Generative AI: A Harbinger of Mass Personalization
AI Overview
AI was born from Large Language Models (LLM), algorithms trained on massive sets of text inputs to perform — at their most rudimentary — a “complete the thought” sentence completion exercise, similar to a Gmail prompt. You’ve probably already heard the spiel so I will keep it brief. As these models ingest, they grow stock of human-authored sentences, and refine their facility to respond with unique and accurate language. LLM’s first major breakthrough was GPT-3, whose principles were applied to produce a synonymous natural language-to-image engine DALLE-2, chatbot ChatGPT, language-to-code, image-to-video, 2D-to-3D, and various other models of that ilk.
Given its lightning speed to generate digital assets, the most readily available inference is that AI will bypass the Iron Triangle of project management (namely: cost-quality-speed, pick any two). Layering AI tech across functionality gives rise to tools that enhance (or in some cases, bestow) creative power to anyone — whether or not they had previously trained that technical muscle. I predict that learning to use AI effectively will become a marketable skill (e.g. AI creative designer) once the tools develop the coherence and stylistic understanding for this to be possible.
I could go further into the possibilities for architecture (model creation), gaming (character generation), film (animation) and how much faster / cheaper / better these business processes are about to become. However I want to explore how AI is set to disrupt our experiences as Consumers through Context Layers in what I will call Personalization 3.0. Where Personalization 2.0 was about providing tailored products and services (e.g. shade-matched makeup), Personalization 3.0 is the ability to lucidly identify and opt into and out of different daily augmentations as well as generate those augmentations on demand. This Next Gen Personalization will allow each person to consume a slightly more personalized version of an asset, tuned to their tastes, language, speech style, height, preference for visuals, etc.
Generative AI tech slices across functions and verticals, bringing an era of mass customization to several media point solutions, including: News, Music, Wellness, Education, Shopping, Gaming
News
If anything indicates that traditional news outlets are in need of disruption, it’s this fact: In 2022, 10% of U.S. adults say they get their news from TikTok. “Citizen Journalism” has been battling the old guard for some time. Grassroots channels, such as Twitter, Discord, and TikTok, are quicker to break news and more flavorful in their delivery. Consumers are looking for personalization in what they consume, but in turning to independent journalists, are forgoing accuracy for custom-tailored entertainment.
News media presents a huge opportunity for a more reliable source to step in and perform this function with real reporting. Traditional reporting houses can use an AI layered solution to maintain veracity while augmenting personal delivery flairs to each member of the audience. Apple News already uses simple “Recommended for You” algorithms to filter different content into our newsfeeds based on prior browsing preferences. AI can take this one step further to augment the content to our exact tastes. For instance, summarization, visual / diagrammatic enhancement, and stylistic and vocabulary adjustments. Or even by interpreting our emotions and current mental states to iterate news stories with personal and contextual grounding. The tools that do this have already been developed (e.g. paraphrasing: Quilbot; image-text matching: GoCharlie.AI) but there is still opportunity for a bundled B2B product made specifically for newsrooms.
Music
Music has already been blessed with AI’s Midas Touch. It was raison d’etre for Spotify’s rise to superiority over other subscription streamers (i.e. Apple Music). Being AI-native is a competitive advantage that trumps more primitive moats, such as content exclusivity (remember Jay Z’s Tidal?). Spotify specifically leans on reinforcement learning, a model that uses environmental signals to optimize toward a long-term reward. The carrot that Spotify dangles in front of its ML is long-term user satisfaction. Because the AI is personalized at the individual level, Spotify is not just one product, but 433 million different products — one for each user.
Spotify has been growing it’s function by acquiring AI companies since 2013, when Spotify acquired Tunigo to better power its music recommendation engine. It is easy to envision the company exploring audio- and music-generating technology next. Generative audio can be used to enhance existing music. For example, real-time DJ mixing to alter the style and BPM of an existing melody can effectively replace the work of a SoundCloud creator. These stylistic edits are similar to what Dall-E is able to do with image (when asked to design “in the style of Monet”), but with sound. Another approach to enhancement is a “continue the melody” tool exemplified by AudioLM, Google’s latest system which can extend a clip of piano music fluidly.
Generative AI could also entirely displace the artist as a full stack production tool. Last September, Harmonai — an organization with backing from Stability AI — released Dance Diffusion, an algorithm and set of tools that can generate clips of music by training on hundreds of hours of existing songs. Streaming services have already commoditized music for us. As mentioned before, we no longer pay for artist’s CDs (or pirate their entire discographies) to make them our own. The removal of that ownership has hampered our connection with artists and severed our obligation to listen to each song on an album. In 2021 Adele infamously fought against new listener behaviors when she successfully requested that Spotify remove the ‘shuffle’ button for album listeners.
But despite listening to generic, machine-curated playlists, fans still have para-social relationships with artists and this facet leaves room for monetization. Generative AI can enter to write and produce entire records based on listener desires, then auto-tune them into, say, Billie Eilish’s voice, auto-license the single, then auto-transfer her the royalties. Boomy’s offering as it describes on its intro video is exactly that: “Make music instantly. Pick a style. Release to the world. Get paid”. This posits a bleak forecast for listeners of the future. Artists effectively could become “Aunt Jemima” brands which churn out ever-so-slightly different music for each fan from behind their public façade, without any involvement in production.
Wellness
Health, fitness, and vitals-monitoring have captured a large contingent of consumer attention of late. Online communities across Discord and Reddit gather to share the techniques (supplements, diets, fasting, light therapy, nootropics) they use for managing their health to achieve longevity and optimized daily performance. “Biohacker” groups need data to quantify their health goals and track their progress. Because each body operates in a silo, each individual’s biohacking journey is hampered by a trial and error period. Enter AI, able to read and combine data then “write” your tailored prescription based on what has worked for you: what has made you go faster, longer, harder, etc.
