhance operational efficiency.

Key Points:

• Deployment Details: Airbnb has integrated an AI customer service bot that currently handles interactions for approximately half of their U.S. users, with plans for a full rollout within the month.

  Host B [05:15]: "It's currently handling interactions for about half of their US users."

• Impact and Efficiency: Early results indicate a 15% reduction in the need for users to escalate issues to human agents, signaling significant gains in operational efficiency and cost management.

  Host B [05:34]: "They're reporting a 15% drop in the number of people who ultimately need to contact a live human agent."

• Strategic Focus: Unlike competitors focusing on comprehensive travel planning, Airbnb is initially concentrating on customer support, aligning with their core business needs and responding to a slowing growth outlook.

  Host A [06:34]: "Airbnb seems to be starting with the support side. Maybe that reflects their business."

Insights: Airbnb's cautious and targeted implementation of AI in customer service underscores a practical approach to leveraging AI for immediate benefits while monitoring user experience. This move is particularly salient as the company navigates a period of slower growth, emphasizing cost efficiency and customer satisfaction.

---

3. Ethical Concerns: Instagram’s Kevin Systrom Criticizes AI Engagement Tactics

The episode then delves into a more philosophical debate sparked by Kevin Systrom, co-founder of Instagram, who voices concerns over AI's focus on maximizing user engagement.

Key Points:

• Critique of AI Chatbots: Systrom criticizes AI chatbots for prioritizing prolonged interactions over providing concise and useful answers, suggesting that some AI systems may be designed to keep users engaged rather than genuinely assisting them.

  Host A [07:24]: "He criticizes chatbots that ask like endless follow-up questions… just to keep the conversation going."

• Comparison to Social Media Strategies: He draws parallels between AI engagement tactics and social media strategies aimed at maximizing user time and activity, raising questions about the underlying motives driving AI development.

  Host B [08:07]: "He draws a direct line to social media tactics, trying to maximize time spent daily, active users, those kinds of metrics."

• Debate on AI Design Philosophy: The discussion highlights a tension in AI development: whether the primary goal should be user utility or engagement metrics, with OpenAI advocating for clarity and directness in responses.

  Host A [09:02]: "AI development, does engagement become the primary goal, does utility suffer?"

Insights: Systrom's perspective introduces a critical viewpoint on the ethical implications of AI design choices. It challenges developers and companies to reflect on the motivations behind AI interactions, advocating for a balance between engagement and genuine usefulness to ensure AI serves the best interests of users.

---

4. AI in Storytelling: Bringing Historical Photos to Life

Concluding the episode on a heartwarming note, the hosts share the story of Mark Edwards from Scotland, who employs AI to animate old photographs, breathing new life into historical images.

Key Points:

• Project Overview: Mark Edwards used AI to animate old black-and-white photos of his hometown, Chedberg, initially to surprise and delight his grandparents. The project resonated widely, leading him to create similar animations for other towns like Kelso and Galashiels.

  Host A [09:36]: "He used AI to animate old photos of his hometown… bringing these old black and white stills to life."

• Community Impact: These animated photos are being utilized in local schools for history lessons and in social groups for older adults, exemplifying AI's potential to foster emotional connections and enhance educational experiences.

  Host A [10:12]: "These videos are actually being used in local schools for history lessons and also at social groups for older people."

• Personal Reactions: Individuals who experienced the animations shared touching stories, such as Monica recalling a personal memory at the old railway station, and Anne expressing joy over seeing historical buildings animated.

  Host A [10:19]: "One woman, Monica, saw the old Galashiels railway station animated and remembered meeting her dad there."

Insights: Mark Edwards' project illustrates the profound and positive ways AI can be harnessed beyond technical applications. By merging technology with storytelling, AI can evoke nostalgia, strengthen community bonds, and make history more engaging and accessible, highlighting its versatility and emotional resonance.

---

Conclusion: Weaving the AI Tapestry

Wrapping up the episode, the hosts reflect on the diverse threads of AI advancements discussed:

• Technological Innovation: The rise of smaller, efficient AI models like Olmo 21B showcases technological progress and the democratization of AI tools.

• Practical Applications: Airbnb's AI customer service bots demonstrate AI's tangible impact on operational efficiency and customer experience.

• Ethical Considerations: Kevin Systrom's critique invites a critical evaluation of AI's role in user engagement versus genuine utility.

