This Week in Startups - E2215

Episode Title: SO MANY THINGS need to go right just so you can watch a TikTok!
Date: November 26, 2025
Host: Jason Calacanis
Guests: Gu Rao (CEO & Co-founder, Newbird), Mike Velardo (Subject AI), Alexander Wang (archival interview, Scale AI), with co-hosts Alex and Lon Harris

Overview of the Episode

This episode dives into the invisible complexity behind everyday technology experiences—like streaming a TikTok—and how AI is being applied in infrastructure, IT operations, and education. Host Jason Calacanis and team talk to Newbird’s Gu Rao about Hawkeye, an AI-powered agentic SRE (“Site Reliability Engineering”) tool, and to Subject AI’s Mike Velardo about scaling personalized education with AI. The episode wraps with an archival interview with Scale AI’s Alexander Wang, reflecting on AI’s past, present, and future, especially in the context of data labeling and autonomous vehicles.

Main Themes:

• The huge and intricate engineering required in digital infrastructure.
• AI’s role in automating and optimizing IT operations.
• Practical applications of generative and agentic AI in education.
• Reflections on the evolution and realistic future of general AI.

Key Segments, Insights & Quotes

1. How Many Things Must Go Right for a Simple TikTok To Play?

Guest: Gu Rao, CEO & Co-Founder, Newbird
Timestamps: 00:00 – 07:39

• Complex Layers of Delivery

  • "The moment you get on a phone and you're looking at a TikTok video, I mean the number of layers that are involved to deliver that... it's a very complex environment and people want more of this faster." (Gu Rao, 00:00)
  • Storage, data quality control, content distribution networks (CDN), edge caching—all must work seamlessly.

• Software & Infrastructure Example

  • "There's software associated with [storage]... quality control... no data loss, no integrity compromise. Then the distribution aspect... you don't want somebody looking at the video to hit the backend server each time—content distribution networks involved." (Gu Rao, 00:39)

• Troubleshooting Complexity

  • Identifying the point of failure (CDN, storage, backend, network) is a daunting, multi-layer problem.

Notable Moment:
Alex voices awe at the fact that Internet infrastructure stays online, despite such vast complexity:

"Is it more surprising that the Internet as we know it works today or is it more surprising when it breaks?... I'm kind of shocked that everything stays online most of the time." (Alex, 07:39)

2. AI for IT Ops: Inside Newbird’s Hawkeye

Guest: Gu Rao
Timestamps: 08:42 – 19:59

• Agentic AI in IT Operations

  • "An agentic system leverages generative AI... applies it to a very specific task domain... we leverage large language models, extract the reasoning capabilities and interface them with the complex array of metrics, logs, alerts, traces." (Gu Rao, 08:46 & 09:48)
  • Hawkeye acts as an “AI IT Ops engineer,” handling both SRE (Site Reliability Engineering) and traditional IT operations.

• Engineer vs. Agent vs. Copilot

  • Hawkeye autonomously finds, diagnoses, and potentially remediates IT issues. It can run both reactively and proactively.
  • "A co-pilot is like you have to distinctly carve out a body of work... what we're really trying to do is streamline your... IT operations where they're getting far fewer alerts or like 80% of the alerts are being handled by Hawkeye and the remediation is done. It's automatically resolved." (Gu Rao, 14:01)

• Context Engineering to Control AI Cost and Quality

  • "Anybody can go in and download a whole bunch of logs, paste it into ChatGPT and say what's wrong. ... That's called context engineering. That's all we do. We specialize in context engineering because... we don't want the customer to incur very large inference costs. More importantly, we don't want LLMs to come up with garbage answers." (Gu Rao, 15:57)

Notable Quote:

"More context—while these large language models have very large context windows, garbage in, garbage out." (Gu Rao, 16:10)

3. Hawkeye in Practice: Plug-and-Play, Reasoning Across Domains

Timestamps: 17:44 – 22:50

• Plug-and-Play AI for IT

  • Hawkeye doesn’t need extensive retraining for each environment; foundational models have seen common IT scenarios.
  • The system’s strength lies in “context engineering”—feeding the right slices of log/metric/alert data to the LLMs.

• Results After One Year

  • "We're achieving close to 90% plus reduction in the meantime to incident response and resolution from the RCAs that we're creating." (Gu Rao, 19:51)

• Autonomous or Consultative Actions

  • Hawkeye can both suggest and, if allowed, autonomously implement remediations (via code, feature flags, or PRs).

• Agentic Systems as Colleagues

  • "You should treat [agentic systems] like it has a Persona because these things are built out of human knowledge. ... I've had evaluations... where an SRE manager came to me and said, your product is doing really good. Hawkeye came up with an answer, but my SRE disagreed... both were correct." (Gu Rao, 22:50)

4. Pace & Plateau of AI Model Improvements

Timestamps: 24:16 – 26:29

• Slower Model Progress is Expected

  • "At some point everything starts to plateau... when you have billions of parameters... at some point adding more may not even make sense... you may just start getting weird answers. ...The knowledge in the models is reaching some sort of convergence." (Gu Rao, 24:54)

• Next Steps: Building Context & Tools Around AI

  • The “core of the brain” is ready; next phase is context, “external EQ” for models.

5. Hawkeye’s Proactive (Not Just Reactive) Capabilities

Timestamps: 26:29 – 28:00

• Retrospective analysis (root cause after an outage), real-time triage, and trend forecasting for preemptive action.
• “A customer will say, I’m planning to roll out new software… is my environment ready to absorb these changes?”
  Hawkeye responds, sometimes bluntly:

  "Hawkeye will tell you, hell no, it's not. You got a lot of work to do." (Alex & Gu Rao, 27:45)

6. Subject AI: AI for Education

Guest: Mike Velardo
Timestamps: 28:38 – 47:37

• From ‘Netflix of Education’ to AI-Driven Learning

  • Originally created cinematic-quality, short-form educational videos; pivoted to AI-powered, personalized classroom tooling.

