Searching for the first great AI app | The Vergecast

Intro (0s)

• The Vergecast is the flagship podcast of the "agentic era", where AI is a major topic of discussion, but it's clarified that many AI companies, including Tesla, rely on human operators in the background to function (13s).
• The podcast hosts, including Nei and David Pierce, discuss various topics, including AI and its applications, with David being asked to memorize a list of names and create new ones (28s).
• Richard Lawler, a self-proclaimed "big AI guy", joins the discussion, and it's joked that he hasn't had AI change a car's brand, but rather, whatever the computer does is considered AI (48s).
• The hosts poke fun at the idea that AI is being defined loosely, with even simple tasks being attributed to AI, and joke that Formula 1 (F1) driver Lewis Hamilton's car is always a Ferrari, regardless of the actual brand (1m1s).
• The conversation touches on the topic of F1 sponsorship, with the hosts noting that there isn't much "dumb AI money" in F1 yet, but new sponsorship opportunities are arising with the new season (1m28s).
• The hosts praise Chrome's sponsorship of the McLaren F1 team, which included adding actual Chrome to the car, as a clever branding move (1m59s).

Google's AI Innovations: Gemini 2.0 (2m10s)

• There's been a lot of AI news recently, particularly from Google, including the release of Gemini 2.0, mixed reality developments, and a new AI for code, Jewels (2m10s).
• Gemini 2.0 is the successor to Gemini 1.5, which was first launched in February, and represents a significant leap in power and efficiency, with roughly the same performance as the Gemini 1.5 Pro tier (3m6s).
• According to Demis Hassabis, who runs DeepMind at Google, Gemini 2.0 is a full step up in terms of efficiency, speed, and latency, making it a major improvement over its predecessor (3m49s).
• Gemini 2.0 has a model called "flash" that is designed to run on phones and is the most commonly encountered model, with a smaller model called "Nano" that runs on actual phones (4m15s).
• Gemini 2.0 can now natively do images and audio, which were previously handled by separate models, making it a more multimodal AI system (5m26s).
• Google has also made announcements about mixed reality, including an Android mixed reality platform, although this is not a new area of development for the company (2m23s).
• Other Google AI developments include Astra and Mariner, which are products that overlap with Gemini 2.0 but are distinct and worth discussing (2m44s).
• Google's Gemini model is expected to be integrated into various Google services such as search, Gmail, and AI overviews, and will also be sold to Cloud customers, marking a shift in the company's approach to AI (5m43s).
• Unlike Open AI, which has multiple models for different tasks, Google is focusing on a single model, Gemini, to be used across various applications (6m2s).
• There is an ongoing debate in the AI industry about scaling laws and whether training larger models will result in more capability, with some arguing that the concept of Artificial General Intelligence (AGI) is being redefined (6m26s).
• Google's newest model has the same capability as previous models but is more efficient, which is seen as important but overshadowed by the AGI conversation (6m52s).
• According to Demis, Google's stance is that there is still headroom for improvement in AI models, but it will require new techniques rather than just increasing the size of the models (7m21s).
• Demis believes that new techniques, such as the Transformer model, are needed to unlock the next step change in AI, and that Google is working on developing these new techniques (7m47s).
• Both Google and Open AI are currently focused on figuring out how to turn their AI models into useful and valuable products that can be monetized (8m10s).
• Google's approach to AI is to work on the technological and infrastructural level, and then use these developments to improve its own services, with Google being its own biggest customer (8m42s).
• Google's Gemini and Gmail are expensive to operate, making efficiency and lower latency crucial for the company, not only to compete with other cloud companies but also to integrate AI into its own products without incurring significant costs (9m2s).
• The focus has shifted from making AI models more capable to making them cheaper and more efficient, allowing for wider integration into products (9m56s).
• The cost reduction is seen as a significant achievement, but it's noted that the AI models are not necessarily better, just cheaper to run (10m4s).
