Suggestions
Daniel Chen
Crypto at Sequoia Capital
Daniel Chen is a multi-time founder and former Visiting Fellow at Sequoia Capital. He is currently the CEO and co-founder of Quilt, a Sequoia-backed company building AI knowledge assistants for GTM teams.2 Chen's career has been marked by significant roles in both the venture capital and startup worlds:
-
He was a Partner at Sequoia Capital from May 2022 to June 2023, following his role as a Visiting Fellow from December 2021 to June 2022.2
-
Prior to Quilt, Chen was the CEO and co-founder of Hero.app, which was acquired by People.ai in April 2021.2
-
Chen also served as a Partner at Andreessen Horowitz from June 2014 to September 2017.2
-
He founded OpenToken, Inc. in 2017 and served as its Chief Executive Officer until 2019.1
-
Chen holds a Bachelor's degree in Business, Economics, and Management, as well as a Bachelor's degree in Computer Science from Caltech.2
Throughout his career, Chen has demonstrated expertise in technical recruiting, startup founding, and venture capital investments, particularly in the areas of AI and blockchain technology.3 His experience spans across various sectors of the tech industry, showcasing his versatility and deep understanding of the startup ecosystem.
Highlights
There are many things that we're not the best at, but one thing we've done particularly well is selling to large enterprises as a seed-stage company (sampling of our customers on our website: https://quilt.app/).
The 2022 downturn, the rise of procurement teams running the show, and the sensitivity around data privacy for AI-native products was the perfect storm to make selling into enterprises uniquely challenging. Here’s how we do it:
1. SOC 2. We started the SOC 2 process right after incorporation. A Type 2 is a must for companies building in AI: a point-in-time Type 1 attestation doesn't cut it. You should plan out at least 4 months in advance of selling to customers (1 for prep, 3 for the observation period). Also, a pentest is a must-have. The great thing is, implementing best practices when you're small is a lot quicker than course-correcting later.
-
AI data privacy. Especially for global organizations with an EU footprint, data privacy is a huge concern. These organizations will often enforce a strict DPA that ensures no training of generative AI models, whether through you or a subprocessor that you use. If using model providers like OpenAI and Anthropic, some products such as their APIs do not train on user data whereas other products such as ChatGPT and Claude do. Make sure you limit your use to the former.
-
Eliminate subprocessors. We have a shortlist of standard subprocessors that we use which makes understanding risks in the flow of data very easy. On the other hand, the more small services you use (e.g. inference wrappers, embeddings persistence layers), the more likely it is for the evaluation to be held up because they don’t have the right DPA or the procurement team is unfamiliar. The fewer layers of developer abstraction the better.
-
SSO/SCIM. If you do a >20 seat deployment, IT will likely want user access control that fits neatly into their existing process. This includes not only enforcing SSO, but also potentially provisioning/deprovisioning accounts via SCIM. Keep in mind that SCIM providers typically charge per-connection.
-
Security questionnaires. Make sure that you are buttoned up as you go through security review. Most often, this will be done through a vendor portal: Coupa, Zip, OneTrust, Graphite, etc. The questions tend to be repetitive for each customer, so once you’ve answered one, all future ones should be dead simple. We have a tool to make this easy.
Curious Quilt can automate your first vendor onboarding? Comment below for access and a quick walkthrough 👇
Low-latency models are the future.
While it is unclear whether or not models are getting better (or hitting an asymptote), it is very clear that models are getting faster and cheaper. In the next 6-12 months, low-latency models will be ubiquitous.
Why does this matter? Low-latency models unlock a new universe of use cases, largely centered around real-time conversational AI: a virtual tutor, psychologist, friend, celebrity avatar, SDR, customer support agent, etc.
But there are still challenging problems to solve:
- Latency is still an issue. A startup recently showed us a really impressive demo with ~1s turn latency (“laggy Zoom”). However, the lag was still noticeable.
- There is still a bit of an uncanny valley in both speech and video, though companies like Tavus and HeyGen are doing amazing work improving this.
- Similar to the lag in enterprise chatbot adoption vs. consumers, we see a strong aversion in companies to deploying Zoom bots in any active opportunity or deal cycle, whereas tech-forward consumers are more accepting of interacting directly with AI.
- For enterprise use cases, it is important to ground all responses on proprietary knowledge and eliminate hallucinations entirely (vs. consumers may not care). This requires entirely different infrastructure than a transcription + Grok/Llama completion setup for conversational use cases.
Follow Quilt for more!
#LowLatencyModels #ConversationalAI #EnterpriseAI
