By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
a16zBeyond Avatars: How AI is Reshaping Online Identity (Danny Postma and Sinead Bovell)
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 3:51
From Lil Miquela to the AI Wave: Setting Up the Episode
Steph Smith frames how AI tools like Midjourney and Stable Diffusion have opened a second wave of digital influencers after CGI-built Lil Miquela. She introduces futurist and model Sinead Bovell and indie builder Danny Postma and previews the topics ahead.
- •Midjourney and Stable Diffusion did not exist a year earlier
- •Lil Miquela was built with CGI in 2016, while AI tools now let almost anyone build a digital influencer
- •Danny has launched seven projects in a few months, including tattoos, memes, headshots and a virtual modeling agency
- •Topics ahead include deflation, what the tools unlock, and building a moat
- 3:51 – 8:17
Sinead's Prediction That AI Would Take Modeling Jobs
Sinead explains why she wrote her 2020 article on AI replacing models: deep fake technology, e-commerce fashion and early research by companies like Zalando pointed that way. She describes the shocked reaction from fashion, advertising and photography, and how people struggled to picture it until the recent generative AI explosion.
- •Sinead tracked trends and saw adversarial networks (deep fakes) as relevant to fashion
- •E-commerce modeling involves small, somewhat robotic variations in poses
- •Zalando was researching the technology and Lil Miquela was a big influencer at the time
- •The article drew responses from modeling, advertising and photography
- •The generative AI explosion has made the idea concrete for people
- 8:17 – 10:58
Deep Agency and the Technology Behind Virtual Models
Danny describes how his Deep Agency tweet went viral and how research tracks are converging in 2023. He explains DreamBooth, StyleGAN and clothing-fitting models, which together can produce a non-existent model wearing real garments.
- •The Deep Agency tweet reached about 30 million views before the product was ready
- •Danny says other tools are better than his as a modeling agency
- •DreamBooth clones a face, and StyleGAN generates faces that do not exist
- •Combining them lets him train a model on a non-existent person
- •Other models can drape top-down clothing photos onto a body, and a company that understands fashion could take it further
- 10:58 – 12:36
How Fashion Companies Are Adopting AI
Sinead says adoption varies by company type, with e-commerce players like Zalando and H&M already having the infrastructure. High fashion may resist more, particularly for design. She suggests viewers may already have seen AI models while shopping online.
- •Zalando showed proof of concept as early as 2019
- •H&M uses data scientists and AI for trend forecasting and supply chain analysis
- •More resistance is expected in high fashion, especially for AI design
- •Companies invested in the technology in 2018, so it is largely here but not yet at scale
- 12:36 – 15:29
How Real Are Fashion Images Already?
Steph asks how real today's photos are given Photoshop and styling. Sinead discusses diversity in representation, pinning and retouching, and a growing cultural push against Photoshopping at the same time as imagery becomes less human. Steph notes AI can be prompted for exactly what you want.
- •Representation in fashion is not very diverse, which limits how useful clothing on a model is
- •Shoots involve pinning, stitching and Photoshop
- •Photoshopping is becoming more taboo as people want to see stretch marks, cellulite and acne
- •AI can be prompted for any look, while the meaning of real will evolve
- 15:29 – 17:13
The Cost of an AI Model and Democratizing Photoshoots
Danny breaks down costs: about 60 cents to train a model on a face and a fraction of a cent per photo. He says the main demand comes from small clothing shops that could never hire models. Steph relates it to her own use of Midjourney for presentation images.
- •Training on a face costs about 60 cents
- •Hundreds of thousands of photos could be generated for under $100
- •Mom-and-pop clothing brands are the top source of live chat questions
- •The tools democratize access for small businesses
- 17:13 – 19:46
Benefits, Deflation and Ethical Gray Areas for Models
Steph asks how models might benefit, such as avatars attending shoots. Sinead agrees a model could be in multiple places and earn more, but says technology tends to be deflationary for wages and raises ethical questions about automated likenesses and communities.