• Google could potentially save money by turning off AI overviews on search results, which might not be a highly valued feature by users (10m35s).
• Google announced that over a billion people are experiencing AI overviews, but it's suggested that this might be due to the feature being forced upon users rather than being a highly sought-after feature (10m51s).
• The conversation has moved away from discussing the potential risks and capabilities of AI models and towards finding ways to monetize them by integrating them into more products (11m25s).
• The products that incorporate AI are not yet compelling, and it's likened to the situation with Bluetooth, where the technology is touted as great, but the products that use it are not necessarily good (12m3s).
• Google has launched four new pairs of headphones, including Project Astra, but the quality and appeal of these products are not discussed in detail (12m12s).
• Google is working on an ambitious AI project that combines visual, listening, and memory capabilities, allowing users to interact with it in various ways, such as asking where they left their glasses, and the AI can respond with the location using its camera capabilities (12m25s).
• This AI project is connected to other Google services like Google Maps, making it more powerful and useful (12m46s).
• The AI is built on top of Gemini, which enables it to integrate with various services and make it more doable (12m53s).
• Project Mariner is a Chrome extension that acts as an agent, browsing the web and performing tasks for users, but it's still in its prototype phase and has limitations (13m3s).
• Project Mariner can perform tasks like finding contact emails on web pages, but it's slow and wonky, and its capabilities are not significantly better than manual browsing (13m30s).
• The requests made to Project Mariner can be absurd, such as memorizing a list, which is a basic computer function, and the results are often unreliable (14m1s).
• The AI's limitations and unreliability can lead to frustration, making users question the value of using AI for certain tasks (14m36s).
• Project Astra's demo video showcases an AI assistant helping a user get into an apartment building, but the demo highlights the AI's limitations and the user's need to remind the AI of previous conversations (14m49s).
• A feature that allows users to pin notes on their phone screen has been discussed, but its relation to AI and its actual accomplishments are unclear (15m38s).
• The feature of hacking into a web browser on behalf of the user is not considered impressive or interesting, and it is suggested that more useful features should be developed (15m58s).
• Apple's iOS 18.2 release includes a feature called "Visual Intelligence" that allows users to point their camera at an object and receive information about it (16m11s).
• Google's Astro feature is seen as the future of Google Lens, allowing users to look at objects and have a conversation about them, but it also raises concerns about the accuracy of information and potential biases (16m29s).
• The "US capital problem" is mentioned, referring to the challenge of providing accurate and unbiased information about sensitive topics, and it is noted that none of the companies have fully addressed this issue (16m51s).
• The potential use cases for Astro and similar features are discussed, including using augmented reality glasses to receive information about objects and environments (17m25s).
• The idea of living in a "custom political reality" is mentioned, where people can choose to see information that aligns with their own biases and perspectives (17m30s).
• The potential for Astro and similar features to raise complex and loaded questions is discussed, and it is suggested that these questions may lead to cultural wars and disagreements (17m43s).
• The idea of using technology to augment everyday life is seen as positive, but it also raises concerns about the potential consequences and challenges of providing accurate and unbiased information (18m12s).
• The concept of augmented reality is discussed, and it is noted that there are two separate pieces to this technology: the logistical aspect of using it to navigate everyday life, and the more complex aspect of providing accurate and unbiased information (18m27s).
• The potential usefulness of Astro and similar features for navigating everyday life is acknowledged, but concerns are also raised about the potential consequences and challenges of providing accurate and unbiased information (18m37s).

