Namasivayam: Transformer models miss edge-case opinions by defaulting to population averages
Vasanth Namasivayam · How NVIDIA, Meta & Dropbox Taught Me To Build Great Products | Vasanth, Founder - Featurely · Jun 27, 2025 · at 27:37
Featurely founder Vasanth Namasivayam explains why standard LLM transformer architectures cannot accurately simulate realistic human user cohorts for market research.
0:00 / 0:47exact quote · 47.9s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“Transformer models, OpenAI, is great at providing feedback on the average, but it doesn't cover the edge cases. Or the edge cases of your opinions. So that's the problem. So ChatGPT is good at persona feedback. It kind of averages it across the population. But if you really want to understand how does the population react? How does the population interact? Not only in the Gaussian middle, but also on the edges. Then, you know, ChatGPT is not built for, you know, that kind of application at all. And so it would require kind of, not just using base transformer models, but more hybrid architectures to be able to pull that off.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Vasanth Namasivayam
Namasivayam: Tech Promotions Incentivize Pointless Feature Factories Over Customer Value
“It's a politically incorrect thing to say, but for a lot of product builders, their focus is as much about getting promoted at the end of the appraisal cycle as it is about building a product which really resonates with the audience. And when that is your ince…”
Namasivayam: Synthetic Humans Must Fail and Complain Unlike AI Agents
“The synthetic human, the best thing which synthetic humans can do, which the AI agents cannot do is they can fail. They can fail at doing a task, which is a reflection of how a real human would fail in using a product. An AI agent will always figure out, or th…”
Namasivayam: AI Democratizes Building, Making User Understanding the Primary Bottleneck
“My belief, my hypothesis is that product building or generating products is going to get more and more democratized thanks to all the advance in you know, generative AI recently. But what is not getting democratized is user understanding.”
Namasivayam: Software will increasingly become hyper-customized, on-demand, and ephemeral
“I suspect that we'll increasingly see a world of more and more hyper-customized, on-demand, ephemeral software, which is kind of spun up on demand, used for a certain application, for a certain person, and then it kind of disappears.”
Namasivayam: Usability, Not Tech Moats, Will Win the Synthetic Human Race
“I think a lot of the companies are going to get it wrong because they are going to be focused, hyper-focused on the technology modes. Technology modes are interesting, but I don't think they are long lasting. I think what is more interesting is, does it just w…”
Namasivayam: AMD and competitors are just catching up to Nvidia's stack
“So they imagined a world where you know, you kind of needed the complexity Of semiconductors. And the entire stack. Even back in 2016. Which is why NVIDIA is at the forefront of technology, even today. AMD and others are just catching up.”