Ventures from The Valley

There are scattered genuine insights—the ChatGPT-as-marketing-moment observation, the browser-as-rendering-mechanism analogy for chatbots, and the referral program failure—but they are buried in long stretches of meandering filler, throat-clearing, and repeated 'um's. The Wabby section is particularly thin, offering little beyond hand-wavy vision talk.

“the simple referral program, like something that was sort of textbook back then, actually didn't work at all”

“browsers are rendering mechanisms for websites and so those chat apps are just rendering mechanisms for for AI models”

Kuyda offers a couple of genuinely counterintuitive frames—chatbots as mere rendering mechanisms incapable of real differentiation, and the argument that the real zero-shot breakthrough was Meena/GPT-3 while ChatGPT was largely a marketing moment—but much of the episode drifts into standard AI-hype territory on jobs, regulation, and big-tech-vs-newcomers narratives.

“it was more of a kind of just a marketing moment for that uh for that instruct model and then it just all blew up”

“browsers are rendering mechanisms for websites and so those chat apps are just rendering mechanisms for for AI models. Um they're not it's impossible to make it different differentiated”

Kuyda is a genuine practitioner with exceptional firsthand credibility: she built production LLM systems before transformers existed, was OpenAI's first API customer, and scaled Replika to tens of millions of users on just $11M raised. However, she is no longer operationally running Replika, and the Wabby venture is extremely early, limiting the depth of current practitioner insight.

“we were the first customer of OpenAI uh which was pretty incredible. We still have like Greg Brockman and I don't know like Merati and Sam in our Slack channel talking about the model that they they're fine-tuning for replica”

“we only raised 11 million including and we actually pivoted. So we spent maybe half of it before we pivoted into something that re actually is called replica”

The episode contains solid specific data points—$11M raised, 1.5M waitlist, four named GPT-3 models, own 5–10B parameter models, GPT-3 API launch 2020 vs. ChatGPT 2022—but the Wabby section is almost entirely abstract, and many claims (e.g., 'dozens of millions' of Replika users, vague missed window of 2021) are hedged or imprecise.

“They provided four models through the API. Um they were called um ADA whatever Babage Cury and Da Vinci”

“we had our own like five to 10 billion brand models uh that we were developing ourselves but we should have gone big big big”

The host is a declared investor and self-described 'close friend' of the guest, which structurally eliminates critical pushback; most questions are compound, leading, or answered by the host before the guest responds. The one genuinely probing question—why Replika isn't at hundreds of millions of users—is quickly let go without follow-up.

“I'm also, um, thanks to you, uh, a proud investor in Wabby”

“why is it not um hundreds of millions? Is it just because uh people do not meet um a personal uh avatar uh like a friend a digital friend or there is any other problem there”