Google's XR Vision (32m23s)

• Android XR is a new platform from Google, similar to how Android was for smartphones, but for smart glasses, offering immersive versions of Google Apps and Gemini, an AI assistant. (32m23s)
• Google has been working on smart glasses for 15 years, including the forgotten Daydream project, and has now announced a developer preview of Android XR. (32m42s)
• The platform will have hardware partners, with Samsung expected to be the first to release a real Android pair of glasses next year, while Apple is working on its own Vision Pro. (33m30s)
• Google's role in making hardware for Android XR is uncertain, but the company has made prototypes and is interested in developing them further. (33m49s)
• The future of AI and augmented reality may rely on smart glasses, with companies racing to develop this technology and answer questions about who gets to augment reality and whether everyone experiences the same one. (34m4s)
• The development of smart glasses raises concerns about the potential for different realities and biases, depending on the hardware and software used, such as Android on Lenovo versus Samsung glasses. (35m14s)
• The question of who gets to control and shape augmented reality is a deep and complex one, and companies may not be fully ready to address it. (35m6s)
• The Washington Post is a trustworthy source for news, offering substantial and considered coverage on various topics, including Capital Hill, the economy, and climate change, and is available for subscription at a discounted rate of 50 cents per week for the first year (35m41s).
• The Washington Post's clean homepage allows users to quickly catch up on the news, making it a great place to find thoughtful stories on a wide array of topics (36m18s).
• Life 360's tile trackers can help eliminate stress by ensuring users know where their items are at all times, and the device features an SOS button that can discreetly send an alert to emergency contacts (36m44s).
• The tile tracker is the first and only tracker with an SOS button, and it does not alert thieves to its presence, making it a reliable option for keeping personal items safe (37m11s).
• Stripe is a payment management software that handles the complexity of financial infrastructure, offering a seamless experience for business owners and their customers, and is used by companies like Uber, BMW, and DoorDash (37m43s).
• Stripe's platform supports millions of businesses around the world, offering a suite of specialized features and tools to power businesses of all sizes, including stripe billing, which makes it easy to handle subscription-based charges and invoicing (38m20s).
• Stripe helps companies of all sizes make progress on their missions and reach more customers globally, and more information can be found at stripe.com (38m48s).
• The latest iOS 18.2 update is out, featuring visual intelligence and chat GPT integration, with the update available for iPhone 16 and up, or iPhone 15 Pro, offering features such as gen Emoji and the ability to invoke chat GPT by using specific phrases (39m50s).
• The update also allows users to change default apps for various tasks, such as tapping on a phone number, which can be customized to open in a different app, a feature that Android has had for a longer time (40m41s).
• The chat GPT integration is interesting, as it tends to punt to chat GPT more often when asked compound questions, providing a relatively consistent experience, and the responses still look like they are from Siri (41m7s).
• Joanna Stern wrote a column about the chat GPT integration, highlighting the humor in invoking chat GPT by using specific phrases, such as asking Siri to write a list (40m11s).
• The ability to change default apps is a significant improvement, making the phone more user-friendly, and is a result of pressure from EU regulators (40m51s).
• The update also includes other features, such as the ability to get a chocolate chip cookie recipe that tastes different from the usual one, by asking a compound question that punts to chat GPT (41m23s).

Salty Moods and Genmoji Musings (41m53s)

• The mental construct of interacting with Siri can be confusing, but the actual user experience is pretty good, providing a way to get better answers from Siri (41m53s).
• The idea of repeatedly sending information to ChatGPT may cause friction, but this friction is expected to decrease over time (42m16s).
• Mapping the action button on a phone to a voice assistant, such as ChaCha BT, can provide a faster way to access information, similar to the new Siri functionality (42m31s).
• The new Siri animation is considered overwrought and annoying, taking over the entire phone screen, which is not preferred by some users (42m45s).
• The previous Siri animation, featuring a little swirling ball, is missed by some users, who find the new animation too flashy and unnecessary (42m50s).

AI Image Generation: Meme or Product? (43m5s)

• The old Siri interface used a spinning wheel to obfuscate its slow loading times, making it appear faster than it actually was, similar to how progress bars are designed to move quickly at first and then slow down (43m17s).
• The new Siri interface is more transparent about its loading times, but it uses glowing visuals to make it seem like something is happening at all times, which can be seen as desperate design on top of a slow system (43m50s).
• AI is particularly good at natural language capabilities, and Siri's understanding of language has improved significantly (44m11s).
• The integration of Siri with other features like CHP is seen as a basic requirement, rather than a standout feature (44m30s).
• The ability of Siri to interact with images through features like Image Playground, Gen Emoji, and Visual Intelligence is seen as a significant development, allowing the phone to become multimodal and communicate through pictures (44m41s).
• Gen Emoji is considered a silly feature, while Image Playground is seen as being 50 steps behind, but this may be a safer approach to implementing the technology (44m52s).

Apple's AI Ambitions and Ad Dilemmas (45m1s)

• Apple's AI ambitions and ad dilemmas are being explored, with a focus on the company's image generation capabilities, such as Gen Emoji, which allows users to create custom emojis (45m1s).
• The effectiveness and practicality of these features are being questioned, with some users finding them to be more of a novelty or meme rather than a useful product (46m32s).
• The use of AI-generated images and emojis is not seen as a necessary or desirable feature by some, who prefer to use existing emojis to convey their emotions (47m21s).
• The idea of generating custom emojis to express complex or nuanced emotions is not seen as a pressing need, with some users feeling that existing emojis are sufficient (47m31s).
• The concept of a "zebra of Many Colors" with anatomical incorrectness is proposed as a humorous example of a custom emoji that could be generated using Apple's AI technology (47m35s).
• Listeners with access to Apple's AI technology are encouraged to experiment with generating custom emojis and sharing the results (47m50s).

