Would you take a live survey conducted by AI avatar? @synthesiaIO CEO Victor Riparbelli (@vriparbelli) thinks so -- for work, at least. At the opening of his $4B video startup's expanded NYC office, we talked about it -- then I made one to try myself 👀 upstartsmedia.com/p/synthesi…

At @synthesiaIO, our product development has always been focused on the enterprise. Yesterday, I spoke with @akashpasricha on @theinformation about our first use case for conversational Avatars: Roleplays for sales training. Now managers can help an entire team practice tough conversations and objections, so when the real calls come, they land deals.

Today on TITV: -How Musk turned SpaceX’s AI unit into an AI cloud firm | @graceihle -Can you trust AI agents with your credit card? | @Yueqi_Yang -Synthesia unveils AI avatar interviewers and coaches | @vriparbelli, CEO of Synthesia -Circle launches ARC, a new blockchain for agents | @jerallaire, Co-founder, Chairman and CEO of Circle 📺 Tune in at 10 am PT / 1 pm ET on thein.fo/4yqdfhx

In the last 2 years, web traffic to news sites has fallen by about 28%.

While last week I pushed back against the nonsense that AI is a threat to humanity, I wanted to start my Monday this week addressing another conspiracy theory making the rounds on tech and business podcasts: the view that the economics of AI don’t work, and that many of the products built around it are little more than wrappers around somebody else’s model. In reality, some of the people speaking about this topic have never built software so their description misses how AI products and services are actually being created and sold. They have an odd tendency (driven to the point of obsession in the case of Ed Zitron) to take the economics of the most expensive part of the stack (i.e. training frontier models) and use them as evidence that the economics of the entire industry are broken. But practically speaking, there are very few cases today where an end user interacts directly with a model. What most businesses and consumers use instead is a harness at the application layer: a model combined with layers of software that turn capabilities into useful applications. At @Synthesiaio, for instance, nobody buys our product just because we happen to have world-leading video and voice models; they buy our technology because we've built a sophisticated platform combining models built by us and by third parties, with layers of orchestration, security, governance, and observability. These layers are difficult to replicate; they are also a large part of what customers are paying for. Therefore, the economics look very different once you separate that application layer from the capital expenditure required to train frontier models. Across the application industry, margins of around 40% are common, with some of the better companies reaching 60%. So the claim that AI products are inherently unprofitable simply doesn’t fit those economics. For example, one of our newer products combines LLMs and agents from companies such as OpenAI or Google with our own video and voice technology to create Interactive Avatars that can talk, listen and take action. The description of this product might sounds simple in a LinkedIn post, but making those components work together reliably and at scale, and then delivering them to 90% of the Fortune 100 requires substantial amount of software engineering. But even if we have the debate just at the model layer, the Ed Zitron argument is full of holes. Yes, it is true that training frontier models can be capital intensive but selling inference to the application layer can be a very attractive business. And the amount of inference is likely to increase quickly as software moves from relatively simple prompts towards agents. A chatbot might make one model call to answer a question; an agent working through a task can make many calls as it reasons, uses tools, checks its work and takes actions. That means much greater token consumption, and therefore more revenue flowing back to the model providers.