Ilya’s latest interview: Current approaches will “go a while and then go downhill” They will continue to

Ilya’s latest interview: Current approaches will “go a while and then go downhill” They will continue to improve, but they will not achieve general artificial intelligence The kind of systems that work are “the kind we don’t know how to build yet” Expected to have a significant impact on the labor market in the coming years Major bottleneck: the ability to generalize.

Models are far inferior to humans at this You can train a model to solve all competitive programming problems, but the model still won't develop real "taste", whereas a teenager has about 10 You can learn to drive in hours.

The evaluation scores look great, but the actual performance is very poor, because the reinforcement learning training will eventually be optimized for the evaluation itself.

Interview address: The following are my views on the big AI infrastructure in the United States.

They are basically based on the continuation of the current route.

It is feasible in five years, but it is really not necessarily true in five to ten years, and it is unlikely that the value created in five years will be greater than what they invested.

Visual summary

Article argument map

Generated from the post's content graph

Mermaid outline

flowchart LR
  format-long_post["longform note"]
  topic-evaluation["evaluation"]
  capability-evaluation["evaluation"]
  format-long_post -->|covers| topic-evaluation
  format-long_post -->|signals| capability-evaluation

Visual structure

Essay structure map

Built from summary and key paragraph positions

Mermaid outline

flowchart LR
  thesis["Ilya’s latest interview: Current approaches will “go a while and then go downhill” They will continue to im..."]
  signal["Ilya’s latest interview: Current approaches will “go a while and then go downhill” They will continue to im..."]
  operator["Interview address: The following are my views on the big AI infrastructure in the United States."]
  implication["It is feasible in five years, but it is really not necessarily true in five to ten years, and it is unlikel..."]
  thesis -->|frames| signal
  signal -->|develops| operator
  operator -->|lands in| implication

Source: View the original post