Sora: The Fascinating AI Experiment (48m10s)

• Apple's AI capabilities make the phone seem more capable, even if it's not actually more capable, and this is reflected in their ads, which can come across as demeaning to the user (48m20s).
• The ads have a problem in that they need to find a scenario where the AI is useful but not so important that the user needs to write the message themselves (48m48s).
• The ads are attention-grabbing but can be off-putting, and the phone's increased capabilities are not as groundbreaking as the "ChatGBT moment" (49m11s).
• The next significant development in AI will likely be agents, which Apple is also working on, and these agents will need access to apps to be truly useful (49m42s).
• For agents to be successful, app developers will need to build features that allow the agents to access and interact with their apps, but this may require financial incentives (50m26s).
• Some companies, like Netflix, may be able to opt out of integrating their services with agents if they have a large enough audience, but smaller companies may not have this option (51m4s).
• The development of agents and their integration with apps will be a crucial factor in the future of AI, and companies are working to build these systems and make them useful and accessible (51m14s).
• Sora is a fascinating AI app that is "super good enough" to be interesting, despite not appearing to be great, and its bizarre nature is due to its ability to create unusual content, such as a person doing gymnastics whose body explodes in a thousand directions every time they do a flip (51m28s).
• The app was quickly overwhelmed with sign-ups and stopped accepting new users due to capacity issues, with Sam Altman stating that they underestimated the demand and will reopen sign-ups at some point (52m29s).
• To use Sora, users need to have either the $20 or $200 ChatGPT Pro Plan, with the latter offering the ability to create 1080p videos up to 20 seconds long, while the former is limited to 720p videos up to 5 seconds long (52m46s).
• The app has raised concerns about its potential misuse, with the ability to create realistic but fake videos that could be used to destabilize a country, and the fact that it has clearly trained on YouTube videos without permission (53m34s).
• Marquez Brownlee's video about Sora highlighted the app's ability to create realistic tech review videos, including one that featured his own plant, demonstrating the app's training on his videos without permission (53m25s).
• The app's training data has raised concerns about copyright infringement and the potential for lawsuits, with some speculating that Google or YouTubers may follow up with a lawsuit (53m44s).
• A key difference between the $20 and $200 ChatGPT subscription plans is that the latter allows users to download videos without watermarks, which could be used to create more convincing fake videos (54m7s).
• Marquez Brownlee also noted that Sora's generated footage is often so bad that people would assume it's real, and intentionally generated low-quality footage that could be mistaken for real CCTV footage (54m23s).
• Open AI's video generation tool, Sor, will include visible watermarks and C2PA metadata to indicate AI-generated content, with the goal of promoting content authenticity and saving civilization from the potential negative impacts of AI-generated misinformation (55m5s).
• The C2PA (Content Authenticity Initiative) is a metadata standard aimed at verifying the authenticity of digital content, but its effectiveness is disputed, with some critics arguing it will not work as intended due to various reasons (55m18s).
• Despite the controversy, C2PA is currently the only available solution, and major companies like Google, Adobe, and Open AI are using it, while platforms like YouTube, Facebook, and TikTok are expected to display this metadata, although not all have agreed to do so (55m51s).
• Apple has not chosen to use C2PA on the creation side, and other platforms are building their own AI tools, which may not have an incentive to point out AI-generated content, potentially leading to a lack of transparency (56m11s).
• The internet may become fragmented into closed platforms that guarantee content from real people and open platforms where AI-generated content is prevalent, which could have both positive and negative consequences (56m32s).
• The audience split between these two types of platforms may not be what is hoped for, with the closed platforms potentially being expensive and having a small audience, while the open platforms may have a larger audience but also more AI-generated content (57m16s).
• The proliferation of AI-generated content is already causing confusion and destabilizing the information economy, with some users questioning the authenticity of videos on platforms like TikTok (57m30s).
• Other AI-related projects, such as Chad's Canvas, are also being developed, but their specifics and implications are not yet clear (57m50s).

