SuperMe

A voice-clone content tool made great content that never spread. Experts talked into a mic, and AI structured the rant into audio and video in their cloned voice. The content was good 10 out of 10 times, but it was not shareable or viral, and people struggled to know what to say.

Winters' diagnosis: "you're prompting the human instead of prompting the AI." A community-first approach also failed, and neither was the come-for-the-tool, stay-for-the-network product they wanted.

Subversive▶ 16:00

Instant crawled profiles worked, but were fulfillment, not a network. An agent built an AI profile from a person's public content in about 10 minutes. People shared them on LinkedIn and were surprised how representative the answers were.

Winters saw it satisfied users without creating network value: "that's not a network product that's a fulfillment product." The lesson is to test whether the core mechanic builds network value, not just satisfaction.

Subversive▶ 16:40

Search across many profiles turned individual advice into a network effect. A homepage question finds the right experts, pulls perspectives from several, summarizes them with attribution, and lets users open a full conversation with any one. Each added profile improves search.

It automates what founders already do: "you ask a bunch of people, you make up your own mind." Profiles answer from the owner's own views, say "I don't know" without a basis, and the owner sees every answer and can add context.

Subversive▶ 13:25

Each failed experiment felt scary, but many tries toward one goal is normal. Two early products fell short before search revealed the network, and each failure raised the question of whether anything would work.

Winters says running many experiments toward a known goal is common in consumer products. SuperMe launched officially in November after months of profiles being shared early.

Subversive

Sourcing only the top 1% of experts beats the internet's average answer. General chatbots return the average of what the internet thinks, so SuperMe sources only the best people in growth, product and engineering. It evaluated its answers against ChatGPT and Claude and found them better.

Showing each expert's background and the companies where they applied the advice adds trust a black-box LLM lacks. The evaluation is internal and no data was shown.

Subversive▶ 19:50

Sourcing edges fade, so sync knowledge that never reaches the public internet. Winters expects LLMs to learn who is credible, making the sourcing edge "not going to be long-term sustainable." SuperMe syncs meeting notes, docs, Slack and email, where most expert thinking lives.

Each item is screened for professional relevance and privacy; an upcoming funding round or a performance review is judged too private. Unique data volume is a key metric, and more experts draw more learners who invest more in their profiles.

Subversive▶ 20:40

Per-item visibility control makes people willing to sync sensitive sources. The system proposes a visibility level for every synced item (private, company only, network only or public) and the owner overrides it. Many users have been keen to sync because they keep that control.

Private items still inform the owner's own SuperMe but are not seen by other people's queries. This also lets knowledgeable but not very public people represent themselves.

Subversive▶ 22:12

Traditional loops still matter, so SuperMe adds tool calls on top. Profile and content sharing brought the first 3,000 users. Some users already call SuperMe as a tool from Claude and Google Docs.

Winters' view is that AI shifted technology a lot but distribution little, so content, virality, paid and sales still apply. "You shouldn't need to be on superme to get value from super profiles."

Subversive▶ 25:00

Use a user's work context to trigger agent outreach to the network. If an agent knows a user is working on debt financing, it can group-chat the best profiles on the topic and report back only if it learned something valuable.

This extrinsic trigger is meant to build the intrinsic habit of trusting the product, which no one gives on day one.

Subversive▶ 27:40

Calibrate agent recommendations to a 20% to 80% open rate. Under 20% of recommendations opened means the agent is spammy. Over 80% means it creates a lot of value and should do far more of it.

Human feedback on value tunes how much engagement the network initiates.

Subversive▶ 29:17

Stay small while searching for a consumer network's spark. SuperMe has eight full-time people plus contractors: one ops growth person, one designer, the rest engineers, and no sales or marketing team.

Enterprise startups can scale from 3 to 15 quickly because they follow a repeatable go-to-market playbook; consumer networks have none, so former consumer founders advised limiting coordination cost.

Subversive▶ 34:50

Onboard every hire like an engineer, including juniors. From day one new hires are told to be engineering first in solutioning, giving designers, marketers and ops people engineering leverage.

Winters argues for still hiring junior talent: with AI-native onboarding they "crush way beyond their years," carry no bias against AI coding tools, and strong CS grads are more available as big companies cut hiring. Shopify has said it still loves junior talent.

Subversive▶ 33:50

Hire for agency and drop people who wait for direction. SuperMe's ops hire laid out his approach and next steps in the interview. At GrubHub, the CTO let a highly credentialed QA hire go after a week for waiting to be told what to do.

Founders' direct reports get thrown in the deep end: "this is a startup. I'm not going to build a 90-day plan for you." People also need judgment on when to ask an expert, as one employee did on email sending and IP warming, and when old knowledge no longer applies and they should test.

Subversive▶ 37:30

Cheap code is no excuse to skip strategy or run too many experiments. Winters saw incumbents flood customers with 12 new features at once, muddying the signal on what works.

He and cofounder Ludo differed (minimum features and maximum design versus the reverse), so they mapped the 2x2 and chose one process. They now require a clear signal of success and debate whether each non-working experiment can compound, after too many unresolved bets kept them from doubling down on a winner.

Subversive▶ 32:30

Growth leaders should warn the CEO before the growth model plateaus. Winters compares AI now to the first mobile era, when short-term winners faded and Uber, Snapchat and DoorDash came in the next wave. Copyable magic tricks leave clones at similar revenue, heading toward zero.

"Are you going to have any revenue in 2030 is my question." SuperMe itself has not decided between subscription and transactional pricing, and nothing yet shows users will pay or keep returning.

Subversive▶ 1:00:50

Winters built his edge by owning growth end to end. He refused to pick between product and marketing, ran growth at GrubHub for five and a half years through the IPO, and at Pinterest rebuilt signup, onboarding and notifications around SEO after Facebook shut off the open graph.

As Eventbrite's CPO he laid off 45% in his first year and reached break even through the pandemic. He cofounded SuperMe with Ludo, his former head of growth engineering at Pinterest, aiming to build the AI era's professional network over 5 to 10 years.

Subversive