Top Story

• Generative AI doesn't have a coherent understanding of the world. (MIT)
  
  No fucking shit. Thanks to the big brains at MIT for bringing us this world-shattering news.

  The researchers found that a popular type of generative AI model can provide turn-by-turn driving directions in New York City with near-perfect accuracy — without having formed an accurate internal map of the city.
  Despite the model's uncanny ability to navigate effectively, when the researchers closed some streets and added detours, its performance plummeted.
  When they dug deeper, the researchers found that the New York maps the model implicitly generated had many nonexistent streets curving between the grid and connecting far away intersections.

  It's a stochastic parrot. We know.

  This could have serious implications for generative AI models deployed in the real world, since a model that seems to be performing well in one context might break down if the task or environment slightly changes.

  Again, anyone who has used AI for more than a couple of minutes is fully aware of this.

  "We needed test beds where we know what the world model is. Now, we can rigorously think about what it means to recover that world model," Vafa explains.

  It doesn't have one.

  The researchers demonstrated the implications of this by adding detours to the map of New York City, which caused all the navigation models to fail.

  Yep.
  
  Years ago, engineers tried using genetic algorithms to optimise a particular electronic circuit to use fewer transistors. They got a result that worked, but nobody could explain how.
  
  Turned out it worked by the coincidental passive properties of the circuit, and not due to the transistors. The moment you made the slightest change to the operating conditions, it failed entirely.

Posted by: Pixy Misa at 06:52 PM | Comments (4) | Add Comment | Trackbacks (Suck)
Post contains 310 words, total size 2 kb.