Reddit Answers: The Best AI Feature? (57m55s)

• Canvas is a feature that allows for interactive and real-time editing with a chatbot, where the bot generates text in a column next to the conversation, and the user can make changes to it, making it a more interactive experience than just blocks of messages (57m55s).
• This feature is a trend among chatbots, as they try to add UI elements and features that don't feel like traditional messaging systems, but still work within that construct (58m43s).
• The idea of Canvas is seen as a good one, whether it's used for writing or coding, and has been tested for a while, with positive feedback from users (59m2s).
• The development of chatbots has reached a point of diminishing returns, where the underlying technology will continue to improve, but at a much slower rate than before (1h0m1s).
• As a result, the focus is shifting from novelty to finding practical uses for the technology, and addressing the question of whether it's actually good for anything (1h0m25s).
• The current state of AI is characterized by disillusionment with its capabilities and limitations, and the need to find new ways to utilize the existing technology (1h0m50s).
• The question now is what else can be done with the current technology, assuming it won't improve dramatically in the near future (1h1m5s).
• One example of a useful application of AI is a feature built by Reddit, which is considered one of the most useful things seen with AI yet (1h1m11s).
• Reddit Answers is a feature that allows users to search Reddit without going through Google, and it uses AI to summarize threads, providing a potential solution for quickly finding answers to specific questions on the platform (1h1m14s).
• The feature has the potential to compile information from various Reddit threads and provide a concise answer, making it a valuable tool for users (1h1m40s).
• Examples of how the feature can be used include finding out what baby gear Reddit users recommend or tips for flying with a baby for the first time (1h2m0s).
• The feature provides a summarized answer based on the collective knowledge of Reddit users, but the quality of the answer may vary (1h2m11s).
• A comparison with Google search results shows that the latter provides more useful and detailed information, such as specific tips for flying with a baby (1h3m12s).
• Personal experiences and anecdotes, such as bringing snacks and using Post-it notes to keep a baby occupied during a flight, can be more valuable than the summarized answers provided by the feature (1h3m51s).
• The feature has the potential to be improved and provide more accurate and useful information to users (1h2m15s).
• The idea of using AI to compile and summarize information from Reddit threads is promising, but the execution may need to be refined to provide better results (1h1m28s).

YouTube's TV Growth and Instagram's New Feature (1h4m3s)