• Human-Centric Uses: Mark Edwards' use of AI for animating historical photos underscores AI's potential to enhance emotional and educational experiences.

Host B [12:00]: "Taken together, it really paints a picture of AI not as one single thing, but as this collection of evolving tools finding niches everywhere."

Final Thoughts: The episode emphasizes that AI is multifaceted, permeating various aspects of technology, business, ethics, and personal life. It underscores the importance of balancing innovation with responsibility, ensuring that AI advancements benefit society holistically.

Host A [12:33]: "What other questions does it raise for you about where this is all heading? It really makes you think about that balance…"

As AI continues to evolve, the conversation highlights the ongoing need to navigate its complexities thoughtfully, fostering advancements that are not only powerful and efficient but also ethical and emotionally meaningful.

---

Stay tuned to AI Deep Dive for more insightful explorations into the ever-evolving world of artificial intelligence.

Transcript

[A]

You know how AI just seems to be, well, everywhere now? It feels like it went from sci fi to daily news almost overnight.

[B]

It really does. Yeah. The pace is something else.

[A]

So today we're doing a quick, deep dive into four recent AI stories that really stood out. Okay, we've got these new, smaller AI models that are surprisingly powerful. Then there's AI showing up in customer service. Also some interesting debate about AI trying maybe too hard to keep us engaged.

[B]

Ah, right, I saw that.

[A]

And finally, a really nice story about AI bringing old photos, like history to life.

[B]

Oh, that sounds interesting.

[A]

Yeah. So the goal is simple. Just pull out the key stuff, the important bits from these developments, nice and quick, without getting bogged down in, you know, jargon.

[B]

Got it. A focused look at what's happening.

[A]

Exactly. Okay, let's jump in. What's striking, I think, is just the sheer range here.

[B]

It really is. You've got the core tech, the models themselves, but then also these very human applications, how it affects us day to day or even connects us to the past. These four topics definitely show how broad AI is becoming.

[A]

Right, so let's start with those smaller models. There's one from AI2, the Allen Institute for AI called Olmo 21B. The 1B is 1 billion parameters. Now, parameters for. For anyone not deep in the weeds, they're kind of like the internal knobs and dials that guide how the AI works, Right?

[B]

Exactly. They're the components that the model learns during training. And 1 billion sounds like a lot.

[A]

It does.

[B]

But in the world of these huge language models we often hear about, it's actually, well, smaller.

[A]

Okay, so it's relative.

[B]

Very relative. Think different engine sizes. More parameters can mean more capability, but they need more fuel, more computing power.

[A]

Right.

[B]

What's interesting about Olmo 21B is that AI2 says it's punching above its weight, performing better on some key tasks than bigger models from Google Meta Alibaba.

[A]

Okay, that's significant. And here's what really caught my eye. They basically open sourced it.

[B]

Yeah, that's a big deal.

[A]

The code, the huge data sets they use to train it, these things called Almo Mix 1124 and Dalmanomics 1124. It's all up on hugging face, free to use. Why is that so important?

[B]

Well, making it open source, it really democratizes things. It lets researchers, developers, even hobbyists look under the hood, replicate the results, build on top of it.

[A]

So more eyes on it, potentially faster progress.

[B]

Exactly. More transparency too.

[A]

Yeah.

[B]

And because it is smaller, it doesn't need a supercomputer it can potentially run on say, a modern laptop or maybe even mobile devices down the line, which.

[A]

Makes it much more accessible for people to actually experiment with.

[B]

Precisely. It puts powerful tools in more hands.

[A]

And it feels like part of a trend. The source has mentioned other smaller models launched recently too.

[B]

Yeah, there have been a few like Microsoft's 5.4 Quinn's models. It's less a total shift away from big models, I think, and more about diversifying the toolbox.

[A]

Right. Tool for the job kind of thing.

[B]

Exactly. You don't always need the massive energy hungry model. Sometimes a smaller, more efficient one is perfect, especially for specific tasks or running locally on a device.

[A]

Makes sense. Now, you mentioned the training data. OMO21B was trained on 4 trillion tokens. What are tokens exactly? We talked parameters.

[B]

Right. So tokens are the basic units of data the AI processes. Think of them like words, or sometimes parts of words, or even punctuation.

[A]

Right.

[B]

4 trillion tokens is just an immense amount of text and code. It comes from all over public websites, books, code repositories, even some AI generated stuff and manually cleaned data just to.

