Skip to content

Inside The MIT AI Study Everyone Misunderstood (And What It Means For Startups)

Lightcone Podcast

Hosted by Diana Hu, Garry Tan, Harj Taggar and Jared Friedman

21m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

The YC partners break down the MIT study on AI project failures that went viral with misleading tweets claiming 95% of AI projects fail. Jared Friedman explains the study actually confirms what they've been saying about AI agents in the real world. The hosts dissect why enterprise AI adoption fails: internal IT systems are poor, consultants like E&Y and Deloitte can't build working software, and the result is a "startup-shaped hole" in enterprise processes. Specific case studies include Tactile building a real-time decision engine for banks in a fraction of the time and budget Citibank and JP Morgan spent on internal builds, Greenlight winning a bank deal after E&Y spent a year building a non-working system, and Redacto closing a FAANG deal 154 days after their YC batch.

The panel also discusses the "startup dreamer" archetype of enterprise champion, the Karpathy interview misinterpretation, and how switching costs after training an AI system create a genuine moat. Garry Tan closes by reframing the 5% success rate not as a reason for despair but as evidence that the top 1% of builders—polymaths who understand both technology and human processes—have an unprecedented opportunity.

  • The MIT study found that two-thirds of surveyed AI projects were internal builds or consulting engagements, and the success rate was significantly higher when enterprises bought from outside vendors like Greenlight or Tactile. 7:15
  • Enterprises would prefer to buy AI solutions from established companies but fundamentally can't build the products themselves, giving startups an unprecedented shot at enterprise deals. 15:55
  • Finding a "startup dreamer" champion inside a large company—someone who wanted to do a startup but was too risk-averse—can unlock enterprise deals because they want to live vicariously through your success. 13:42
  • A CIO of a $5 billion financial services firm confirmed that once time is invested in training an AI system, switching costs become prohibitive, creating a genuine moat for AI startups. 19:52
  • The Karpathy interview about agents requiring significant tooling and context was misread as "AI is overhyped" when it was actually a call to build better infrastructure around AI agents. 18:29

Highlights

  • 3:36

    Apple Calendar App as Software Quality Benchmarkinsight

    Garry Tan argues that if Apple, with infinite capital and the smartest people, cannot make a good calendar app, then no internal IT team or consulting firm can build good software either.

  • 5:40

    Tactile vs. Bank Internal Buildsstory

    Harj Taggar describes how Citibank and JP Morgan each spent three to five years and tens of millions building decision engines internally, while Tactile built a REST API that does the same in real time for a fraction of the budget.

  • 6:37

    Greenlight Wins After E&Y Failsstory

    Jared Friedman recounts how a bank chose E&Y to build an AI system, E&Y spent a year and it didn't work at all, and the bank then came back to Greenlight who deployed a working system.

  • 7:15

    Internal vs. External AI Project Success Ratesdata

    Jared Friedman reports that two-thirds of the MIT study's surveyed projects were internal builds or consulting engagements, and the success rate was much higher for the one-third that bought from outside vendors.

  • 11:23

    Redacto Closes FAANG Deal in 154 Daysstory

    Diana Hu describes how Redacto, a document processing company, closed a FAANG deal just 154 days after their YC batch by outperforming the company's years of failed internal OCR attempts.

  • 13:42

    The Startup Dreamer Champion Archetypeinsight

    Harj Taggar identifies a specific type of enterprise employee who always wanted to do a startup but was too risk-averse, and who will champion a startup they get along with because it lets them live vicariously.

  • 18:29

    Karpathy Interview Misread as AI Doomerismdebate

    Harj Taggar explains that the viral tweets about Karpathy saying agents are overhyped actually misrepresent his point: agents need proper data, context, and tooling, which is an opportunity for builders rather than a death sentence for AI.

  • 19:52

    Switching Costs as a Genuine Moatinsight

    Jared Friedman reads a quote from a CIO of a $5 billion financial services firm saying that once time is invested in training a system, switching costs become prohibitive—directly contradicting the "no moat" argument for AI startups.

About this episode

MIT's new State of AI in Business report went viral for claiming that 95% of enterprise AI projects fail. But the real story isn't that AI doesn't work — it's just big companies can't build it. In this episode of the Lightcone, Garry, Harj, Diana, and Jared break down what the study really says, why in-house enterprise AI efforts keep stalling, and how startups are filling the gap with products that learn, integrate, and actually deliver value.

Transcript

0:00
21:43

4,206 words · 4 speakers

Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.

Topics

  • Enterprise AI Adoption
  • AI Startups
  • MIT AI Study
  • Enterprise Sales
  • AI Agents
  • Go-To-Market Strategy
  • Switching Costs
  • Product-Market Fit

Page updated · transcript and summary are machine-generated