In health, Biobetter is an AI-powered biohacking assistant that provides blood test-based recommendations to increase productivity, slow down the aging process, and prevent diseases. Wild.ai uses A.I.-sourced data sets from active women to analyze vitals and performance to make training and nutritional recommendations adapted to their physiologies.
Generative AI can add a consumer-interactive layer atop the recommendations. For instance, imagine running a marathon connected to your Garmin watch. The software is able to read your heart rate and pace and determine if you are running better or worse than peak performance. Based on readings it can notify you when to drink more water or consume an energy gel. It may hijack your AirPods for music with higher BPM, or use audio-generative function to remix your song into something livelier which historically has pushed you to sprint. The entire physical experience of your race can be governed by wearables, manual inputs, and software that reads and optimizes right-in-time. I can imagine similar applications for other endurance sports or power-lifting.
Mustard is an AI-driven sports coaching app whereby users capture their training on video and the app evaluates the mechanics of their performance and provides them with coaching. It uses a computer vision technique called Human Pose Estimation, which trains algorithms to identify human joint position using keypoint skeleton models and offers guidance on how to improve exercise form. This technology has varied applications which include optimizing form in tennis and swimming or predicting players’ next moves in football.
For the most part, populations have returned to workout classes post-pandemic despite all forecasts for exercise shifting from the studio to the home. Humans like the experiential, motivational aspect of sweating in a room altogether while a trainer repeats inspiring phrases to us (to the point that we’d pay for it). Personalization, social connection and (to an extent) surveillance are important to us. Replicating that requires authenticity that was not captured over Zoom in 2020, but could be replicated with generative AI personal trainers in 2023.
Education
The same “coaching” premise can be extended to education. One-to-one instruction (as opposed to traditional “classroom style” one-to-many) offers deeper learning. Eight-out-of-ten pupils who receive private tutoring get better grades because it allows instructors to adjust the pace of new lessons and shift focus to concepts that require deeper explanation. For students, the benefits (confidence, engagement and performance) are many, but the financial cost is high. This is where tech can offer scale. Cousins of AI and ML have crept into EdTech for this purpose. If you have been on GMAT Reddit, you’ll have seen how highly ranked Target Test Prep’s product is. It’s not AI, but it has attempted to replicate 1:1 tutor attention with analytics and insights to guide the student’s own performance. The GMAT itself is adaptive, meaning that it outputs question difficulty based on it’s test-taker’s demonstrated facility with previous difficulty levels. Lalia is an English-learning platform scaled to the proficiency of each individual learner. Artificial Intelligence has been customizing learning to a degree, but LLM will improve it manifold.
Picture a 3rd Grade math education platform that is fully reactive. Generative AI can tackle syllabus content and course design. The software can ingest real-time feedback such as: student’s performance (struggling with fractions), problem-solving techniques (long division), focus level, and even expression. It can then generate and personalize new learning materials: reorganizing the lesson on converting decimals to fractions until the learner has grasped them, quizzes that begin at their natural capability level and climb, problems tailored to focus areas, generated avatar’s explaining weak spots on video lectures, auto-produced flashcards. The interface can optimize what it outputs against each student’s needs and parameters. TutorAI, which launched this past September, is working its way in this direction.
Shopping
Commerce, while technically not consumer media itself, is still sustained by a variety of media forms, from eComm websites to socials. Advertising is one such pillar. As online spaces inundate us with more and more ads, we have become increasingly discerning when we engage with and tune out the noise. Shifting privacy laws mean that CACs have risen, and companies need to find new ways to stretch a marketing dollar to full efficacy. AI is already in employ for companies optimizing which consumers to target and when in their buying windows (e.g. Aiadvertising, Persado)
Generative AI introduces the possibility of on-demand personalized advertising in a relatively sci-fictive version of our reality (e.g. adcreative.ai). Imagine scrolling through BestBuy.com and finding the camera page come to life and direct you towards the best model for your environmental photography job, plus the best lens for the occasional lowlight zooms you post to your Instagram. Or imagine walking past a digital billboard and seeing yourself reflected in your size and favorite color of the jacket on the model.
This type of real-time triage between data and visual assets might also be the answer to the decade-old virtual fitting room problem. Despite a cohort of strong teams (Truefit, Fitanalytics) and big-name acquisitions in the space (Zeekit by Walmart, Bodylabs by Amazon), the whole category has struggled and failed because of what I think is root tech issues. Generative AI trained against enough footage of users in different garments of known tech packs might finally uncover the capability to “try” clothes on virtually and save the fashion industry $760B in annual returns costs.
Gaming
Games are the most consumed form of media entertainment but also the most complex. A quality game integrates 2D art, 3D art, sound effects, music, dialog, characters, levels and real-time experiences. Game production is a heavy lift so cost and time-to-market have been steep barriers to entry in the space.
With AI assistance in design and asset generation, this barrier is set to drop, sending fresh game developers onto the market and dramatically reducing the price of content. Think about when music became digitized, and we could buy a ton of $0.99 singles on iTunes rather than the few vinyls we could afford. Now think about when Spotify offered us a subscription and we could listen to infinite music. Both affordability and accessibility will impact gaming similarly. As more games are churned out to market, discovery function (currently weak) will improve, and as a result personalization will take off. The stories, avatars, voices and worlds visited in games will be increasingly tailored to their player. Whether that’s through spoken language, dialect, character appearance, world features, or rendering style.
—
Are you building something in this space? I work with early-stage Retail and Consumer startups and would love to hear from you! My DMs are open on Linkedin or Twitter @aylajeiroudi