[A]

Give a sense of scale. Like how many words is that roughly?

[B]

Well, a million tokens is roughly maybe 750,000 words. So 4 trillion is vast.

[A]

Okay, Vast is the word. So a ton of information packed into this relatively smaller AI brain.

[B]

And apparently it paid off. AI2's benchmarks show it beating some well known models on things like arithmetic reasoning.

[A]

Like math problems.

[B]

Yeah, exactly. And also factual accuracy, like how truthful its answers are on certain tests. Which is, as you said, quite remarkable for its size.

[A]

It really is the performance on benchmarks like GSM8K for reasoning and truthful QA. It's just good data and smart design can go a long way even without billions more parameters.

[B]

Definitely points towards efficiency in design and training.

[A]

But, and this feels really important, AI2 themselves are putting a big warning label on it. They're saying don't use this for commercial stuff yet because it can still produce what they call problematic outputs.

[B]

Yeah, that includes harmful content, sensitive information, or just plain inaccurate stuff.

[A]

That seems like a major caveat.

[B]

It absolutely is. And it's good they're being upfront about it. It highlights this core tension in AI. The capabilities are racing ahead, but making them safe and reliable is still a huge challenge.

[A]

So powerful, accessible, open, but handle with care.

[B]

That sums it up well. It's progress, but responsible progress requires acknowledging the risks. They're basically saying, here's the research, but it needs more Work before you deploy it widely.

[A]

Okay, let's shift gears. How is this tech actually being used now, like in the real world? Airbnb is a good example.

[B]

Right. They've quietly rolled out an AI customer service bot.

[A]

Quietly?

[B]

Yeah. Not a huge splashy announcement initially. CEO Brian Chesky mentioned it on their recent earnings call. It's currently handling interactions for about half of their US users.

[A]

Wow. And the plan is. Full rollout soon.

[B]

This month, apparently. And the early results, according to them, are pretty, pretty significant.

[A]

How so?

[B]

They're reporting a 15% drop in the number of people who ultimately need to contact a live human agent.

[A]

15%? That's actually a big chunk of support requests potentially being handled by the AI.

[B]

It is. It points directly to, you know, operational efficiency gains, fewer human agent hours needed for certain types of queries, and it.

[A]

Seems like Airbnb is being quite specific, focusing just on customer service for now.

[B]

That seems to be the strategy. Yeah. Unlike some competitors like Expedia or booking.com who are reportedly going big on AI for things like building your whole travel itinerary or giving real time updates.

[A]

Right. More complex planning tasks.

[B]

Exactly. Airbnb seems to be starting with the support side. Maybe that reflects their business. So much depends on resolving host or guest issues smoothly.

[A]

That makes sense. Sort the problems out first. It's interesting seeing these different approaches within the travel industry.

[B]

Definitely. Everyone's experimenting, but prioritizing different areas based on their own models and likely where they see the quickest or safest wins.

[A]

And this comes at a time when Airbnb mentioned their growth might be slowing. A bit. Bit of caution about travel demand that.

[B]

Was mentioned in their outlook. Yes. Slower growth expected in the current quarter.

[A]

So finding efficiencies through AI in customer service probably looks pretty appealing right now.

[B]

It's a very logical connection to make. If demand growth slows, managing costs and keeping existing customers happy becomes even more crucial. AI for support could tick both boxes.

[A]

But it does raise that question for us as users.

[B]

Absolutely. How good is the AI bot? Will it handle complex problems well? Or will you end up frustrated just wanting to talk to a person? That quality of interaction is key.

[A]

Yeah. That 15% figure is great for efficiency, but the user experience has to be there too. Okay, let's pivot again. This next one is more philosophical, maybe.

[B]

Okay.

[A]

It's about concerns raised by Kevin Systrom, one of Instagram's co founders. He's worried AI companies are too focused on juicing engagement.

[B]

Ah, right. Using AI to keep you hooked, basically, rather than just being useful.

[A]

Exactly. He criticizes chatbots that ask like endless follow up questions. Not necessarily because they need more info, but just to keep the conversation going.

[B]

Yeah, he draws a direct line to social media tactics, doesn't he? Trying to maximize time spent daily, active users, those kinds of metrics.

[A]

He rings a bell, that feeling. Yeah, you ask a question and instead of straight answer, the bot asks you three more questions. First, he even brought up how ChatGPT got criticized for being too nice or chatty sometimes.

