In this week’s conversation, I talk to Brennan Woodruff of GoCharlie about how AI services based on content generation can contend with a ChatGPT-dominated world. Plus John Biggs and I break down the week’s news: YouTube’s guidelines for synthetic content, a new study rating the big models on hallucinations, and an inflection point on the thorny issue of fair use.
Subscribe to the podcast here.
John Biggs and I offer a crash course on using AI for marketing and media. Learn more about the 3-hour session here.
When you’re running a startup, you’re already in a race. When you’re running an AI startup, you’re essentially in the New York City marathon. It’s already a slog, and you’re wall-to-wall with thousands of competitors of all stripes. Whether or not you succeed depends on the kind of race you’re running: Do you want to win the whole thing, beat your personal best, or be top in your category?
GoCharlie appears to be aiming for the third option. The AI startup is one of many that specializes in creating marketing copy, images, and other material, but it differentiates itself by applying its own large language model (LLM) trained specifically for that use case. That would seem to give the young company an advantage, but now that OpenAI had made it easy for anyone to create task-specific GPTs with assistants — and is creating a platform to sell them — can GoCharlie get past this “extinction-level event” for AI startups?
I spoke with co-founder Brennan Woodruff about GoCharlie, what it brings to the table to marketers and media people, and how AI entrepreneurs can stay in the race even when running alongside a ChatGPT that’s wearing rocket boots.
In this week’s AI news that’s most relevant to media…
What even is fair use anyway? Ed Newton-Rex, the VP of Audio at Stability AI — the creator of the Stable Diffusion image generator — very publicly resigned from the company, arguing strongly against the perspective, common among tech companies, that training AI models on copyrighted material constitutes fair use. It may be one of the most important public stances on the issue, and it hopefully helps force Big Media and Big Tech to make some kind of uberdeal that sets the standard for how to compensate copyright holders. My buddy Ricky Sutton has more to say about what such a deal would look like.
Let’s put “Hail Hydra” at the end of every deepfake: YouTube kinda-sorta took a stand on deepfakes, introducing new requirements for creators to label “synthetic” content made to look realistic, but allowing a parody/satire exception. It’s a important step, though still leaves a lot up to YouTube’s human moderators. Also: anyone making bank off of songs made from cloned voices of various artists is on notice now that those artists can force synthetic songs to be taken down. Progress? Probably. But other platforms (a certain single-letter network comes to mind) will likely have different standards.
Wait, people use Notion? This week Notion launched Q&A, an AI-powered feature that can scan all the material you’ve put on the service to inform answers to specific queries — essentially letting you have a conversation with your work. This is the dream of Google Bard’s feature that connects with all your Gmail and Google Docs, but Notion’s thingie probably has a better chance of giving useful answers since it probably won’t have every grocery list you’ve made since 2006 in there.
The Hallucination Olympics: Rankings for which generative AI model hallucinates the most are out, and boy, Google’s Gemini upgrade can’t come fast enough — Google Palm, which powers Bard, was dead last. Perhaps not surprisingly, OpenAI’s models lead the pack, though some smart folks were able to get Llama 2 into the same category. Hallucinations will never go away entirely, but we’re optimistic that the robots will continue to get better at, you know, facts. Now if we could just to the same with bias…
You can subscribe to The Media Copilot podcast on your favorite platform here, or via Apple Podcasts below.
Leave a Reply