AI Insider: Things Are About to Get Much Worse

Tijdlijn van de aflevering

1. 0:00

   AI's thinking speed and one-shot pattern recognition versus human cognition

   • AI can process roughly 100 tokens per second (about 0.7 words each), making it around a thousand times faster than human thinking, which the guest frames as Kahneman-style Type 1 pattern recognition rather than slow deliberate reasoning.
   • Diffusion and transformer models do everything in one shot with no chain of reasoning, so they can analyze an entire legal corpus instantly, like a savant giving an immediate answer on steroids.
   • The guest argues the bottleneck was never input (reading leases, shuffling paper) but thinking itself, and AI now outperforms humans at that core cognitive task.
2. 6:16

   Job displacement across white-collar professions and the collapsing career ladder

   • The guest claims all jobs are going, and that people would react very differently if blue-collar jobs were disappearing versus lawyers, doctors, and mathematicians, who are now the ones at risk.
   • The traditional ladder from paralegal to first-year associate to partner is being pulled up from the ground floor, since second-year associates can now direct AI agents to do first-year work.
   • Humanoid robots costing about $150 an hour, with super-genius IQs that can read and follow a recipe, represent the next wave of physical labor replacement.
3. 11:02

   The intelligence inversion and why this technological wave differs from prior ones

   • Previous transitions moved humans up the value curve from land to labor to capital to cognition, but AI now substitutes for the graduate-level cognitive work itself.
   • The average person speaks 10 million words a year; GPT-3 cost $600 per million words while the latest model costs $1 per million, meaning all your words can be replaced for about $100.
   • The guest argues the difference is speed, quality, and the impossibility of retraining truck drivers or accountants fast enough to beat AI at anything except physical contact and empathy.

Kernbegrippen

• AI replacing cognitive work— The core argument that AI now substitutes not just muscle but thinking itself, eliminating the historical human advantage of moving up to cognitive labor.
• one-shot analysis— Diffusion and transformer models can analyze an entire legal corpus or video in a single pass without any chain of reasoning, like a savant's instant answer.
• 100 tokens per second— AI thinks roughly a thousand times faster than a human, at about 0.7 words per token, making human deliberation look glacial.

Opvallende citaten

• You can think at 100 tokens per second. A token is about 0.7 words. Like I said, if you use the chat Jimmy thing, it's 15,000 and it'll go to 100,000. So it's about a thousand times faster than a human.

  🤯

  — The concrete number reveals AI thinks roughly a thousand times faster than a human, a scale that is hard to intuitively grasp.

  Ga naar 0:42

• When GPT3 came out, it was $600 per million words. So that's like $6,000 or something like that multiplied by the mistakes you make. The latest uh GPT Luna model is $1 per million words. So it's $10 for your total words or 10 times that's 100 bucks to replace all your words.

  🤯

  — It costs about $100 to replace an entire person's yearly speech output, showing how absurdly cheap cognitive replacement has become.

  Ga naar 14:33

Praktische conclusies