- •A model's avatar could be in multiple places at once for more income
- •Technology is historically deflationary, especially for wages
- •More supply of human-passing identities could push prices down
- •Downward wage pressure will span fields from modeling to programming
- •Modeling and acting uniquely involve automating likeness and represented communities, with someone profiting from that identity
- 19:46 – 23:26
Danny's Projects and What Users Respond To
Steph lists Danny's projects, and Danny explains why they spread: the results feel like magic and solve real problems, like people lacking good photos. Headshot Pro has been his biggest success, especially for remote teams, and he sees AI enabling new uses rather than just replacing photographers.
- •Projects include Tattoos AI, Profile Picture AI, Deep Agency, MemeMorph and Headshot Pro
- •Uploading a few photos and getting hundreds back drives sharing on Instagram and TikTok
- •Headshot Pro serves people who dislike photoshoots and remote teams wanting consistent style
- •Lensa reportedly made $40 million
- •Headshot Pro's counter went from about 450,000 to roughly a million headshots in two weeks
- 23:26 – 25:03
Willingness to Pay and the Pace of New Models
Steph asks what people will pay for. Danny says Profile Picture prices dropped to near cost, about $5, after Lensa collapsed the market. He says new models appear so fast he barely sleeps.
- •Users surprised Danny by requesting LinkedIn-style photos, leading to Headshots
- •Profile Picture started expensive and is now near cost price
- •Lensa effectively killed the market for this product
- •New model releases can make existing technology obsolete overnight
- 25:03 – 30:26
Unlocking New Creativity: Design Tools and AI as Pre-Production
Steph asks how AI can go beyond replicating the old. Sinead describes AI-assisted design for people outside luxury markets, digital application of clothes and IP questions, while Danny frames AI as pre-production alongside Photoshop. They discuss virtual twins in campaigns and social platforms integrating try-ons.
- •Sinead: design tools reach people who couldn't previously enter the market
- •Copyright and IP are unresolved, such as asking an AI for clothes inspired by Celine
- •Danny: AI could be pre-production, with photographers shooting training images and AI generating the settings
- •Sinead imagines dropping your virtual twin into campaign shoots
- •Danny expects social platforms to integrate clothing try-on within a year
- 30:26 – 33:43
Keeping Up and Finding a Moat
Steph asks how they keep up. A speaker describes reading academic white papers and making conversations accessible. Danny recounts Lensa swooping in a month after his launch, concluding that distribution and marketing matter, and that builders should avoid hype and focus on small problems and SEO.
- •Reading academic white papers helps anticipate trends
- •Everyone should have a right to shape their own future
- •Profile Picture AI launched around October, less than half a year earlier
- •Lensa took the market share; indie hackers must prioritize marketing and distribution
- •Danny now focuses on SEO, tools and backlinks to outlast hype
- 33:43 – 36:54
Stacking Models and Getting Started with Building
Danny says Headshot Pro stacks 15 AI models, which is the moat, whereas Profile Picture was just DreamBooth. Hosting and running GPUs has become easy thanks to new startups. He recommends Replicate and no-code tools like Zapier, saying 99% of apps can be built with API wrappers.
- •Headshot Pro stacks 15 custom models for quality, training, upscaling and generation
- •Danny had to learn Python and machine learning to build them
- •Deploying models and running GPUs is now much easier
- •Replicate hosts models like Stable Diffusion with an API
- •Most apps can be built with API wrappers and no-code tools
- 36:54 – 40:47
Product Ideas and What to Build Next
Danny says he would avoid image generation and chatbots because Adobe, Canva and ChatGPT already cover them, and suggests picking an annoying problem AI could speed up. Steph proposes a brand-trained image tool, and Danny proposes a personalized morning podcast built from Hacker News and Twitter. They agree on avoiding hype-driven builds.
- •Avoid image generation and chatbots given Adobe, Canva and ChatGPT
- •Pick a problem that annoys you and consider AI to speed it up or save money in a recession
- •Steph suggests a self-serve tool trained on a brand's colors and past images; Danny bets Canva does it within three months
- •Danny wants a customized AI podcast from Hacker News and favorite Twitter accounts
- •Novelty projects fade, while clear jobs to be done last
