The Future of AI: Personal Agents, Taste & Private Data | Lin Qiao & Demi Guo | E9
TL;DR
Private data is the next frontier, not more public web scraping — Lin Qiao argues less than 5% of the world’s data is in public internet + labeling pipelines, while 95% sits inside enterprises and apps, so the real unlock is tuning models on private data rather than waiting for bigger base models.
AI is flattening org charts before it fully replaces workers — Fireworks runs tens of trillions of tokens per day with just 150 people, and Qiao says companies are already cutting layers of middle management because AI makes information discovery, summaries, and alignment dramatically easier.
Taste is becoming the scarce human skill — Demi Guo’s core thesis is that future human value comes from judgment, taste, and identity, which is why she shifted Pika from creation tools to “humanized agents” that users train like a creative assistant, not a one-shot software tool.
Open models are converging fast, but frontier labs still lead on the hardest work — Qiao says simple day-to-day tasks like spreadsheet work, routing, writing, and document processing are already near parity, while complex systems-level agent tasks remain roughly 6–12 months ahead in proprietary models.
The anti-AI backlash is partly about authenticity, not just jobs — One of the most memorable moments is Qiao describing her high-school daughter’s Gen Z view: they’re cynical about AI because it doesn’t feel authentic, and using AI-generated art for a school magazine is seen as worse than drawing it yourself.
The creators think agents should feel more like collaborators—or even children—than tools — Guo repeatedly says the wrong mental model is “productivity software”; the right one is an agent you raise through feedback, emotional attachment, and personalization, which is why some users are already spending up to $10,000 per month on them.
The Breakdown
Fireworks AI’s origin story: from Meta infrastructure to inference at scale
Lin Qiao opens with the backstory: seven Meta veterans started Fireworks after helping build AI infrastructure and PyTorch during Facebook’s mobile-first era, back when “my first app on iPhone was a flashlight.” Her pitch is that many companies now face the same AI-first transition Meta once did—no AI hardware, no AI software, no AI team—and Fireworks exists to bridge that.
Why private data matters more than who you trust
Asked whether enterprises want open models mainly for cost or privacy, Qiao goes deeper: the real issue is that foundation models only reflect a tiny slice of the world’s intelligence. She says public internet and labeled data account for less than 5% of global data, while more than 95% lives in private enterprise systems, and activating that data through tuning is the next phase of frontier intelligence.
Pika’s pivot: from creative tool to “humanized agent”
Demi Guo explains that Pika started as a video creation tool, but the team found web interfaces and prompting still felt too hard for normal people—even some teammates couldn’t comfortably use them. That pushed them toward a more accessible interface: a creative agent you can talk to, video call, and screen-share with, more like working with a person than operating software.
The AI layoff trap—and why both founders are more optimistic than alarmist
The show pivots to a paper arguing firms may automate too aggressively and destroy consumer demand in the process, but Qiao’s on-the-ground view is more nuanced. She says AI has compressed the path from idea to production from multiple quarters to multiple days, creating a boom in experimentation and startup formation even as large companies flatten orgs and reduce middle management.
Middle management gets squeezed as AI becomes the company memory
This is one of the sharpest sections: Qiao describes a future where managers oversee 20 to 50 people instead of 7 to 10 because AI now handles context gathering, summaries, and information flow. At Fireworks, she says they do fewer one-on-ones, rely heavily on Slack, and use AI to summarize context by individual need—making the org flatter and faster.
Taste, authenticity, and the “human in the loop” fight
Guo argues the durable human advantage is judgment and taste, while Qiao adds a surprisingly personal observation from her daughter: Gen Z is skeptical of AI because it feels inauthentic and uncreative. That leads into a shared view that agents still need human steering; otherwise, as one quoted clip puts it, you get “the ultimate slop.”
Open vs. closed models: convergence is real, but not equal
Qiao gives one of the cleanest heuristics in the episode: when models keep swapping places on leaderboards every week or two, that’s a sign they’re converging. Her practical read is that open models are already very close on routine tasks, but for difficult agentic work—like designing complex distributed systems—frontier proprietary models still lead by roughly six months to a year.
Why the public fears AI while insiders see possibility
The final stretch contrasts public negativity with insider optimism, from U.S. polling to examples in China where AI and robotics show up in daily life—like robots delivering still-frozen ice cream in minutes. Qiao says medicine is a great example of AI’s upside, from ambient medical scribes to preventive-care software that can navigate fragmented record systems, while Guo says the industry needs to stop framing AI mainly as replacement and start framing it as self-expression and empowerment.