• The world's body of information, including personal experiences on platforms like Reddit, can be useful for certain types of queries where there isn't a definitive answer, but trying to summarize these experiences into a single "right" answer can be problematic (1h4m4s).
• The issue with AI-powered notification summaries is that they often try to extract further value from already well-written and concise posts, which can result in uninteresting summaries (1h4m22s).
• This approach is an example of the "lowest common denominator" of AI, where complex information is reduced to simplistic summaries (1h4m53s).
• Some AI models, like Google's Notebook LM, take a different approach by providing users with a list of relevant sources rather than trying to summarize the information into a few sentences (1h5m8s).
• This approach is more effective for platforms like Reddit, where users often want to read a variety of perspectives and experiences rather than a single summary (1h5m16s).
• Reddit's vast and diverse community makes it an ideal platform for AI-powered tools that can help users navigate and find relevant information (1h5m46s).
• The value of Reddit lies in its ability to provide a sense of community and validation, where users can share their experiences and connect with others who have similar interests or experiences (1h6m18s).
• The goal of AI-powered tools should be to enhance this experience, rather than trying to simplify or summarize it (1h6m24s).
• The pstar 3, a luxury electric SUV, is an example of a product that integrates technology in a way that is seamless and intuitive, with features like an aerodynamic exterior, Scandinavian minimalist interior, and optimized powertrain (1h6m51s).
• The Porsche 3 has a lot to offer, providing spacious comfort and agile handling, making it a great option for drivers who don't want to choose between the two, and it can be test-driven at a local Porsche space (1h7m33s).
• Ipsy is a beauty subscription service that delivers personalized beauty products to customers' doors, offering monthly beauty subscriptions such as Glam Bag and BoxyCharm, which include products from top brands like Benefit, Elf, and Bare Minerals (1h7m51s).
• Ipsy's Glam Bag is a curated selection of makeup, skincare, and hair care products, worth more than the subscription cost, providing a high-end beauty experience without the high price tag (1h8m6s).
• Ipsy is a great gift option, allowing the recipient to redeem the subscription whenever they choose and potentially discover new favorite products (1h8m34s).
• Amazon Q is a generative AI assistant from AWS that helps streamline work, allowing tasks like summarizing monthly results to be done quickly (1h8m59s).
• The lightning round is presented by Amazon Q, and the topic of discussion is YouTube's growth on TVs and living rooms (1h9m29s).
• YouTube is still growing fast on TVs and living rooms, with 400 million hours a month of podcast viewing happening on TVs, and sports content is also seeing significant growth (1h10m34s).
• YouTube has launched a new feature called "Watch with," which allows creators to combine live commentary with sporting events, providing a new way for viewers to engage with content (1h10m47s).
• YouTube is introducing a feature that allows users to create their own commentary over top of something they're watching, starting with sports, and potentially expanding to other areas like Apple keynotes, with multiple commentary streams available (1h11m2s).
• This feature is part of YouTube's shift towards becoming a primary platform for TVs, with the company building features specifically for TV, such as the "shows" feature, which allows users to put up a bunch of videos in seasons and episodes, making it feel more like a streaming service (1h11m36s).
• YouTube is also working on syncing phones with TVs, and is pushing to be a more premium, higher-end, and TV-centric platform, while trying not to kill YouTube as a product elsewhere (1h12m11s).
• There is a tension between YouTube's desire to be a premium platform and its existing product, which is fascinating to observe (1h12m29s).
• YouTube has disabled links on its videos that allow users to click the title and go to YouTube, and instead requires publishers to choose between reenabling the links and making less money, or not using YouTube at all (1h13m12s).
• This change was made by YouTube to "remove their branding" and was implemented at the start of the year, despite previously working fine since 2016 (1h14m3s).
• The Verge is a pro-links website and would prefer to have the links reenabled, but is not willing to make less money to do so (1h13m9s).
• The decision to disable the links has been met with frustration, as it may lead to a return to bespoke video players, which are often of poor quality (1h13m41s).
• Instagram is testing a feature that allows creators to test their reels on people who don't follow them, to gauge performance before publishing to their followers (1h14m22s).
• This feature is seen as a way for creators to focus group their content and optimize it for reach, rather than creating for their existing audience (1h14m56s).
• This approach is viewed as a commercialization of creative work, where the goal is to perform well in the algorithm rather than creating art for its own sake (1h17m4s).
• The feature is seen as a reflection of how Instagram and its CEO, Adam Mosseri, view the world, prioritizing reach and engagement over community and creative expression (1h15m36s).
• The idea that creators can't trust their followers to judge the quality of their content is seen as a weird way to think about things, and may lead to a permanent downranking in the algorithm (1h16m44s).
• The feature is also seen as a way for Instagram to further prioritize content that performs well with strangers, rather than content that resonates with a creator's existing community (1h16m13s).
• The commercialization of creator platforms is seen as a broader trend, where the goal is to make money rather than create art (1h17m20s).
• Instagram's algorithm is more creator-friendly, allowing users to build a following and understand their audience, but it also puts creators at risk of losing everything if they make two videos in a row that their audience doesn't like, as the platform cares more about who follows and is followed by the creator (1h17m45s).
• Adam Mosseri, the head of Instagram, is giving creators more tools to prevent this from happening, which is seen as a response to the pressure of constantly producing content that resonates with their audience (1h18m35s).
• TikTok lost a court case challenging a law that would force it to either ban itself or be sold, with the court ruling that Congress made a national security decision that the court would not override (1h18m43s).
• The case raises complicated First Amendment questions that may be addressed in a future episode (1h18m57s).
• Donald Trump, the incoming president, was asked about TikTok and stated that he used the platform successfully in his campaign, but also expressed concerns about Chinese companies owning the app (1h19m11s).
• Trump suggested that if TikTok is banned, another platform will take its place, and expressed his desire to prevent Facebook from becoming an even bigger monopoly (1h19m35s).
• Trump's comments are seen as a walk back from his previous statements about protecting TikTok, and instead focus on preventing Facebook from gaining more power (1h20m8s).
• The possibility of a company like Amazon, Walmart, or Subway buying TikTok is seen as a potential outcome, as a way for Trump to negotiate a deal that would allow TikTok to continue operating in the US (1h20m22s).