In search of the perfect movie recommendation | The Vergecast

The Problem with Movie Recommendations

• David Pierce, host of The Vergecast, discusses the difficulty of finding a good movie to watch on streaming services. He notes that despite personalized recommendations, the options often feel overwhelming and not truly tailored to his preferences.
• He highlights the lack of truly personalized recommendations on platforms like Netflix, Hulu, and Max, where the suggestions often feel generic or based on popular trends rather than individual tastes.

AI as a Solution

• Pierce believes that AI could be a solution to this problem, as it could analyze user preferences and provide more accurate and personalized recommendations. He acknowledges that AI systems sometimes make up movies, but argues that this is a minor issue compared to the frustration of endless scrolling.
• He mentions that chatGPT and other AI systems can already provide movie recommendations based on specific criteria, such as genre or similar movies. He gives examples of using chatGPT to find movies similar to "Mission: Impossible" and "Crazy Stupid Love."
• Pierce also mentions Google's Gemini as another AI system capable of providing movie recommendations, demonstrating its ability to suggest classic 1980s action movies.

The Power of AI in Movie Recommendations

• AI-powered movie recommendation tools can process and organize vast amounts of data, including genres, actor names, and user feedback from social media platforms.
• AI can structure this data into a usable database, allowing users to make specific requests, such as "a movie with a dog that's not sad."
• Diana Piscu, a developer and entrepreneur, believes that AI can revolutionize movie recommendations by analyzing data that wouldn't otherwise be accessible.
• There are four types of information about movies and shows: metadata (basic information like title, actors, director), watch data (how a user interacted with the content), reviews (user feedback), and social media data (comments, tweets, etc.).
• AI can analyze all of this data to provide more personalized and accurate recommendations than traditional methods.

Challenges in Movie Recommendation

• Streaming services lack access to crucial data about how viewers interact with content, such as whether they finish a show, watch it multiple times, or abandon it halfway through. This information is valuable for improving recommendations but is not readily available at scale.
• The amount of content available for recommendation is limited, even for large streaming services. While platforms like IMDb have a vast library, it pales in comparison to the ever-growing content on platforms like TikTok, YouTube, and Instagram Reels.
• Understanding the "traits" or deeper characteristics of a movie or show beyond basic genres is crucial for effective recommendations. This includes aspects like mood, pacing, soundtrack quality, and suitability for different viewing contexts.

Advanced AI for Movie Recommendations

• Spotify's recommendation system, as described by Gustaf Söderström, demonstrates how machine listening and knowledge graphs can be used to analyze and categorize content, enabling more accurate and nuanced recommendations.
• Spotify's system utilizes both traditional collaborative filtering techniques (based on user behavior) and advanced machine learning techniques that analyze content itself, creating a more comprehensive understanding of user preferences and content characteristics.
• The text discusses the importance of understanding content at a deeper level for accurate movie recommendations.
• It highlights the recent breakthrough of Google's Gemini 1.5 AI model, which boasts a large context window, allowing it to process vast amounts of information, like an entire movie, at once.
• The text showcases two demos of Gemini 1.5: one where it identifies a specific moment in a Buster Keaton movie based on a textual description, and another where it finds a scene in the movie based on a simple drawing.
• The text acknowledges that while these demos are impressive, they are still early stages and the technology faces challenges like copyright issues.

The Complexity of Movie Preferences

• The text then delves into the complex question of what constitutes a good recommendation, emphasizing the difficulty in understanding why people like what they like.
• It uses the example of "The Crown" to illustrate how recommendations can be inaccurate because they fail to capture intangible aspects of a viewer's preferences.

Real Good: A Company Focused on Movie Recommendations

• The text introduces David Sanderson, CEO of Real Good, a company providing data about TV shows and movies, highlighting the lack of a universal source for such information.
• Real Good is a company that provides movie recommendations, but it doesn't have access to the same amount of data as streaming services like Netflix.
• Real Good uses AI to analyze movies and shows and predict whether a user will enjoy them.
• The AI considers factors like genre, storyline, and audience scores, but it also tries to understand the underlying themes and archetypes of the content.
• Pablo Alesia, who runs engineering and data for Real Good, believes that AI is not yet capable of fully understanding movies and shows in the same way that humans do.
• Real Good is working on improving its AI capabilities, but it faces challenges in accessing watch data and understanding the complex nuances of human taste.

Limitations of AI in Movie Recommendations

• The speaker discusses the limitations of AI in providing movie recommendations, highlighting that AI models lack the context and emotional understanding that humans possess when choosing a movie.
• Pablo, a person mentioned in the video, argues that AI's ability to analyze minute details like a bird appearing in a movie at a specific time is irrelevant to the user's enjoyment of the film.

A Two-Level Approach to Movie Recommendations

• The speaker proposes a two-level approach to movie recommendations: the first level focuses on genre and tags, while the second level considers the mood of the content and the viewer's mood at the time of watching.
• The speaker emphasizes that mood plays a significant role in movie selection, as people's preferences change based on their current emotional state.

AI's Ability to Understand Complex Requests

• Diana Piscu, mentioned in the video, describes two ways people use the AI movie recommendation service "Movie Vanders": finding a specific movie they can't remember the name of and receiving recommendations based on desired elements.
• The speaker is impressed by the ability of AI to understand and respond to complex requests, such as finding movies with similar plots but different characters.
• The speaker believes that AI can analyze reviews, synopses, and tweets to identify movies with similar themes and styles to a given movie.

Personal Experience with AI Movie Recommendations

• The speaker has been using AI tools for movie recommendations for over a year and has found that they can be effective, especially when used with specific prompts.
• The speaker suggests using prompts like "show me stuff like" with a list of movies they enjoy, or asking for "underrated" or "lesser-known" titles.
• The speaker also recommends asking for specific vibes, such as "a good romcom that's ideally under 90 minutes and doesn't require much brain power."
• The speaker acknowledges that AI recommendations are not always perfect and may sometimes suggest bad movies, but they still find the process helpful for finding new things to watch.
• The speaker appreciates that AI recommendations are low-stakes, as they can simply ask for another suggestion if they don't like the first one.
• The speaker believes that AI recommendations have helped them to reduce aimless browsing and find more enjoyable movies.

The Best Way to Get Good Movie Recommendations

• The best way to get good movie recommendations is to watch as much as possible on as few streaming services as possible.
• This is because streaming services use your watch history to predict what you'll like next.
• If you want to get good recommendations from services like Real Good, Just Watch, or Letterboxd, you need to keep track of everything you've watched.
• While AI may eventually be able to understand us and movies well enough to give perfect recommendations, that technology is still a ways off.
• For now, the best way to get good recommendations from streaming services is to watch things you like all the way through, over and over again.