What surprised PG about AI
Paul Graham discusses how AI development surprised him by starting with human-level but flawed performance rather than the expected progression from simple to complex. Early AI systems like ChatGPT emerged as capable but unreliable, similar to an undergrad bullshitting through a paper, then improved toward perfection—the opposite of his predicted trajectory.
Summary
Paul Graham reflects on how his understanding of AI development in the 1980s was fundamentally wrong. He expected AI to progress incrementally, starting with perfect fly-level intelligence and gradually working up through mice, cats, monkeys, and eventually humans—maintaining perfection at each level. However, the actual development path surprised him completely. Instead of starting simple and scaling up perfectly, the first viable AI systems (particularly early versions of ChatGPT) emerged as fully human-like but deeply flawed—plausible-sounding but often incorrect, similar to an undergraduate trying to bluff their way through a paper with fabricated answers. This represents an inverted quality trajectory: rather than beginning perfect and working toward human capability, AI started at human complexity but with poor reliability, then worked toward perfection. Graham notes the persistent nature of this flaw, describing current AI as retaining much of that 'bullshitting' character. He illustrates this with the 'jagged frontier' concept—AI can solve famous open math problems while struggling with simple factual queries like restaurant hours, suggesting an uneven competency landscape rather than a smooth progression.
Key Insights
- Graham expected AI to progress from simple perfect intelligence (like a fly) gradually scaling to human-level, but instead AI emerged as full human-level complexity that was deeply flawed
- Early versions of ChatGPT behaved like an undergraduate bullshitting through a paper—producing plausible-sounding but often false content rather than simple, reliable outputs
- The actual optimization path was the opposite of what was expected: starting with human-level performance and working toward perfection, rather than starting perfect and scaling up
- Current AI systems retain the 'bullshitting' flavor of their early versions despite improvements
- AI exhibits a 'jagged frontier' where it can solve famous open math problems but fails at simple factual queries like restaurant hours, revealing uneven capability rather than smooth progression
Topics
Transcript
[0:00] I'm curious what surprised you about AI. >> Well, I was studying AI back in the 1980s, a very different kind of AI that would never have worked. I mean, it was a joke. So, I've been thinking about this a lot and I had a lot I spent a lot of time thinking about what AI would eventually be like. And the way we thought it was all going to play out was you start out with like a perfect fly, right? You know, it doesn't do much. It's like a fly. like a fly's brain, right? It can only do what a fly does, [0:32] but it does it as well as a fly, right? And then…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator
A camera that can see through walls
Bill Abster from Applied Electromagnetics discusses his company's breakthrough technology that uses radio waves and high-frequency antennas to create 3D images of objects behind opaque materials. The system processes echoes in real-time using GPUs to generate detailed depth maps, and the company has achieved significant progress in just 4 months by maintaining rapid iteration cycles.
What If We Stopped Using GPUs? | YC Paper Club
A YC Paper Club discussion exploring alternatives to GPU-based computing, featuring presentations on optical computing, neuromorphic chips, and biological computing approaches. Speakers argue that GPU efficiency (FLOPS per joule) has plateaued, necessitating fundamentally different computational substrates inspired by the brain's 20-watt energy efficiency.
Robotic factories that build robots
Tensor is building a 12,000 sq ft robotic factory that manufactures robots, including humanoids and specialized grippers. The company, founded by Berkeley researchers with autonomous vehicle and manufacturing backgrounds, treats the entire factory as a learning system that improves manufacturing processes with each customer order.
New LLMs Are Unlocking Robot-Use Agents
This discussion explores how large language models like Claude (Astra) are enabling robot control through code generation and tool use, building on foundational work like RT2 and visual language-action models. The speakers from Waddle Labs and RoboCurve explain how models trained on diverse data modalities—including web text, computer usage, and coding—are converging toward general-purpose robot agents that can perform complex physical tasks within the next 2 years.
Rocket cargo delivery anywhere on Earth in minutes
Matthew, CEO of Hover Arrow, discusses the company's mission to deliver cargo anywhere on Earth within minutes using rocket technology. The startup has secured $1.37 billion in commercial letters of intent, with primary focus on military applications and emerging interest from industries requiring rapid, temperature-controlled delivery of perishable goods.