• Flip Classroom & Teacher Empowerment

  • Subject AI enables “flipped classroom”—students cover basics with videos/quizzes; teachers provide high-value one-on-one.
  • "I'm really here to amplify the heroes of society, which are teachers and coaches... [Subject AI] allows for them to have more one on one interactions." (Mike Velardo, 33:45)

• AI Video Games and Engagement

  • Gamified, interactive, short-form curriculum: “All of our clips around five minutes or less now... 96% of our students complete at least one course." (Mike Velardo, 37:39)

• Teacher Tools: AI-Driven Grading and Planning

  • The platform auto-grades, plans lessons, and tracks student progress, aiming to triple student coverage while keeping teacher admin below an hour/day.

• AI Model Choice & Academic Integrity

  • Uses Claude, has used GPT and Gemini; models can be swapped modularly.
  • AI grading is “recommended”—teachers have ultimate say; for “teacher of record” cases, there’s an accredited human backstop.
  • "We want to make sure that students are actually learning and prepare for the next step... We don't want to be a credit mill." (Mike Velardo, 44:11)

7. Flashback: Scale AI’s Alexander Wang (2019)

Timestamps: 47:43 – end

• Alexander’s Early Journey

  • Raised $100M by 22; parents were physicists at Los Alamos. “I have a fun little history. I grew up in Los Alamos, New Mexico... both my parents are physicists.” (Alexander Wang, 50:32)

• What Scale AI Actually Does

  • Core business: providing high-quality labeled data for ML/AI projects, especially self-driving cars.
  • "Machines don't know what to do unless they have data that actually tells them what they're supposed to be doing... we are this data refinery." (Alexander Wang, 54:35)

• The Future of Self-Driving and Regulation

  • Alexander predicted regulation wouldn’t be a major blocker; most obstacles are technical and data-driven.
  • "Humans are terrible drivers and if we're going to allow 16 year olds to drive tanks down the highways at 100 miles an hour, maybe letting the smart computers do it's better idea." (Alex, 60:51)

• AI’s Impact on Labor & Progress

  • Autonomous trucks will take the “boring parts” of the job; AI augments, not eliminates, higher-value work.
  • "The automated truck driving systems, actually what they would do is automate the, the long haul middles ... and allow the current truck drivers to focus on these like higher value trips." (Alexander Wang, 62:17)

• AI Danger, Oversight, and Path to AGI

  • Alexander is skeptical of sci-fi fears; argues for practical oversight and responsible engineering, not dystopian anxieties.
  • "You only think that's not far fetched because you watched a lot of these sci fi movies... In reality, there's a lot of oversight over these machine learning systems." (Alexander Wang, 65:44)

  On general AI:

  • "It's very overblown, the timelines that people are talking about general AI happening... not even clear that if you have infinite compute, you'll be able to produce general AI. I think that's very unclear." (Alexander Wang, 74:19)

Memorable Quotes & Moments

• “I’m surprised any part of [this] infrastructure works... like when my train comes in on time, I’m really happy with it.”
  — Gu Rao (07:49)

• “We specialize in context engineering because... we don't want LLMs to come up with garbage answers.”
  — Gu Rao (15:57)

• “Treat agentic systems like a colleague. Sometimes they will surprise you, even out-innovate your human SREs.”
  — Gu Rao (22:50)

• “We want teachers to spend LESS time on the product and more with students. Let us do the grading.”
  — Mike Velardo, Subject AI (39:59)

• “Machines don't know what to do unless they have data that actually tells them what they're supposed to be doing... we are this data refinery.”
  — Alexander Wang, Scale AI (54:35)

Segment Timestamps & Topics

| Timestamp | Segment | |----------------|-------------------------------------------------------------| | 00:00-07:39 | The unseen complexity behind serving a TikTok video | | 07:39-19:59 | How AI can act as an IT Ops engineer (Newbird Hawkeye) | | 19:59-26:29 | Hawkeye's performance, context engineering, and evolution | | 26:29-28:00 | Hawkeye’s proactive capabilities in IT environments | | 28:38-47:37 | Subject AI’s evolution, tools, and impact in education | | 47:43-end | Archival interview with Alexander Wang (Scale AI) |

Episode Takeaways

• The Backbone of Modern Life is Vastly Complex and Fragile:
  A TikTok video relies on dozens of technical layers—any could fail.

• Agentic and Applied AI is Now Essential, Not Experimental:
  Whether diagnosing IT ops or personalizing learning, AI is moving from theoretical extra to operational core.

• Context is King in AI Ops:
  Feeding the right context to language models is the keystone to reliable, affordable automation.

• Education is Evolving Rapidly with AI:
  Making lesson planning, grading, and differentiation scalable is possible, but still needs a human touch and oversight.

• Long-Term AI Hype Should Be Tempered:
  Major advances plateau, more compute ≠ AGI. Oversight matters—and there’s more continuity than disruption in the AI story.

For Listeners Who Haven’t Tuned In

This episode is rich with insights for anyone curious about the "machine behind the curtain" in tech, how AI is being practically deployed in mission-critical business and education settings, what makes for robust and responsible AI adoption, and how even the AI titans themselves urge a blend of optimism, caution, and practicality.

Listen to this episode for:

• Real-life examples of AI augmenting teams (not just replacing them)
• Candid founder perspectives on deploying next-gen tools in enterprise and B2B SaaS
• A look back at how the AI conversation has changed in just five years