[B]

And OpenAI's response to that, I recall was something about it being due to short term feedback during training.

[A]

Right. But Systrom's take is sharper. He suggests this over engagement might not.

[B]

Be an accident, that it could be an intentional feature designed to inflate those metrics that, you know, look good to investors.

[A]

That's the implication. His main point is AI should focus on giving high quality, direct answers. Stop trying to just keep us talking.

[B]

It's a strong critique. And while he didn't call out specific.

[A]

Companies by name, the target seemed pretty clear. It makes you wonder about the design philosophy behind these AI assistants. Is the goal maximum helpfulness or maximum interaction time?

[B]

It really gets to the heart of the values driving AI development, doesn't it? If engagement becomes the primary goal, does utility suffer? Does accuracy suffer? Systrom's call for quality answers is a pretty crucial point.

[A]

And OpenAI's counterargument is what exactly?

[B]

Their response generally is that the model might genuinely lack information or context and needs to ask clarifying questions. But they they say it should try to answer directly first and only ask for more details if it's truly necessary.

[A]

So a bit of a gray area there. Is it needing info or just trying to chat?

[B]

It's probably a mix, depending on the situation and the specific AI. But Systrom's warning about the incentives driving the design is definitely worth thinking about.

[A]

Okay, Absolutely. Now, for our last story, let's end on something completely different and honestly, just really nice.

[B]

Okay, I'm ready for nice.

[A]

It's about a guy named Mark Edwards in Scotland. He used AI to animate old photos of his hometown, Chedberg.

[B]

Oh, cool. Like making the people move.

[A]

Yeah, bringing these old black and white stills to life. He did it initially just to show his grandparents, George and Isabel.

[B]

That's lovely. How did they react?

[A]

They loved it. And apparently so did everyone else who saw it. The reaction was so positive, he ended up doing more videos for nearby towns like Kelso and Galashiels.

[B]

Wow, that's fantastic. Taking a personal project and sharing it.

[A]

Wider and it's not just, you know, a cool gimmick online. These videos are actually being used.

[B]

How so?

[A]

In local schools for history lessons and also at social groups for older people, like the Opal Project in Galashiels.

[B]

That's brilliant. Using AI to connect generations with their local history in such a visual, engaging way.

[A]

Totally. And the article shared some reactions from people who saw them. One woman, Monica, saw the old Galashiels railway station animated and remembered meeting her dad there.

[B]

Oh, wow. That kind of personal connection is powerful.

[A]

Another person, Anne, just loves seeing the old buildings come alive again. It taps into that nostalgia, that sense of place.

[B]

It really does. It makes history feel immediate and human.

[A]

Mark himself said he was kind of amazed at how well it worked. And the response online, thousands of views, shares, likes. He's even getting requests from other towns and sports clubs now.

[B]

It really struck a chord. It shows AI doesn't just have to be about, you know, productivity or complex problem solving. It can be about emotion, connection, storytelling.

[A]

Yeah. And his grandfather George's reaction was apparently just delight and wonder asking, how does it do that? It's that simple joy of seeing the past in a new light.

[B]

A really heartwarming use of the technology.

[A]

Definitely. So, wrapping this up, it's been a quick tour, but you see the threads, right?

[B]

Yeah, absolutely. We've got these smaller, more accessible AI models pushing boundaries, but they come with important safety warnings.

[A]

Right. Then there's the practical side. AI rolling out in customer service, aiming for efficiency, but raising questions about the actual user experience.

[B]

And that critical perspective from Systrom. Is AI being designed purely for usefulness or our engagement metrics muddying the waters?

[A]

And finally, that lovely example of AI not as a complex tool, but as a way to connect with history and community on an emotional level.

[B]

Taken together, it really paints a picture of AI not as one single thing, but as this collection of evolving tools finding niches everywhere.

[A]

Some technical, some practical, some raising ethical questions, some just. Human?

[B]

Exactly. It's becoming part of the fabric in so many different ways, each with its own implications.

[A]

So maybe the final thought for you, listening in, is which of these developments feels like it will touch your life most directly? Is it the behind the scenes tech? The customer service bots? The way AI interacts with you, or these more creative, connective uses?

[B]

Yeah. And what other questions does it raise for you about where this is all heading? It really makes you think about that balance, doesn't it, between the shiny new innovation and actual practical, maybe even ethical usefulness in our lives.

[A]

Finding that balance in the age of AI a good place to leave it for today.